Safe Haskell	None
Language	Haskell2010

TensorFlow.GenOps.Core

Synopsis

Documentation

_HostRecv Source

Arguments

:: TensorType tensor_type
=> Int64	send_device_incarnation: The current incarnation of send_device.
-> Build (Tensor Value tensor_type)	tensor: The tensor to receive.

Receives the named tensor from send_device on recv_device.

_HostRecv requires its input on host memory whereas _Recv requires its input on device memory.

_HostSend Source

Arguments

:: TensorType t
=> Int64	send_device_incarnation: The current incarnation of send_device.
-> Tensor v1 t	tensor: The tensor to send.
-> Build ControlNode

Sends the named tensor from send_device to recv_device.

_HostSend requires its input on host memory whereas _Send requires its input on device memory.

_Recv Source

Arguments

:: TensorType tensor_type
=> Int64	send_device_incarnation: The current incarnation of send_device.
-> Build (Tensor Value tensor_type)	tensor: The tensor to receive.

Receives the named tensor from send_device on recv_device.

_Send Source

Arguments

:: TensorType t
=> Int64	send_device_incarnation: The current incarnation of send_device.
-> Tensor v1 t	tensor: The tensor to send.
-> Build ControlNode

Sends the named tensor from send_device to recv_device.

noOp :: ControlNode Source

Does nothing. Only useful as a placeholder for control edges.

_Retval Source

Arguments

:: TensorType t
=> Int64	index: This return value is the index-th return value of the function.
-> Tensor v1 t	input: The return value.
-> Build ControlNode

A graph node which represents a return value of a function.

_Arg Source

Arguments

:: TensorType t
=> Int64	index: This argument is the index-th argument of the function.
-> Build (Tensor Value t)	output: The argument.

A graph node which represents an argument to a function.

quantizedBatchNormWithGlobalNormalization Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Bool	scale_after_normalization: A bool indicating whether the resulted tensor needs to be multiplied with gamma.
-> Float	variance_epsilon: A small float number to avoid dividing by 0.
-> Tensor v1 tinput	t: A 4D input Tensor.
-> Tensor v2 Float	t_min: The value represented by the lowest quantized input.
-> Tensor v3 Float	t_max: The value represented by the highest quantized input.
-> Tensor v4 tinput	m: A 1D mean Tensor with size matching the last dimension of t. This is the first output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v5 Float	m_min: The value represented by the lowest quantized mean.
-> Tensor v6 Float	m_max: The value represented by the highest quantized mean.
-> Tensor v7 tinput	v: A 1D variance Tensor with size matching the last dimension of t. This is the second output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v8 Float	v_min: The value represented by the lowest quantized variance.
-> Tensor v9 Float	v_max: The value represented by the highest quantized variance.
-> Tensor v10 tinput	beta: A 1D beta Tensor with size matching the last dimension of t. An offset to be added to the normalized tensor.
-> Tensor v11 Float	beta_min: The value represented by the lowest quantized offset.
-> Tensor v12 Float	beta_max: The value represented by the highest quantized offset.
-> Tensor v13 tinput	gamma: A 1D gamma Tensor with size matching the last dimension of t. If "scale_after_normalization" is true, this tensor will be multiplied with the normalized tensor.
-> Tensor v14 Float	gamma_min: The value represented by the lowest quantized gamma.
-> Tensor v15 Float	gamma_max: The value represented by the highest quantized gamma.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(result, result_min, result_max) result result_min result_max

Quantized Batch normalization.

This op is deprecated and will be removed in the future. Prefer `tf.nn.batch_normalization`.

quantizedRelu6 Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	features
-> Tensor v2 Float	min_features: The float value that the lowest quantized value represents.
-> Tensor v3 Float	max_features: The float value that the highest quantized value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(activations, min_activations, max_activations) activations: Has the same output shape as "features". min_activations: The float value that the lowest quantized value represents. max_activations: The float value that the highest quantized value represents.

Computes Quantized Rectified Linear 6: `min(max(features, 0), 6)`

quantizedBiasAdd Source

Arguments

:: (TensorType t1, OneOf `[Int16, Int32, Word16, Word8]` t1, TensorType t2, OneOf `[Int16, Int32, Word16, Word8]` t2, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 t1	input
-> Tensor v2 t2	bias: A 1D bias Tensor with size matching the last dimension of `input`.
-> Tensor v3 Float	min_input: The float value that the lowest quantized input value represents.
-> Tensor v4 Float	max_input: The float value that the highest quantized input value represents.
-> Tensor v5 Float	min_bias: The float value that the lowest quantized bias value represents.
-> Tensor v6 Float	max_bias: The float value that the highest quantized bias value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(output, min_out, max_out) output min_out: The float value that the lowest quantized output value represents. max_out: The float value that the highest quantized output value represents.

Adds Tensor bias to Tensor input for Quantized types.

Broadcasts the values of bias on dimensions 0..N-2 of input.

fractionalAvgPoolGrad Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)
=> Tensor v1 Int64	orig_input_tensor_shape: Original input tensor shape for `fractional_avg_pool`
-> Tensor v2 t	out_backprop: 4-D with shape `[batch, height, width, channels]`. Gradients w.r.t. the output of `fractional_avg_pool`.
-> Tensor v3 Int64	row_pooling_sequence: row pooling sequence, form pooling region with col_pooling_sequence.
-> Tensor v4 Int64	col_pooling_sequence: column pooling sequence, form pooling region with row_pooling sequence.
-> Tensor Value t	output: 4-D. Gradients w.r.t. the input of `fractional_avg_pool`.

Computes gradient of the FractionalAvgPool function.

Unlike FractionalMaxPoolGrad, we don't need to find arg_max for FractionalAvgPoolGrad, we just need to evenly back-propagate each element of out_backprop to those indices that form the same pooling cell. Therefore, we just need to know the shape of original input tensor, instead of the whole tensor.

fractionalMaxPoolGrad Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)
=> Tensor v1 t	orig_input: Original input for `fractional_max_pool`
-> Tensor v2 t	orig_output: Original output for `fractional_max_pool`
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, height, width, channels]`. Gradients w.r.t. the output of `fractional_max_pool`.
-> Tensor v4 Int64	row_pooling_sequence: row pooling sequence, form pooling region with col_pooling_sequence.
-> Tensor v5 Int64	col_pooling_sequence: column pooling sequence, form pooling region with row_pooling sequence.
-> Tensor Value t	output: 4-D. Gradients w.r.t. the input of `fractional_max_pool`.

Computes gradient of the FractionalMaxPool function.

fractionalMaxPool Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)

=> Tensor v1 t

value: 4-D with shape `[batch, height, width, channels]`.

-> (Tensor Value t, Tensor Value Int64, Tensor Value Int64)

(output, row_pooling_sequence, col_pooling_sequence)

output: output tensor after fractional max pooling.
row_pooling_sequence: row pooling sequence, needed to calculate gradient.
col_pooling_sequence: column pooling sequence, needed to calculate gradient.

Performs fractional max pooling on the input.

Fractional max pooling is slightly different than regular max pooling. In regular max pooling, you downsize an input set by taking the maximum value of smaller N x N subsections of the set (often 2x2), and try to reduce the set by a factor of N, where N is an integer. Fractional max pooling, as you might expect from the word "fractional", means that the overall reduction ratio N does not have to be an integer.

The sizes of the pooling regions are generated randomly but are fairly uniform. For example, let's look at the height dimension, and the constraints on the list of rows that will be pool boundaries.

First we define the following:

input_row_length : the number of rows from the input set
output_row_length : which will be smaller than the input
alpha = input_row_length / output_row_length : our reduction ratio
K = floor(alpha)
row_pooling_sequence : this is the result list of pool boundary rows

Then, row_pooling_sequence should satisfy:

a[0] = 0 : the first value of the sequence is 0
a[end] = input_row_length : the last value of the sequence is the size
K <= (a[i+1] - a[i]) <= K+1 : all intervals are K or K+1 size
length(row_pooling_sequence) = output_row_length+1

For more details on fractional max pooling, see this paper: Benjamin Graham, Fractional Max-Pooling

topK Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Int64	k: Number of top elements to look for along the last dimension (along each row for matrices).
-> Tensor v1 t	input: 1-D or higher with last dimension at least `k`.
-> (Tensor Value t, Tensor Value Int32)	(values, indices) values: The `k` largest elements along each last dimensional slice. indices: The indices of `values` within the last dimension of `input`.

Finds values and indices of the k largest elements for the last dimension.

If the input is a vector (rank-1), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus `values[j]` is the j-th largest entry in input, and its index is `indices[j]`.

For matrices (resp. higher rank input), computes the top k entries in each row (resp. vector along the last dimension). Thus,

values.shape = indices.shape = input.shape[:-1] + [k]

If two elements are equal, the lower-index element appears first.

If k varies dynamically, use TopKV2 below.

inTopK Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64]` t)
=> Int64	k: Number of top elements to look at for computing precision.
-> Tensor v1 Float	predictions: A `batch_size` x `classes` tensor.
-> Tensor v2 t	targets: A `batch_size` vector of class ids.
-> Tensor Value Bool	precision: Computed Precision at `k` as a `bool Tensor`.

Says whether the targets are in the top K predictions.

This outputs a batch_size bool array, an entry `out[i]` is true if the prediction for the target class is among the top k predictions among all predictions for example i. Note that the behavior of InTopK differs from the TopK op in its handling of ties; if multiple classes have the same prediction value and straddle the top-k boundary, all of those classes are considered to be in the top k.

More formally, let

$predictions_i$ be the predictions for all classes for example i, $targets_i$ be the target class for example i, $out_i$ be the output for example i,

$$out_i = predictions_{i, targets_i} in TopKIncludingTies(predictions_i)$$

sparseSoftmaxCrossEntropyWithLogits Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t, TensorType tlabels, OneOf `[Int32, Int64]` tlabels)
=> Tensor v1 t	features: batch_size x num_classes matrix
-> Tensor v2 tlabels	labels: batch_size vector with values in [0, num_classes). This is the label for the given minibatch entry.
-> (Tensor Value t, Tensor Value t)	(loss, backprop) loss: Per example loss (batch_size vector). backprop: backpropagated gradients (batch_size x num_classes matrix).

Computes softmax cross entropy cost and gradients to backpropagate.

Unlike SoftmaxCrossEntropyWithLogits, this operation does not accept a matrix of label probabilities, but rather a single label per row of features. This label is considered to have probability 1.0 for the given row.

Inputs are the logits, not probabilities.

softmaxCrossEntropyWithLogits Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	features: batch_size x num_classes matrix
-> Tensor v2 t	labels: batch_size x num_classes matrix The caller must ensure that each batch of labels represents a valid probability distribution.
-> (Tensor Value t, Tensor Value t)	(loss, backprop) loss: Per example loss (batch_size vector). backprop: backpropagated gradients (batch_size x num_classes matrix).

Computes softmax cross entropy cost and gradients to backpropagate.

Inputs are the logits, not probabilities.

logSoftmax Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	logits: 2-D with shape `[batch_size, num_classes]`.
-> Tensor Value t	logsoftmax: Same shape as `logits`.

Computes log softmax activations.

For each batch i and class j we have

logsoftmax[i, j] = logits[i, j] - log(sum(exp(logits[i])))

softsignGrad Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	gradients: The backpropagated gradients to the corresponding softsign operation.
-> Tensor v2 t	features: The features passed as input to the corresponding softsign operation.
-> Tensor Value t	backprops: The gradients: `gradients / (1 + abs(-features)) ** 2`.

Computes softsign gradients for a softsign operation.

softplus Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	features
-> Tensor Value t	activations

Computes softplus: `log(exp(features) + 1)`.

eluGrad Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	gradients: The backpropagated gradients to the corresponding Elu operation.
-> Tensor v2 t	outputs: The outputs of the corresponding Elu operation.
-> Tensor Value t	backprops: The gradients: `gradients * (outputs + 1)` if outputs < 0, `gradients` otherwise.

Computes gradients for the exponential linear (Elu) operation.

elu Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	features
-> Tensor Value t	activations

Computes exponential linear: `exp(features) - 1` if < 0, features otherwise.

See Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

relu6 Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	features
-> Tensor Value t	activations

Computes rectified linear 6: `min(max(features, 0), 6)`.

reluGrad Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	gradients: The backpropagated gradients to the corresponding Relu operation.
-> Tensor v2 t	features: The features passed as input to the corresponding Relu operation, OR the outputs of that operation (both work equivalently).
-> Tensor Value t	backprops: `gradients * (features > 0)`.

Computes rectified linear gradients for a Relu operation.

dilation2DBackpropInput Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, depth]`.
-> Tensor v2 t	filter: 3-D with shape `[filter_height, filter_width, depth]`.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, depth]`.
-> Tensor Value t	in_backprop: 4-D with shape `[batch, in_height, in_width, depth]`.

Computes the gradient of morphological 2-D dilation with respect to the input.

maxPoolGrad Source

Arguments

:: (TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	orig_input: The original input tensor.
-> Tensor v2 t	orig_output: The original output tensor.
-> Tensor v3 t	grad: 4-D. Gradients w.r.t. the output of `max_pool`.
-> Tensor Value t	output: Gradients w.r.t. the input to `max_pool`.

Computes gradients of the maxpooling function.

lRNGrad Source

Arguments

:: (TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	input_grads: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 t	input_image: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v3 t	output_image: 4-D with shape `[batch, height, width, channels]`.
-> Tensor Value t	output: The gradients for LRN.

Gradients for Local Response Normalization.

maxPool3DGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Float	orig_input: The original input tensor.
-> Tensor v2 Float	orig_output: The original output tensor.
-> Tensor v3 t	grad: Output backprop of shape `[batch, depth, rows, cols, channels]`.
-> Tensor Value t	output

Computes gradients of max pooling function.

conv3DBackpropFilterV2 Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: Shape `[batch, depth, rows, cols, in_channels]`.
-> Tensor v2 Int32	filter_sizes: An integer vector representing the tensor shape of `filter`, where `filter` is a 5-D `[filter_depth, filter_height, filter_width, in_channels, out_channels]` tensor.
-> Tensor v3 t	out_backprop: Backprop signal of shape `[batch, out_depth, out_rows, out_cols, out_channels]`.
-> Tensor Value t	output

Computes the gradients of 3-D convolution with respect to the filter.

conv3DBackpropFilter Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: Shape `[batch, depth, rows, cols, in_channels]`.
-> Tensor v2 t	filter: Shape `[depth, rows, cols, in_channels, out_channels]`. `in_channels` must match between `input` and `filter`.
-> Tensor v3 t	out_backprop: Backprop signal of shape `[batch, out_depth, out_rows, out_cols, out_channels]`.
-> Tensor Value t	output

Computes the gradients of 3-D convolution with respect to the filter.

conv3D Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: Shape `[batch, in_depth, in_height, in_width, in_channels]`.
-> Tensor v2 t	filter: Shape `[filter_depth, filter_height, filter_width, in_channels, out_channels]`. `in_channels` must match between `input` and `filter`.
-> Tensor Value t	output

Computes a 3-D convolution given 5-D input and filter tensors.

In signal processing, cross-correlation is a measure of similarity of two waveforms as a function of a time-lag applied to one of them. This is also known as a sliding dot product or sliding inner-product.

Our Conv3D implements a form of cross-correlation.

depthwiseConv2dNativeBackpropFilter Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, in_channels]`.
-> Tensor v2 Int32	filter_sizes: An integer vector representing the tensor shape of `filter`, where `filter` is a 4-D `[filter_height, filter_width, in_channels, depthwise_multiplier]` tensor.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, out_channels]`. Gradients w.r.t. the output of the convolution.
-> Tensor Value t	output: 4-D with shape `[filter_height, filter_width, in_channels, out_channels]`. Gradient w.r.t. the `filter` input of the convolution.

Computes the gradients of depthwise convolution with respect to the filter.

conv2DBackpropFilter Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, in_channels]`.
-> Tensor v2 Int32	filter_sizes: An integer vector representing the tensor shape of `filter`, where `filter` is a 4-D `[filter_height, filter_width, in_channels, out_channels]` tensor.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, out_channels]`. Gradients w.r.t. the output of the convolution.
-> Tensor Value t	output: 4-D with shape `[filter_height, filter_width, in_channels, out_channels]`. Gradient w.r.t. the `filter` input of the convolution.

Computes the gradients of convolution with respect to the filter.

conv2DBackpropInput Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 Int32	input_sizes: An integer vector representing the shape of `input`, where `input` is a 4-D `[batch, height, width, channels]` tensor.
-> Tensor v2 t	filter: 4-D with shape `[filter_height, filter_width, in_channels, out_channels]`.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, out_channels]`. Gradients w.r.t. the output of the convolution.
-> Tensor Value t	output: 4-D with shape `[batch, in_height, in_width, in_channels]`. Gradient w.r.t. the input of the convolution.

Computes the gradients of convolution with respect to the input.

conv2D Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	input
-> Tensor v2 t	filter
-> Tensor Value t	output

Computes a 2-D convolution given 4-D input and filter tensors.

Given an input tensor of shape `[batch, in_height, in_width, in_channels]` and a filter / kernel tensor of shape `[filter_height, filter_width, in_channels, out_channels]`, this op performs the following:

Flattens the filter to a 2-D matrix with shape `[filter_height * filter_width * in_channels, output_channels]`.
Extracts image patches from the input tensor to form a *virtual* tensor of shape `[batch, out_height, out_width, filter_height * filter_width * in_channels]`.
For each patch, right-multiplies the filter matrix and the image patch vector.

In detail, with the default NHWC format,

output[b, i, j, k] = sum_{di, dj, q} input[b, strides[1] * i + di, strides[2] * j + dj, q] * filter[di, dj, q, k]

Must have `strides[0] = strides[3] = 1`. For the most common case of the same horizontal and vertices strides, `strides = [1, stride, stride, 1]`.

biasAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	value: Any number of dimensions.
-> Tensor v2 t	bias: 1-D with size the last dimension of `value`.
-> Tensor Value t	output: Broadcasted sum of `value` and `bias`.

Adds bias to value.

This is a special case of `tf.add` where bias is restricted to be 1-D. Broadcasting is supported, so value may have any number of dimensions.

fusedBatchNorm Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x: A 4D Tensor for input data.
-> Tensor v2 t	scale: A 1D Tensor for scaling factor, to scale the normalized x.
-> Tensor v3 t	offset: A 1D Tensor for offset, to shift to the normalized x.
-> Tensor v4 t	mean: A 1D Tensor for population mean. Used for inference only; must be empty for training.
-> Tensor v5 t	variance: A 1D Tensor for population variance. Used for inference only; must be empty for training.
-> (Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t)	(y, batch_mean, batch_variance, reserve_space_1, reserve_space_2) y: A 4D Tensor for output data. batch_mean: A 1D Tensor for the computed batch mean, to be used by TensorFlow to compute the running mean. batch_variance: A 1D Tensor for the computed batch variance, to be used by TensorFlow to compute the running variance. reserve_space_1: A 1D Tensor for the computed batch mean, to be reused in the gradient computation. reserve_space_2: A 1D Tensor for the computed batch variance (inverted variance in the cuDNN case), to be used in the gradient computation.

Batch normalization.

Note that the size of 4D Tensors are defined by either NHWC or NCHW. The size of 1D Tensors matches the dimension C of the 4D Tensors.

batchNormWithGlobalNormalizationGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Bool	scale_after_normalization: A bool indicating whether the resulted tensor needs to be multiplied with gamma.
-> Float	variance_epsilon: A small float number to avoid dividing by 0.
-> Tensor v1 t	t: A 4D input Tensor.
-> Tensor v2 t	m: A 1D mean Tensor with size matching the last dimension of t. This is the first output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v3 t	v: A 1D variance Tensor with size matching the last dimension of t. This is the second output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v4 t	gamma: A 1D gamma Tensor with size matching the last dimension of t. If "scale_after_normalization" is true, this Tensor will be multiplied with the normalized Tensor.
-> Tensor v5 t	backprop: 4D backprop Tensor.
-> (Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t)	(dx, dm, dv, db, dg) dx: 4D backprop tensor for input. dm: 1D backprop tensor for mean. dv: 1D backprop tensor for variance. db: 1D backprop tensor for beta. dg: 1D backprop tensor for gamma.

Gradients for batch normalization.

This op is deprecated. See `tf.nn.batch_normalization`.

batchFFT3D Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

batchIFFT2D Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

avgPool Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	value: 4-D with shape `[batch, height, width, channels]`.
-> Tensor Value t	output: The average pooled output tensor.

Performs average pooling on the input.

Each entry in output is the mean of the corresponding size ksize window in value.

batchFFT2D Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

batchFFT Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

requantizationRange Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput)
=> Tensor v1 tinput	input
-> Tensor v2 Float	input_min: The float value that the minimum quantized input value represents.
-> Tensor v3 Float	input_max: The float value that the maximum quantized input value represents.
-> (Tensor Value Float, Tensor Value Float)	(output_min, output_max) output_min: The computed min output. output_max: the computed max output.

Given a quantized tensor described by (input, input_min, input_max), outputs a

range that covers the actual values present in that tensor. This op is typically used to produce the requested_output_min and requested_output_max for Requantize.

requantize Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	input
-> Tensor v2 Float	input_min: The float value that the minimum quantized input value represents.
-> Tensor v3 Float	input_max: The float value that the maximum quantized input value represents.
-> Tensor v4 Float	requested_output_min: The float value that the minimum quantized output value represents.
-> Tensor v5 Float	requested_output_max: The float value that the maximum quantized output value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(output, output_min, output_max) output output_min: The requested_output_min value is copied into this output. output_max: The requested_output_max value is copied into this output.

Convert the quantized input tensor into a lower-precision output, using the

output range specified with requested_output_min and requested_output_max.

input_min, input_max: are scalar floats that specify the range for the float interpretation of the input data. For example, if input_min is -1.0f and input_max is 1.0f, and we are dealing with quint16 quantized data, then a 0 value in the 16-bit data should be interpreted as -1.0f, and a 65535 means 1.0f.

quantizeDownAndShrinkRange Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	input
-> Tensor v2 Float	input_min: The float value that the minimum quantized input value represents.
-> Tensor v3 Float	input_max: The float value that the maximum quantized input value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(output, output_min, output_max) output output_min: The float value that the minimum quantized output value represents. output_max: The float value that the maximum quantized output value represents.

Convert the quantized input tensor into a lower-precision output, using the

actual distribution of the values to maximize the usage of the lower bit depth and adjusting the output min and max ranges accordingly.

input_min, input_max: are scalar floats that specify the range for the float interpretation of the input data. For example, if input_min is -1.0f and input_max is 1.0f, and we are dealing with quint16 quantized data, then a 0 value in the 16-bit data should be interpreted as -1.0f, and a 65535 means 1.0f.

This operator tries to squeeze as much precision as possible into an output with a lower bit depth by calculating the actual min and max values found in the data. For example, maybe that quint16 input has no values lower than 16,384 and none higher than 49,152. That means only half the range is actually needed, all the float interpretations are between -0.5f and 0.5f, so if we want to compress the data into a quint8 output, we can use that range rather than the theoretical -1.0f to 1.0f that is suggested by the input min and max.

In practice, this is most useful for taking output from operations like QuantizedMatMul that can produce higher bit-depth outputs than their inputs and may have large potential output ranges, but in practice have a distribution of input values that only uses a small fraction of the possible range. By feeding that output into this operator, we can reduce it from 32 bits down to 8 with minimal loss of accuracy.

quantizedMatMul Source

Arguments

:: (TensorType t1, OneOf `[Int16, Int32, Word16, Word8]` t1, TensorType t2, OneOf `[Int16, Int32, Word16, Word8]` t2, TensorType toutput, OneOf `[Int16, Int32, Word16, Word8]` toutput)
=> Tensor v1 t1	a: Must be a two-dimensional tensor.
-> Tensor v2 t2	b: Must be a two-dimensional tensor.
-> Tensor v3 Float	min_a: The float value that the lowest quantized `a` value represents.
-> Tensor v4 Float	max_a: The float value that the highest quantized `a` value represents.
-> Tensor v5 Float	min_b: The float value that the lowest quantized `b` value represents.
-> Tensor v6 Float	max_b: The float value that the highest quantized `b` value represents.
-> (Tensor Value toutput, Tensor Value Float, Tensor Value Float)	(out, min_out, max_out) out min_out: The float value that the lowest quantized output value represents. max_out: The float value that the highest quantized output value represents.

Perform a quantized matrix multiplication of a by the matrix b.

The inputs must be two-dimensional matrices and the inner dimension of a (after being transposed if transpose_a is non-zero) must match the outer dimension of b (after being transposed if transposed_b is non-zero).

cumprod Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	x
-> Tensor v2 tidx	axis
-> Tensor Value t	out

Compute the cumulative product of the tensor x along axis.

By default, this op performs an inclusive cumprod, which means that the first element of the input is identical to the first element of the output: ```prettyprint tf.cumprod([a, b, c]) ==> [a, a * b, a * b * c] ```

By setting the exclusive kwarg to True, an exclusive cumprod is performed instead: ```prettyprint tf.cumprod([a, b, c], exclusive=True) ==> [0, a, a * b] ```

By setting the reverse kwarg to True, the cumprod is performed in the opposite direction: ```prettyprint tf.cumprod([a, b, c], reverse=True) ==> [a * b * c, b * c, c] ``` This is more efficient than using separate `tf.reverse` ops.

The reverse and exclusive kwargs can also be combined: ```prettyprint tf.cumprod([a, b, c], exclusive=True, reverse=True) ==> [b * c, c, 0] ```

cumsum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	x
-> Tensor v2 tidx	axis
-> Tensor Value t	out

Compute the cumulative sum of the tensor x along axis.

By default, this op performs an inclusive cumsum, which means that the first element of the input is identical to the first element of the output: ```prettyprint tf.cumsum([a, b, c]) ==> [a, a + b, a + b + c] ```

By setting the exclusive kwarg to True, an exclusive cumsum is performed instead: ```prettyprint tf.cumsum([a, b, c], exclusive=True) ==> [0, a, a + b] ```

By setting the reverse kwarg to True, the cumsum is performed in the opposite direction: ```prettyprint tf.cumsum([a, b, c], reverse=True) ==> [a + b + c, b + c, c] ``` This is more efficient than using separate `tf.reverse` ops.

The reverse and exclusive kwargs can also be combined: ```prettyprint tf.cumsum([a, b, c], exclusive=True, reverse=True) ==> [b + c, c, 0] ```

cross Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	a: A tensor containing 3-element vectors.
-> Tensor v2 t	b: Another tensor, of same type and shape as `a`.
-> Tensor Value t	product: Pairwise cross product of the vectors in `a` and `b`.

Compute the pairwise cross product.

a and b must be the same shape; they can either be simple 3-element vectors, or any shape where the innermost dimension is 3. In the latter case, each pair of corresponding 3-element vectors is cross-multiplied independently.

iFFT3D Source

Arguments

:: Tensor v1 (Complex Float)

input: A complex64 tensor.

-> Tensor Value (Complex Float)

output: A complex64 tensor of the same shape as input. The inner-most 3 dimensions of input are replaced with their inverse 3D Fourier Transform.

compatibility(numpy) Equivalent to np.fft3end_compatibility

Compute the inverse 3-dimensional discrete Fourier Transform over the inner-most

3 dimensions of input.

fFT3D Source

Arguments

:: Tensor v1 (Complex Float)

input: A complex64 tensor.

-> Tensor Value (Complex Float)

output: A complex64 tensor of the same shape as input. The inner-most 3 dimensions of input are replaced with their 3D Fourier Transform.

compatibility(numpy) Equivalent to np.fft3end_compatibility

Compute the 3-dimensional discrete Fourier Transform over the inner-most 3

dimensions of input.

maxPoolGradWithArgmax Source

Arguments

:: (TensorType targmax, OneOf `[Int32, Int64]` targmax, TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	input: The original input.
-> Tensor v2 t	grad: 4-D with shape `[batch, height, width, channels]`. Gradients w.r.t. the output of `max_pool`.
-> Tensor v3 targmax	argmax: The indices of the maximum values chosen for each output of `max_pool`.
-> Tensor Value t	output: Gradients w.r.t. the input of `max_pool`.

Computes gradients of the maxpooling function.

fFT2D Source

Arguments

:: Tensor v1 (Complex Float)

input: A complex64 tensor.

-> Tensor Value (Complex Float)

output: A complex64 tensor of the same shape as input. The inner-most 2 dimensions of input are replaced with their 2D Fourier Transform.

compatibility(numpy) Equivalent to np.fft2end_compatibility

Compute the 2-dimensional discrete Fourier Transform over the inner-most

2 dimensions of input.

iFFT Source

Arguments

:: Tensor v1 (Complex Float)	input: A complex64 tensor.
-> Tensor Value (Complex Float)	output: A complex64 tensor of the same shape as `input`. The inner-most dimension of `input` is replaced with its inverse 1D Fourier Transform.

Compute the inverse 1-dimensional discrete Fourier Transform over the inner-most

dimension of input.

fFT Source

Arguments

:: Tensor v1 (Complex Float)	input: A complex64 tensor.
-> Tensor Value (Complex Float)	output: A complex64 tensor of the same shape as `input`. The inner-most dimension of `input` is replaced with its 1D Fourier Transform.

Compute the 1-dimensional discrete Fourier Transform over the inner-most

dimension of input.

conj Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float]` t)
=> Tensor v1 t	input
-> Tensor Value t	output

Returns the complex conjugate of a complex number.

Given a tensor input of complex numbers, this operation returns a tensor of complex numbers that are the complex conjugate of each element in input. The complex numbers in input must be of the form $a + bj$, where *a* is the real part and *b* is the imaginary part.

The complex conjugate returned by this operation is of the form $a - bj$.

For example:

``` # tensor input is [-2.25 + 4.75j, 3.25 + 5.75j] tf.conj(input) ==> [-2.25 - 4.75j, 3.25 - 5.75j] ```

real Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float]` t, TensorType tout, OneOf `[Double, Float]` tout)
=> Tensor v1 t	input
-> Tensor Value tout	output

Returns the real part of a complex number.

Given a tensor input of complex numbers, this operation returns a tensor of type float that is the real part of each element in input. All elements in input must be complex numbers of the form $a + bj$, where *a* is the real part returned by this operation and *b* is the imaginary part.

For example:

``` # tensor input is [-2.25 + 4.75j, 3.25 + 5.75j] tf.real(input) ==> [-2.25, 3.25] ```

complex Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t, TensorType tout, OneOf `[Complex Double, Complex Float]` tout)
=> Tensor v1 t	real
-> Tensor v2 t	imag
-> Tensor Value tout	out

Converts two real numbers to a complex number.

Given a tensor real representing the real part of a complex number, and a tensor imag representing the imaginary part of a complex number, this operation returns complex numbers elementwise of the form $a + bj$, where *a* represents the real part and *b* represents the imag part.

The input tensors real and imag must have the same shape.

For example:

``` # tensor real is [2.25, 3.25] # tensor imag is [4.75, 5.75] tf.complex(real, imag) ==> [[2.25 + 4.75j], [3.25 + 5.75j]] ```

range Source

Arguments

:: (TensorType tidx, OneOf `[Int32, Int64, Double, Float]` tidx)
=> Tensor v1 tidx	start: 0-D (scalar). First entry in the sequence.
-> Tensor v2 tidx	limit: 0-D (scalar). Upper limit of sequence, exclusive.
-> Tensor v3 tidx	delta: 0-D (scalar). Optional. Default is 1. Number that increments `start`.
-> Tensor Value tidx	output: 1-D.

Creates a sequence of numbers.

This operation creates a sequence of numbers that begins at start and extends by increments of delta up to but not including limit.

For example:

``` # start is 3 # limit is 18 # delta is 3 tf.range(start, limit, delta) ==> [3, 6, 9, 12, 15] ```

any Source

Arguments

:: (TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 Bool	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value Bool	output: The reduced tensor.

Computes the "logical or" of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

sparseSegmentMean Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	data
-> Tensor v2 tidx	indices: A 1-D tensor. Has same rank as `segment_ids`.
-> Tensor v3 Int32	segment_ids: A 1-D tensor. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the mean along sparse segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Like SegmentMean, but segment_ids can have rank less than `data`'s first dimension, selecting a subset of dimension 0, specified by indices.

sparseSegmentSum Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	data
-> Tensor v2 tidx	indices: A 1-D tensor. Has same rank as `segment_ids`.
-> Tensor v3 Int32	segment_ids: A 1-D tensor. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the sum along sparse segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Like SegmentSum, but segment_ids can have rank less than `data`'s first dimension, selecting a subset of dimension 0, specified by indices.

For example:

```prettyprint c = tf.constant([[1,2,3,4], [-1,-2,-3,-4], [5,6,7,8]])

# Select two rows, one segment. tf.sparse_segment_sum(c, tf.constant([0, 1]), tf.constant([0, 0])) ==> [[0 0 0 0]]

# Select two rows, two segment. tf.sparse_segment_sum(c, tf.constant([0, 1]), tf.constant([0, 1])) ==> [[ 1 2 3 4] [-1 -2 -3 -4]]

# Select all rows, two segments. tf.sparse_segment_sum(c, tf.constant([0, 1, 2]), tf.constant([0, 0, 1])) ==> [[0 0 0 0] [5 6 7 8]]

# Which is equivalent to: tf.segment_sum(c, tf.constant([0, 0, 1])) ```

unsortedSegmentSum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A tensor whose shape is a prefix of `data.shape`.
-> Tensor v3 Int32	num_segments
-> Tensor Value t	output: Has same shape as data, except for the first `segment_ids.rank` dimensions, which are replaced with a single dimension which has size `num_segments`.

Computes the sum along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that `(output[i] = sum_{j...} data[j...]` where the sum is over tuples `j...` such that `segment_ids[j...] == i`. Unlike SegmentSum, segment_ids need not be sorted and need not cover all values in the full range of valid values.

If the sum is empty for a given segment ID i, `output[i] = 0`.

num_segments should equal the number of distinct segment IDs.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/UnsortedSegmentSum.png" alt /div

segmentMin Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A 1-D tensor whose rank is equal to the rank of `data`'s first dimension. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the minimum along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that $output_i = min_j(data_j)$ where min is over j such that `segment_ids[j] == i`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/SegmentMin.png" alt /div

segmentProd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A 1-D tensor whose rank is equal to the rank of `data`'s first dimension. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the product along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that $output_i = prod_j data_j$ where the product is over j such that `segment_ids[j] == i`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/SegmentProd.png" alt /div

segmentMean Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A 1-D tensor whose rank is equal to the rank of `data`'s first dimension. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the mean along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that $output_i = frac{sum_j data_j}{N}$ where mean is over j such that `segment_ids[j] == i` and N is the total number of values summed.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/SegmentMean.png" alt /div

segmentSum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A 1-D tensor whose rank is equal to the rank of `data`'s first dimension. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the sum along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that $output_i = sum_j data_j$ where sum is over j such that `segment_ids[j] == i`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/SegmentSum.png" alt /div

argMin Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input
-> Tensor v2 tidx	dimension: int32, 0 <= dimension < rank(input). Describes which dimension of the input Tensor to reduce across. For vectors, use dimension = 0.
-> Tensor Value Int64	output

Returns the index with the smallest value across dimensions of a tensor.

max Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value t	output: The reduced tensor.

Computes the maximum of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

min Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value t	output: The reduced tensor.

Computes the minimum of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

prod Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value t	output: The reduced tensor.

Computes the product of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

sum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value t	output: The reduced tensor.

Computes the sum of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

sparseMatMul Source

Arguments

:: (TensorType ta, OneOf `[Word16, Float]` ta, TensorType tb, OneOf `[Word16, Float]` tb)
=> Tensor v1 ta	a
-> Tensor v2 tb	b
-> Tensor Value Float	product

Multiply matrix "a" by matrix "b".

The inputs must be two-dimensional matrices and the inner dimension of "a" must match the outer dimension of "b". This op is optimized for the case where at least one of "a" or "b" is sparse. The breakeven for using this versus a dense matrix multiply on one platform was 30% zero values in the sparse matrix.

matMul Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Word16, Double, Float]` t)
=> Tensor v1 t	a
-> Tensor v2 t	b
-> Tensor Value t	product

Multiply the matrix "a" by the matrix "b".

The inputs must be two-dimensional matrices and the inner dimension of "a" (after being transposed if transpose_a is true) must match the outer dimension of "b" (after being transposed if transposed_b is true).

Note*: The default kernel implementation for MatMul on GPUs uses cublas.

logicalAnd Source

Arguments

:: Tensor v1 Bool	x
-> Tensor v2 Bool	y
-> Tensor Value Bool	z

Returns the truth value of x AND y element-wise.

NOTE*: LogicalAnd supports broadcasting. More about broadcasting here

equal Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Bool, ByteString, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x == y) element-wise.

NOTE*: Equal supports broadcasting. More about broadcasting here

greaterEqual Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x >= y) element-wise.

NOTE*: GreaterEqual supports broadcasting. More about broadcasting here

lessEqual Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x <= y) element-wise.

NOTE*: LessEqual supports broadcasting. More about broadcasting here

less Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x < y) element-wise.

NOTE*: Less supports broadcasting. More about broadcasting here

polygamma Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	a
-> Tensor v2 t	x
-> Tensor Value t	z

Compute the polygamma function $psi^{(n)}(x)$.

The polygamma function is defined as:

``` psi^{(n)}(x) = frac{d^n}{dx^n} psi(x) ``` where $psi(x)$ is the digamma function.

igamma Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	a
-> Tensor v2 t	x
-> Tensor Value t	z

Compute the lower regularized incomplete Gamma function `Q(a, x)`.

The lower regularized incomplete Gamma function is defined as:

``` P(a, x) = gamma(a, x) / Gamma(a) = 1 - Q(a, x) ``` where ``` gamma(a, x) = int_{0}^{x} t^{a-1} exp(-t) dt ``` is the lower incomplete Gamma function.

Note, above `Q(a, x)` (Igammac) is the upper regularized complete Gamma function.

igammac Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	a
-> Tensor v2 t	x
-> Tensor Value t	z

Compute the upper regularized incomplete Gamma function `Q(a, x)`.

The upper regularized incomplete Gamma function is defined as:

``` Q(a, x) = Gamma(a, x) / Gamma(a) = 1 - P(a, x) ``` where ``` Gamma(a, x) = int_{x}^{infty} t^{a-1} exp(-t) dt ``` is the upper incomplete Gama function.

Note, above `P(a, x)` (Igamma) is the lower regularized complete Gamma function.

mod Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns element-wise remainder of division.

NOTE*: Mod supports broadcasting. More about broadcasting here

minimum Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns the min of x and y (i.e. x < y ? x : y) element-wise.

NOTE*: Minimum supports broadcasting. More about broadcasting here

maximum Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns the max of x and y (i.e. x > y ? x : y) element-wise.

NOTE*: Maximum supports broadcasting. More about broadcasting here

squaredDifference Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns (x - y)(x - y) element-wise.

NOTE*: SquaredDifference supports broadcasting. More about broadcasting here

softplusGrad Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	gradients: The backpropagated gradients to the corresponding softplus operation.
-> Tensor v2 t	features: The features passed as input to the corresponding softplus operation.
-> Tensor Value t	backprops: The gradients: `gradients / (1 + exp(-features))`.

Computes softplus gradients for a softplus operation.

batchToSpace Source

Arguments

:: (TensorType t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Int64	block_size
-> Tensor v1 t	input: 4-D tensor with shape `[batchblock_sizeblock_size, height_padblock_size, width_padblock_size, depth]`. Note that the batch size of the input tensor must be divisible by `block_size * block_size`.
-> Tensor v2 tidx	crops: 2-D tensor of non-negative integers with shape `[2, 2]`. It specifies how many elements to crop from the intermediate result across the spatial dimensions as follows: crops = [[crop_top, crop_bottom], [crop_left, crop_right]]
-> Tensor Value t	output: 4-D with shape `[batch, height, width, depth]`, where: height = height_pad - crop_top - crop_bottom width = width_pad - crop_left - crop_right The attr `block_size` must be greater than one. It indicates the block size. Some examples: For the following input of shape `[4, 1, 1, 1]` and block_size of 2: ```prettyprint [[[[1]]], [[[2]]], [[[3]]], [[[4]]]] ``` The output tensor has shape `[1, 2, 2, 1]` and value: ```prettyprint x = [[[[1], [2]], [[3], [4]]]] ``` For the following input of shape `[4, 1, 1, 3]` and block_size of 2: ```prettyprint [[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]] ``` The output tensor has shape `[1, 2, 2, 3]` and value: ```prettyprint x = [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]] ``` For the following input of shape `[4, 2, 2, 1]` and block_size of 2: ```prettyprint x = [[[[1], [3]], [[5], [7]]], [[[2], [4]], [[10], [12]]], [[[5], [7]], [[13], [15]]], [[[6], [8]], [[14], [16]]]] ``` The output tensor has shape `[1, 4, 4, 1]` and value: ```prettyprint x = [[[1], [2], [3], [4]], [[5], [6], [7], [8]], [[9], [10], [11], [12]], [[13], [14], [15], [16]]] ``` For the following input of shape `[8, 1, 2, 1]` and block_size of 2: ```prettyprint x = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]], [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]] ``` The output tensor has shape `[2, 2, 4, 1]` and value: ```prettyprint x = [[[[1], [3]], [[5], [7]]], [[[2], [4]], [[10], [12]]], [[[5], [7]], [[13], [15]]], [[[6], [8]], [[14], [16]]]] ```

BatchToSpace for 4-D tensors of type T.

This is a legacy version of the more general BatchToSpaceND.

Rearranges (permutes) data from batch into blocks of spatial data, followed by cropping. This is the reverse transformation of SpaceToBatch. More specifically, this op outputs a copy of the input tensor where values from the batch dimension are moved in spatial blocks to the height and width dimensions, followed by cropping along the height and width dimensions.

mul Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x * y element-wise.

NOTE*: Mul supports broadcasting. More about broadcasting here

rint Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Returns element-wise integer closest to x.

If the result is midway between two representable values, the even representable is chosen. For example:

``` rint(-1.5) ==> -2.0 rint(0.5000001) ==> 1.0 rint([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0]) ==> [-2., -2., -0., 0., 2., 2., 2.] ```

ceil Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Returns element-wise smallest integer in not less than x.

floor Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Returns element-wise largest integer not greater than x.

maxPool3D Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: Shape `[batch, depth, rows, cols, channels]` tensor to pool over.
-> Tensor Value t	output: The max pooled output tensor.

Performs 3D max pooling on the input.

isInf Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value Bool	y

Returns which elements of x are Inf.

compatibility(numpy) Equivalent to np.isinfend_compatibility

depthwiseConv2dNativeBackpropInput Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 Int32	input_sizes: An integer vector representing the shape of `input`, where `input` is a 4-D `[batch, height, width, channels]` tensor.
-> Tensor v2 t	filter: 4-D with shape `[filter_height, filter_width, in_channels, depthwise_multiplier]`.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, out_channels]`. Gradients w.r.t. the output of the convolution.
-> Tensor Value t	output: 4-D with shape `[batch, in_height, in_width, in_channels]`. Gradient w.r.t. the input of the convolution.

Computes the gradients of depthwise convolution with respect to the input.

isNan Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value Bool	y

Returns which elements of x are NaN.

compatibility(numpy) Equivalent to np.isnanend_compatibility

log1p Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes natural logarithm of (1 + x) element-wise.

I.e., $y = log_e (1 + x)$.

asin Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes asin of x element-wise.

topKV2 Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: 1-D or higher with last dimension at least `k`.
-> Tensor v2 Int32	k: 0-D. Number of top elements to look for along the last dimension (along each row for matrices).
-> (Tensor Value t, Tensor Value Int32)	(values, indices) values: The `k` largest elements along each last dimensional slice. indices: The indices of `values` within the last dimension of `input`.

Finds values and indices of the k largest elements for the last dimension.

If the input is a vector (rank-1), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus `values[j]` is the j-th largest entry in input, and its index is `indices[j]`.

For matrices (resp. higher rank input), computes the top k entries in each row (resp. vector along the last dimension). Thus,

values.shape = indices.shape = input.shape[:-1] + [k]

If two elements are equal, the lower-index element appears first.

This is the same as TopK, but takes k as in input rather than an attr.

cos Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes cos of x element-wise.

sin Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes sin of x element-wise.

randomUniformInt Source

Arguments

:: (TensorType tout, OneOf `[Int32, Int64]` tout, TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	shape: The shape of the output tensor.
-> Tensor v2 tout	minval: 0-D. Inclusive lower bound on the generated integers.
-> Tensor v3 tout	maxval: 0-D. Exclusive upper bound on the generated integers.
-> Build (Tensor Value tout)	output: A tensor of the specified shape filled with uniform random integers.

Outputs random integers from a uniform distribution.

The generated values are uniform integers in the range `[minval, maxval)`. The lower bound minval is included in the range, while the upper bound maxval is excluded.

The random integers are slightly biased unless `maxval - minval` is an exact power of two. The bias is small for values of `maxval - minval` significantly smaller than the range of the output (either `2^32` or `2^64`).

erfc Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the complementary error function of x element-wise.

digamma Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes Psi, the derivative of Lgamma (the log of the absolute value of

`Gamma(x)`), element-wise.

fusedResizeAndPadConv2D Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, in_channels]`.
-> Tensor v2 Int32	size: A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The new size for the images.
-> Tensor v3 Int32	paddings: A two-column matrix specifying the padding sizes. The number of rows must be the same as the rank of `input`.
-> Tensor v4 t	filter: 4-D with shape `[filter_height, filter_width, in_channels, out_channels]`.
-> Tensor Value t	output

Performs a resize and padding as a preprocess during a convolution.

It's often possible to do spatial transformations more efficiently as part of the packing stage of a convolution, so this op allows for an optimized implementation where these stages are fused together. This prevents the need to write out the intermediate results as whole tensors, reducing memory pressure, and we can get some latency gains by merging the transformation calculations. The data_format attribute for Conv2D isn't supported by this op, and defaults to NHWC order. Internally this op uses a single per-graph scratch buffer, which means that it will block if multiple versions are being run in parallel. This is because this operator is primarily an optimization to minimize memory usage.

sub Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x - y element-wise.

NOTE*: Sub supports broadcasting. More about broadcasting here

sign Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Returns an element-wise indication of the sign of a number.

`y = sign(x) = -1` if `x 0 if `x == 0`; 1 if `x 0`.

For complex numbers, `y = sign(x) = x / |x|` if `x != 0`, otherwise `y = 0`.

lgamma Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the log of the absolute value of `Gamma(x)` element-wise.

log Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes natural logarithm of x element-wise.

I.e., $y = log_e x$.

exp Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes exponential of x element-wise. $y = e^x$.

dilation2D Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, depth]`.
-> Tensor v2 t	filter: 3-D with shape `[filter_height, filter_width, depth]`.
-> Tensor Value t	output: 4-D with shape `[batch, out_height, out_width, depth]`.

Computes the grayscale dilation of 4-D input and 3-D filter tensors.

The input tensor has shape `[batch, in_height, in_width, depth]` and the filter tensor has shape `[filter_height, filter_width, depth]`, i.e., each input channel is processed independently of the others with its own structuring function. The output tensor has shape `[batch, out_height, out_width, depth]`. The spatial dimensions of the output tensor depend on the padding algorithm. We currently only support the default NHWC data_format.

In detail, the grayscale morphological 2-D dilation is the max-sum correlation (for consistency with conv2d, we use unmirrored filters):

output[b, y, x, c] = max_{dy, dx} input[b, strides[1] * y + rates[1] * dy, strides[2] * x + rates[2] * dx, c] + filter[dy, dx, c]

Max-pooling is a special case when the filter has size equal to the pooling kernel size and contains all zeros.

Note on duality: The dilation of input by the filter is equal to the negation of the erosion of `-input` by the reflected filter.

rsqrtGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the gradient for the rsqrt of x wrt its input.

Specifically, `grad = dy * -0.5 * y^3`, where `y = rsqrt(x)`, and dy is the corresponding input gradient.

rsqrt Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes reciprocal of square root of x element-wise.

I.e., $y = 1 / sqrt{x}$.

quantizedMaxPool Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Word16, Word8]` t)
=> Tensor v1 t	input: The 4D (batch x rows x cols x depth) Tensor to MaxReduce over.
-> Tensor v2 Float	min_input: The float value that the lowest quantized input value represents.
-> Tensor v3 Float	max_input: The float value that the highest quantized input value represents.
-> (Tensor Value t, Tensor Value Float, Tensor Value Float)	(output, min_output, max_output) output min_output: The float value that the lowest quantized output value represents. max_output: The float value that the highest quantized output value represents.

Produces the max pool of the input tensor for quantized types.

sqrt Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes square root of x element-wise.

I.e., $y = sqrt{x} = x^{1/2}$.

identityReader Source

Arguments

:: Build (Tensor Ref ByteString)

reader_handle: The handle to reference the Reader.

A Reader that outputs the queued work as both the key and value.

To use, enqueue strings in a Queue. ReaderRead will take the front work string and output (work, work).

square Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes square of x element-wise.

I.e., $y = x * x = x^2$.

quantizedReshape Source

Arguments

:: (TensorType t, TensorType tshape, OneOf `[Int32, Int64]` tshape)
=> Tensor v1 t	tensor
-> Tensor v2 tshape	shape: Defines the shape of the output tensor.
-> Tensor v3 Float	input_min: The minimum value of the input.
-> Tensor v4 Float	input_max: The maximum value of the input.
-> (Tensor Value t, Tensor Value Float, Tensor Value Float)	(output, output_min, output_max) output output_min: This value is copied from input_min. output_max: This value is copied from input_max.

Reshapes a quantized tensor as per the Reshape op.

```

reciprocalGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the gradient for the inverse of x wrt its input.

Specifically, `grad = -dy * y*y`, where `y = 1/x`, and dy is the corresponding input gradient.

invGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the gradient for the inverse of x wrt its input.

Specifically, `grad = -dy * y*y`, where `y = 1/x`, and dy is the corresponding input gradient.

inv Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the reciprocal of x element-wise.

I.e., $y = 1 / x$.

tensorArrayConcatV2 Source

Arguments

:: TensorType dtype
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> (Tensor Value dtype, Tensor Value Int64)	(value, lengths) value: All of the elements in the TensorArray, concatenated along the first axis. lengths: A vector of the row sizes of the original T elements in the value output. In the example above, this would be the values: `(n1, n2, ..., n(T-1))`.

Concat the elements from the TensorArray into value value.

Takes T elements of shapes

``` (n0 x d0 x d1 x ...), (n1 x d0 x d1 x ...), ..., (n(T-1) x d0 x d1 x ...) ```

and concatenates them into a Tensor of shape:

```(n0 + n1 + ... + n(T-1) x d0 x d1 x ...)```

All elements must have the same shape (excepting the first dimension).

complexAbs Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float]` t, TensorType tout, OneOf `[Double, Float]` tout)
=> Tensor v1 t	x
-> Tensor Value tout	y

Computes the complex absolute value of a tensor.

Given a tensor x of complex numbers, this operation returns a tensor of type float or double that is the absolute value of each element in x. All elements in x must be complex numbers of the form $a + bj$. The absolute value is computed as $ sqrt{a^2 + b^2}$.

For example:

``` # tensor x is [[-2.25 + 4.75j], [-3.25 + 5.75j]] tf.complex_abs(x) ==> [5.25594902, 6.60492229] ```

_HostCast Source

Arguments

:: (TensorType srcT, TensorType dstT)
=> Tensor v1 srcT	x
-> Tensor Value dstT	y

Cast x of type SrcT to y of DstT.

_HostCast requires its input and produces its output in host memory.

resizeNearestNeighbor Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	images: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 Int32	size: = A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The new size for the images.
-> Tensor Value t	resized_images: 4-D with shape `[batch, new_height, new_width, channels]`.

Resize images to size using nearest neighbor interpolation.

adjustContrast Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word8, Double, Float]` t)
=> Tensor v1 t	images
-> Tensor v2 Float	contrast_factor
-> Tensor v3 Float	min_value
-> Tensor v4 Float	max_value
-> Tensor Value Float	output

Deprecated. Disallowed in GraphDef version >= 2.

batchMatrixDiagPart Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor Value t	diagonal

batchMatrixSetDiag Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor v2 t	diagonal
-> Tensor Value t	output

batchMatrixDiag Source

Arguments

:: TensorType t
=> Tensor v1 t	diagonal
-> Tensor Value t	output

fakeQuantWithMinMaxVarsPerChannelGradient Source

Arguments

:: Tensor v1 Float	gradients: Backpropagated gradients above the FakeQuantWithMinMaxVars operation, shape one of: `[d]`, `[b, d]`, `[b, h, w, d]`.
-> Tensor v2 Float	inputs: Values passed as inputs to the FakeQuantWithMinMaxVars operation, shape same as `gradients`. min, max: Quantization interval, floats of shape `[d]`.
-> Tensor v3 Float	min
-> Tensor v4 Float	max
-> (Tensor Value Float, Tensor Value Float, Tensor Value Float)	(backprops_wrt_input, backprop_wrt_min, backprop_wrt_max) backprops_wrt_input: Backpropagated gradients w.r.t. inputs, shape same as `inputs`: `gradients * (inputs >= min && inputs <= max)`. backprop_wrt_min: Backpropagated gradients w.r.t. min parameter, shape `[d]`: `sum_per_d(gradients * (inputs < min))`. backprop_wrt_max: Backpropagated gradients w.r.t. max parameter, shape `[d]`: `sum_per_d(gradients * (inputs > max))`.

Compute gradients for a FakeQuantWithMinMaxVarsPerChannel operation.

sparseSegmentSqrtNGrad Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	grad: gradient propagated to the SparseSegmentSqrtN op.
-> Tensor v2 tidx	indices: indices passed to the corresponding SparseSegmentSqrtN op.
-> Tensor v3 Int32	segment_ids: segment_ids passed to the corresponding SparseSegmentSqrtN op.
-> Tensor v4 Int32	output_dim0: dimension 0 of "data" passed to SparseSegmentSqrtN op.
-> Tensor Value t	output

Computes gradients for SparseSegmentSqrtN.

Returns tensor "output" with same shape as grad, except for dimension 0 whose value is output_dim0.

fakeQuantWithMinMaxVarsPerChannel Source

Arguments

:: Tensor v1 Float	inputs
-> Tensor v2 Float	min
-> Tensor v3 Float	max
-> Tensor Value Float	outputs

Fake-quantize the inputs tensor of type float and one of the shapes: `[d]`,

`[b, d]` `[b, h, w, d]` via per-channel floats min and max of shape `[d]` to outputs tensor of same shape as inputs.

min; max: is the clamping range for the inputs data in the corresponding depth channel. Op divides this range into 255 steps (total of 256 values), then replaces each inputs value with the closest of the quantized step values.

This operation has a gradient and thus allows for training min and max values.

scalarSummary Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 ByteString	tags: Tags for the summary.
-> Tensor v2 t	values: Same shape as `tags. Values for the summary.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Outputs a Summary protocol buffer with scalar values.

The input tags and values must have the same shape. The generated summary has a summary value for each tag-value pair in tags and values.

neg Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes numerical negative value element-wise.

I.e., $y = -x$.

fakeQuantWithMinMaxArgsGradient Source

Arguments

:: Tensor v1 Float	gradients: Backpropagated gradients above the FakeQuantWithMinMaxArgs operation.
-> Tensor v2 Float	inputs: Values passed as inputs to the FakeQuantWithMinMaxArgs operation.
-> Tensor Value Float	backprops: Backpropagated gradients below the FakeQuantWithMinMaxArgs operation: `gradients * (inputs >= min && inputs <= max)`.

Compute gradients for a FakeQuantWithMinMaxArgs operation.

debugNanCount Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Input tensor, non-Reference type.
-> Tensor Value Int64	output: An integer output tensor that is the number of NaNs in the input.

Debug NaN Value Counter Op

Counts number of NaNs in the input tensor, for debugging.

debugIdentity Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Input tensor, non-Reference type.
-> Tensor Value t	output: Output tensor that equals the input tensor.

Debug Identity Op.

Provides an identity mapping of the non-Ref type input tensor for debugging.

bitcast Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType type', OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` type')
=> Tensor v1 t	input
-> Tensor Value type'	output

Bitcasts a tensor from one type to another without copying data.

Given a tensor input, this operation returns a tensor that has the same buffer data as input with datatype `type`.

If the input datatype T is larger than the output datatype `type` then the shape changes from [...] to [..., sizeof(T)/sizeof(`type`)].

If T is smaller than `type`, the operator requires that the rightmost dimension be equal to sizeof(`type`)/sizeof(T). The shape then goes from [..., sizeof(`type`)/sizeof(T)] to [...].

NOTE*: Bitcast is implemented as a low-level cast, so machines with different endian orderings will give different results.

sigmoid Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes sigmoid of x element-wise.

Specifically, `y = 1 / (1 + exp(-x))`.

copy Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Input tensor.
-> Tensor Value t	output: Output tensor, deep-copied from input.

Copy Op.

Performs CPU-to-CPU or GPU-to-GPU deep-copying of tensor, depending on the device on which the tensor is allocated.

Unlike the CopyHost Op, this op does not have HostMemory constraint on its input or output.

fixedUnigramCandidateSampler Source

Arguments

:: Int64	num_sampled: Number of candidates to randomly sample per batch.
-> Int64	num_true: Number of true labels per context.
-> Int64	range_max: The sampler will sample integers from the interval [0, range_max).
-> Bool	unique: If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
-> Tensor v1 Int64	true_classes: A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.
-> (Tensor Value Int64, Tensor Value Float, Tensor Value Float)	(sampled_candidates, true_expected_count, sampled_expected_count) sampled_candidates: A vector of length num_sampled, in which each element is the ID of a sampled candidate. true_expected_count: A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability. sampled_expected_count: A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

Generates labels for candidate sampling with a learned unigram distribution.

A unigram sampler could use a fixed unigram distribution read from a file or passed in as an in-memory array instead of building up the distribution from data on the fly. There is also an option to skew the distribution by applying a distortion power to the weights.

The vocabulary file should be in CSV-like format, with the last field being the weight associated with the word.

For each batch, this op picks a single set of sampled candidate labels.

The advantages of sampling candidates per-batch are simplicity and the possibility of efficient dense matrix multiplication. The disadvantage is that the sampled candidates must be chosen independently of the context and of the true labels.

listDiff Source

Arguments

:: (TensorType t, TensorType out_idx, OneOf `[Int32, Int64]` out_idx)
=> Tensor v1 t	x: 1-D. Values to keep.
-> Tensor v2 t	y: 1-D. Values to remove.
-> (Tensor Value t, Tensor Value out_idx)	(out, idx) out: 1-D. Values present in `x` but not in `y`. idx: 1-D. Positions of `x` values preserved in `out`.

Computes the difference between two lists of numbers or strings.

Given a list x and a list y, this operation returns a list out that represents all values that are in x but not in y. The returned list out is sorted in the same order that the numbers appear in x (duplicates are preserved). This operation also returns a list idx that represents the position of each out element in x. In other words:

`out[i] = x[idx[i]] for i in [0, 1, ..., len(out) - 1]`

For example, given this input:

```prettyprint x = [1, 2, 3, 4, 5, 6] y = [1, 3, 5] ```

This operation would return:

```prettyprint out ==> [2, 4, 6] idx ==> [1, 3, 5] ```

extractImagePatches Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	images: 4-D Tensor with shape `[batch, in_rows, in_cols, depth]`.
-> Tensor Value t	patches: 4-D Tensor with shape `[batch, out_rows, out_cols, ksize_rows * ksize_cols * depth]` containing image patches with size `ksize_rows x ksize_cols x depth` vectorized in the "depth" dimension.

Extract patches from images and put them in the "depth" output dimension.

spaceToDepth Source

Arguments

:: TensorType t
=> Int64	block_size: The size of the spatial block.
-> Tensor v1 t	input
-> Tensor Value t	output

SpaceToDepth for tensors of type T.

Rearranges blocks of spatial data, into depth. More specifically, this op outputs a copy of the input tensor where values from the height and width dimensions are moved to the depth dimension. The attr block_size indicates the input block size and how the data is moved.

Non-overlapping blocks of size `block_size x block size` are rearranged into depth at each location.
The depth of the output tensor is `input_depth * block_size * block_size`.
The input tensor's height and width must be divisible by block_size.

That is, assuming the input is in the shape: `[batch, height, width, depth]`, the shape of the output will be: `[batch, heightblock_size, widthblock_size, depth*block_size*block_size]`

This operation requires that the input tensor be of rank 4, and that block_size be >=1 and a divisor of both the input height and width.

This operation is useful for resizing the activations between convolutions (but keeping all data), e.g. instead of pooling. It is also useful for training purely convolutional models.

For example, given this input of shape `[1, 2, 2, 1]`, and block_size of 2:

```prettyprint x = [[[[1], [2]], [[3], [4]]]] ```

This operation will output a tensor of shape `[1, 1, 1, 4]`:

```prettyprint [[[[1, 2, 3, 4]]]] ```

Here, the input has a batch of 1 and each batch element has shape `[2, 2, 1]`, the corresponding output will have a single element (i.e. width and height are both 1) and will have a depth of 4 channels (1 * block_size * block_size). The output element shape is `[1, 1, 4]`.

For an input tensor with larger depth, here of shape `[1, 2, 2, 3]`, e.g.

```prettyprint x = [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]] ```

This operation, for block_size of 2, will return the following tensor of shape `[1, 1, 1, 12]`

```prettyprint [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]] ```

Similarly, for the following input of shape `[1 4 4 1]`, and a block size of 2:

```prettyprint x = [[[[1], [2], [5], [6]], [[3], [4], [7], [8]], [[9], [10], [13], [14]], [[11], [12], [15], [16]]]] ```

the operator will return the following tensor of shape `[1 2 2 4]`:

```prettyprint x = [[[[1, 2, 3, 4], [5, 6, 7, 8]], [[9, 10, 11, 12], [13, 14, 15, 16]]]] ```

cropAndResizeGradBoxes Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Float	grads: A 4-D tensor of shape `[num_boxes, crop_height, crop_width, depth]`.
-> Tensor v2 t	image: A 4-D tensor of shape `[batch, image_height, image_width, depth]`. Both `image_height` and `image_width` need to be positive.
-> Tensor v3 Float	boxes: A 2-D tensor of shape `[num_boxes, 4]`. The `i`-th row of the tensor specifies the coordinates of a box in the `box_ind[i]` image and is specified in normalized coordinates `[y1, x1, y2, x2]`. A normalized coordinate value of `y` is mapped to the image coordinate at `y * (image_height - 1)`, so as the `[0, 1]` interval of normalized image height is mapped to `[0, image_height - 1] in image height coordinates. We do allow y1 > y2, in which case the sampled crop is an up-down flipped version of the original image. The width dimension is treated similarly. Normalized coordinates outside the `[0, 1]` range are allowed, in which case we use `extrapolation_value` to extrapolate the input image values.
-> Tensor v4 Int32	box_ind: A 1-D tensor of shape `[num_boxes]` with int32 values in `[0, batch)`. The value of `box_ind[i]` specifies the image that the `i`-th box refers to.
-> Tensor Value Float	output: A 2-D tensor of shape `[num_boxes, 4]`.

Computes the gradient of the crop_and_resize op wrt the input boxes tensor.

batchToSpaceND Source

Arguments

:: (TensorType t, TensorType tblock_shape, OneOf `[Int32, Int64]` tblock_shape, TensorType tcrops, OneOf `[Int32, Int64]` tcrops)
=> Tensor v1 t	input: N-D with shape `input_shape = [batch] + spatial_shape + remaining_shape`, where spatial_shape has M dimensions.
-> Tensor v2 tblock_shape	block_shape: 1-D with shape `[M]`, all values must be >= 1.
-> Tensor v3 tcrops	crops: 2-D with shape `[M, 2]`, all values must be >= 0. `crops[i] = [crop_start, crop_end]` specifies the amount to crop from input dimension `i + 1`, which corresponds to spatial dimension `i`. It is required that `crop_start[i] + crop_end[i] <= block_shape[i] * input_shape[i + 1]`. This operation is equivalent to the following steps: Reshape `input` to `reshaped` of shape: [block_shape[0], ..., block_shape[M-1], batch / prod(block_shape), input_shape[1], ..., input_shape[N-1]] Permute dimensions of `reshaped` to produce `permuted` of shape [batch / prod(block_shape), input_shape[1], block_shape[0], ..., input_shape[M], block_shape[M-1], input_shape[M+1], ..., input_shape[N-1]] Reshape `permuted` to produce `reshaped_permuted` of shape [batch / prod(block_shape), input_shape[1] * block_shape[0], ..., input_shape[M] * block_shape[M-1], input_shape[M+1], ..., input_shape[N-1]] Crop the start and end of dimensions `[1, ..., M]` of `reshaped_permuted` according to `crops` to produce the output of shape: [batch / prod(block_shape), input_shape[1] * block_shape[0] - crops[0,0] - crops[0,1], ..., input_shape[M] * block_shape[M-1] - crops[M-1,0] - crops[M-1,1], input_shape[M+1], ..., input_shape[N-1]] Some examples: For the following input of shape `[4, 1, 1, 1]`, `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`: ```prettyprint [[[[1]]], [[[2]]], [[[3]]], [[[4]]]] ``` The output tensor has shape `[1, 2, 2, 1]` and value: ```prettyprint x = [[[[1], [2]], [[3], [4]]]] ``` For the following input of shape `[4, 1, 1, 3]`, `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`: ```prettyprint [[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]] ``` The output tensor has shape `[1, 2, 2, 3]` and value: ```prettyprint x = [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]] ``` For the following input of shape `[4, 2, 2, 1]`, `block_shape = [2, 2]`, and `crops = [[0, 0], [0, 0]]`: ```prettyprint x = [[[[1], [3]], [[5], [7]]], [[[2], [4]], [[10], [12]]], [[[5], [7]], [[13], [15]]], [[[6], [8]], [[14], [16]]]] ``` The output tensor has shape `[1, 4, 4, 1]` and value: ```prettyprint x = [[[1], [2], [3], [4]], [[5], [6], [7], [8]], [[9], [10], [11], [12]], [[13], [14], [15], [16]]] ``` For the following input of shape `[8, 1, 3, 1]`, `block_shape = [2, 2]`, and `crops = [[0, 0], [2, 0]]`: ```prettyprint x = [[[[0], [1], [3]]], [[[0], [9], [11]]], [[[0], [2], [4]]], [[[0], [10], [12]]], [[[0], [5], [7]]], [[[0], [13], [15]]], [[[0], [6], [8]]], [[[0], [14], [16]]]] ``` The output tensor has shape `[2, 2, 4, 1]` and value: ```prettyprint x = [[[[1], [2], [3], [4]], [[5], [6], [7], [8]]], [[[9], [10], [11], [12]], [[13], [14], [15], [16]]]] ```
-> Tensor Value t	output

BatchToSpace for N-D tensors of type T.

This operation reshapes the "batch" dimension 0 into `M + 1` dimensions of shape `block_shape + [batch]`, interleaves these blocks back into the grid defined by the spatial dimensions `[1, ..., M]`, to obtain a result with the same rank as the input. The spatial dimensions of this intermediate result are then optionally cropped according to crops to produce the output. This is the reverse of SpaceToBatch. See below for a precise description.

spaceToBatch Source

Arguments

:: (TensorType t, TensorType tpaddings, OneOf `[Int32, Int64]` tpaddings)
=> Int64	block_size
-> Tensor v1 t	input: 4-D with shape `[batch, height, width, depth]`.
-> Tensor v2 tpaddings	paddings: 2-D tensor of non-negative integers with shape `[2, 2]`. It specifies the padding of the input with zeros across the spatial dimensions as follows: paddings = [[pad_top, pad_bottom], [pad_left, pad_right]] The effective spatial dimensions of the zero-padded input tensor will be: height_pad = pad_top + height + pad_bottom width_pad = pad_left + width + pad_right The attr `block_size` must be greater than one. It indicates the block size. Non-overlapping blocks of size `block_size x block size` in the height and width dimensions are rearranged into the batch dimension at each location. The batch of the output tensor is `batch * block_size * block_size`. Both height_pad and width_pad must be divisible by block_size. The shape of the output will be: [batchblock_sizeblock_size, height_padblock_size, width_padblock_size, depth] Some examples: For the following input of shape `[1, 2, 2, 1]` and block_size of 2: ```prettyprint x = [[[[1], [2]], [[3], [4]]]] ``` The output tensor has shape `[4, 1, 1, 1]` and value: ```prettyprint [[[[1]]], [[[2]]], [[[3]]], [[[4]]]] ``` For the following input of shape `[1, 2, 2, 3]` and block_size of 2: ```prettyprint x = [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]] ``` The output tensor has shape `[4, 1, 1, 3]` and value: ```prettyprint [[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]] ``` For the following input of shape `[1, 4, 4, 1]` and block_size of 2: ```prettyprint x = [[[[1], [2], [3], [4]], [[5], [6], [7], [8]], [[9], [10], [11], [12]], [[13], [14], [15], [16]]]] ``` The output tensor has shape `[4, 2, 2, 1]` and value: ```prettyprint x = [[[[1], [3]], [[5], [7]]], [[[2], [4]], [[10], [12]]], [[[5], [7]], [[13], [15]]], [[[6], [8]], [[14], [16]]]] ``` For the following input of shape `[2, 2, 4, 1]` and block_size of 2: ```prettyprint x = [[[[1], [2], [3], [4]], [[5], [6], [7], [8]]], [[[9], [10], [11], [12]], [[13], [14], [15], [16]]]] ``` The output tensor has shape `[8, 1, 2, 1]` and value: ```prettyprint x = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]], [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]] ``` Among others, this operation is useful for reducing atrous convolution into regular convolution.
-> Tensor Value t	output

SpaceToBatch for 4-D tensors of type T.

This is a legacy version of the more general SpaceToBatchND.

Zero-pads and then rearranges (permutes) blocks of spatial data into batch. More specifically, this op outputs a copy of the input tensor where values from the height and width dimensions are moved to the batch dimension. After the zero-padding, both height and width of the input must be divisible by the block size.

adjustHue Source

Arguments

:: Tensor v1 Float	images: Images to adjust. At least 3-D.
-> Tensor v2 Float	delta: A float delta to add to the hue.
-> Tensor Value Float	output: The hue-adjusted image or images.

Adjust the hue of one or more images.

images is a tensor of at least 3 dimensions. The last dimension is interpretted as channels, and must be three.

The input image is considered in the RGB colorspace. Conceptually, the RGB colors are first mapped into HSV. A delta is then applied all the hue values, and then remapped back to RGB colorspace.

spaceToBatchND Source

Arguments

:: (TensorType t, TensorType tblock_shape, OneOf `[Int32, Int64]` tblock_shape, TensorType tpaddings, OneOf `[Int32, Int64]` tpaddings)
=> Tensor v1 t	input: N-D with shape `input_shape = [batch] + spatial_shape + remaining_shape`, where spatial_shape has `M` dimensions.
-> Tensor v2 tblock_shape	block_shape: 1-D with shape `[M]`, all values must be >= 1.
-> Tensor v3 tpaddings	paddings: 2-D with shape `[M, 2]`, all values must be >= 0. `paddings[i] = [pad_start, pad_end]` specifies the padding for input dimension `i + 1`, which corresponds to spatial dimension `i`. It is required that `block_shape[i]` divides `input_shape[i + 1] + pad_start + pad_end`. This operation is equivalent to the following steps: Zero-pad the start and end of dimensions `[1, ..., M]` of the input according to `paddings` to produce `padded` of shape `padded_shape`. Reshape `padded` to `reshaped_padded` of shape: batch + padded_shape[1 / block_shape[0], block_shape[0], ..., padded_shape[M] / block_shape[M-1], block_shape[M-1]] + remaining_shape Permute dimensions of `reshaped_padded` to produce `permuted_reshaped_padded` of shape: block_shape + [batch] + [padded_shape[1] / block_shape[0], ..., padded_shape[M] / block_shape[M-1]] + remaining_shape Reshape `permuted_reshaped_padded` to flatten `block_shape` into the batch dimension, producing an output tensor of shape: batch * prod(block_shape) + padded_shape[1 / block_shape[0], ..., padded_shape[M] / block_shape[M-1]] + remaining_shape Some examples: For the following input of shape `[1, 2, 2, 1]`, `block_shape = [2, 2]`, and `paddings = [[0, 0], [0, 0]]`: ```prettyprint x = [[[[1], [2]], [[3], [4]]]] ``` The output tensor has shape `[4, 1, 1, 1]` and value: ```prettyprint [[[[1]]], [[[2]]], [[[3]]], [[[4]]]] ``` For the following input of shape `[1, 2, 2, 3]`, `block_shape = [2, 2]`, and `paddings = [[0, 0], [0, 0]]`: ```prettyprint x = [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]] ``` The output tensor has shape `[4, 1, 1, 3]` and value: ```prettyprint [[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]] ``` For the following input of shape `[1, 4, 4, 1]`, `block_shape = [2, 2]`, and `paddings = [[0, 0], [0, 0]]`: ```prettyprint x = [[[[1], [2], [3], [4]], [[5], [6], [7], [8]], [[9], [10], [11], [12]], [[13], [14], [15], [16]]]] ``` The output tensor has shape `[4, 2, 2, 1]` and value: ```prettyprint x = [[[[1], [3]], [[5], [7]]], [[[2], [4]], [[10], [12]]], [[[5], [7]], [[13], [15]]], [[[6], [8]], [[14], [16]]]] ``` For the following input of shape `[2, 2, 4, 1]`, block_shape = `[2, 2]`, and paddings = `[[0, 0], [2, 0]]`: ```prettyprint x = [[[[1], [2], [3], [4]], [[5], [6], [7], [8]]], [[[9], [10], [11], [12]], [[13], [14], [15], [16]]]] ``` The output tensor has shape `[8, 1, 3, 1]` and value: ```prettyprint x = [[[[0], [1], [3]]], [[[0], [9], [11]]], [[[0], [2], [4]]], [[[0], [10], [12]]], [[[0], [5], [7]]], [[[0], [13], [15]]], [[[0], [6], [8]]], [[[0], [14], [16]]]] ``` Among others, this operation is useful for reducing atrous convolution into regular convolution.
-> Tensor Value t	output

SpaceToBatch for N-D tensors of type T.

This operation divides "spatial" dimensions `[1, ..., M]` of the input into a grid of blocks of shape block_shape, and interleaves these blocks with the "batch" dimension (0) such that in the output, the spatial dimensions `[1, ..., M]` correspond to the position within the grid, and the batch dimension combines both the position within a spatial block and the original batch position. Prior to division into blocks, the spatial dimensions of the input are optionally zero padded according to paddings. See below for a precise description.

diagPart Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Double, Float]` t)
=> Tensor v1 t	input: Rank k tensor where k is 2, 4, or 6.
-> Tensor Value t	diagonal: The extracted diagonal.

Returns the diagonal part of the tensor.

This operation returns a tensor with the diagonal part of the input. The diagonal part is computed as follows:

Assume input has dimensions `[D1,..., Dk, D1,..., Dk]`, then the output is a tensor of rank k with dimensions `[D1,..., Dk]` where:

`diagonal[i1,..., ik] = input[i1, ..., ik, i1,..., ik]`.

For example:

```prettyprint # input is [[1, 0, 0, 0] [0, 2, 0, 0] [0, 0, 3, 0] [0, 0, 0, 4]]

tf.diag_part(input) ==> [1, 2, 3, 4] ```

placeholderV2 Source

Arguments

:: TensorType dtype
=> Shape	shape: The shape of the tensor. The shape can be any partially-specified shape. To be unconstrained, pass in a shape with unknown rank.
-> Tensor Value dtype	output: A placeholder tensor that must be replaced using the feed mechanism.

A placeholder op for a value that will be fed into the computation.

N.B. This operation will fail with an error if it is executed. It is intended as a way to represent a value that will always be fed, and to provide attrs that enable the fed value to be checked at runtime.

acos Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes acos of x element-wise.

placeholder Source

Arguments

:: TensorType dtype
=> Tensor Value dtype	output: A placeholder tensor that must be replaced using the feed mechanism.

A placeholder op for a value that will be fed into the computation.

N.B. This operation will fail with an error if it is executed. It is intended as a way to represent a value that will always be fed, and to provide attrs that enable the fed value to be checked at runtime.

controlTrigger :: ControlNode Source

Does nothing. Serves as a control trigger for scheduling.

Only useful as a placeholder for control edges.

atan Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes atan of x element-wise.

mirrorPad Source

Arguments

:: (TensorType t, TensorType tpaddings, OneOf `[Int32, Int64]` tpaddings)
=> Tensor v1 t	input: The input tensor to be padded.
-> Tensor v2 tpaddings	paddings: A two-column matrix specifying the padding sizes. The number of rows must be the same as the rank of `input`.
-> Tensor Value t	output: The padded tensor.

Pads a tensor with mirrored values.

This operation pads a input with mirrored values according to the paddings you specify. paddings is an integer tensor with shape `[n, 2]`, where n is the rank of input. For each dimension D of input, `paddings[D, 0]` indicates how many values to add before the contents of input in that dimension, and `paddings[D, 1]` indicates how many values to add after the contents of input in that dimension. Both `paddings[D, 0]` and `paddings[D, 1]` must be no greater than `input.dim_size(D)` (or `input.dim_size(D) - 1`) if copy_border is true (if false, respectively).

The padded size of each dimension D of the output is:

`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`

For example:

```prettyprint # t is [[1, 2, 3], [4, 5, 6]]. # paddings is [[1, 1]], [2, 2]]. # mode is SYMMETRIC. # rank of t is 2. pad(t, paddings) ==> [[2, 1, 1, 2, 3, 3, 2] [2, 1, 1, 2, 3, 3, 2] [5, 4, 4, 5, 6, 6, 5] [5, 4, 4, 5, 6, 6, 5]] ```

where' Source

Arguments

:: Tensor v1 Bool	input
-> Tensor Value Int64	index

Returns locations of true values in a boolean tensor.

This operation returns the coordinates of true elements in input. The coordinates are returned in a 2-D tensor where the first dimension (rows) represents the number of true elements, and the second dimension (columns) represents the coordinates of the true elements. Keep in mind, the shape of the output tensor can vary depending on how many true values there are in input. Indices are output in row-major order.

For example:

```prettyprint # input tensor is [[True, False] # [True, False]] # input has two true values, so output has two coordinates. # input has rank of 2, so coordinates have two indices. where(input) ==> [[0, 0], [1, 0]]

# input tensor is [[[True, False] # [True, False]] # [[False, True] # [False, True]] # [[False, False] # [False, True]]] # input has 5 true values, so output has 5 coordinates. # input has rank of 3, so coordinates have three indices. where(input) ==> [[0, 0, 0], [0, 1, 0], [1, 0, 1], [1, 1, 1], [2, 1, 1]] ```

avgPool3DGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int32	orig_input_shape: The original input dimensions.
-> Tensor v2 t	grad: Output backprop of shape `[batch, depth, rows, cols, channels]`.
-> Tensor Value t	output: The backprop for input.

Computes gradients of average pooling function.

readerReset Source

Arguments

:: Tensor Ref ByteString	reader_handle: Handle to a Reader.
-> Build ControlNode

Restore a Reader to its initial clean state.

tileGrad Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor v2 Int32	multiples
-> Tensor Value t	output

Returns the gradient of Tile.

Since Tile takes an input and repeats the input multiples times along each dimension, TileGrad takes in multiples and aggregates each repeated tile of input into output.

expandDims Source

Arguments

:: (TensorType t, TensorType tdim, OneOf `[Int32, Int64]` tdim)
=> Tensor v1 t	input
-> Tensor v2 tdim	dim: 0-D (scalar). Specifies the dimension index at which to expand the shape of `input`.
-> Tensor Value t	output: Contains the same data as `input`, but its shape has an additional dimension of size 1 added.

Inserts a dimension of 1 into a tensor's shape.

Given a tensor input, this operation inserts a dimension of 1 at the dimension index dim of input's shape. The dimension index dim starts at zero; if you specify a negative number for dim it is counted backward from the end.

This operation is useful if you want to add a batch dimension to a single element. For example, if you have a single image of shape `[height, width, channels]`, you can make it a batch of 1 image with `expand_dims(image, 0)`, which will make the shape `[1, height, width, channels]`.

Other examples:

```prettyprint # t is a tensor of shape [2] shape(expand_dims(t, 0)) ==> [1, 2] shape(expand_dims(t, 1)) ==> [2, 1] shape(expand_dims(t, -1)) ==> [2, 1]

# t2 is a tensor of shape [2, 3, 5] shape(expand_dims(t2, 0)) ==> [1, 2, 3, 5] shape(expand_dims(t2, 2)) ==> [2, 3, 1, 5] shape(expand_dims(t2, 3)) ==> [2, 3, 5, 1] ```

This operation requires that:

`-1-input.dims() <= dim <= input.dims()`

This operation is related to `squeeze()`, which removes dimensions of size 1.

tensorSummary Source

Arguments

:: TensorType t
=> Tensor v1 t	tensor: A tensor to serialize.
-> Tensor Value ByteString	summary

Outputs a Summary protocol buffer with a tensor.

tile Source

Arguments

:: (TensorType t, TensorType tmultiples, OneOf `[Int32, Int64]` tmultiples)
=> Tensor v1 t	input: 1-D or higher.
-> Tensor v2 tmultiples	multiples: 1-D. Length must be the same as the number of dimensions in `input`
-> Tensor Value t	output

Constructs a tensor by tiling a given tensor.

This operation creates a new tensor by replicating input multiples times. The output tensor's i'th dimension has `input.dims(i) * multiples[i]` elements, and the values of input are replicated `multiples[i]` times along the ith dimension. For example, tiling `[a b c d]` by `[2]` produces `[a b c d a b c d]`.

stridedSlice Source

Arguments

:: (TensorType t, TensorType index, OneOf `[Int32, Int64]` index)
=> Tensor v1 t	input
-> Tensor v2 index	begin: `begin[k]` specifies the offset into the `k`th range specification. The exact dimension this corresponds to will be determined by context. Out-of-bounds values will be silently clamped. If the `k`th bit of `begin_mask` then `begin[k]` is ignored and the full range of the appropriate dimension is used instead. Negative values causes indexing to start from the highest element e.g. If `foo==[1,2,3]` then `foo[-1]==3`.
-> Tensor v3 index	end: `end[i]` is like `begin` with the exception that `end_mask` is used to determine full ranges.
-> Tensor v4 index	strides: `strides[i]` specifies the increment in the `i`th specification after extracting a given element. Negative indices will reverse the original order. Out or range values are clamped to `[0,dim[i]) if slice[i]>0` or `[-1,dim[i]-1] if slice[i] < 0`
-> Tensor Value t	output

Return a strided slice from input.

Note, most python users will want to use the Python __getitem__ or __getitem__ rather than this op directly.

The goal of this op is to produce a new tensor with a subset of the elements from the n dimensional input tensor. The subset is chosen using a sequence of m sparse range specifications encoded into the arguments of this function. Note, in some cases m could be equal to n, but this need not be the case. Each range specification entry can be one of the following:

An ellipsis (...). Ellipses are used to imply zero or more dimensions of full-dimension selection and are produced using ellipsis_mask. For example, `foo[...]` is the identity slice.
A new axis. This is used to insert a new shape=1 dimension and is produced using new_axis_mask. For example, `foo[:, ...]` where foo is shape `(3, 4)` produces a `(1, 3, 4)` tensor.
A range `begin:end:stride`. This is used to specify how much to choose from a given dimension. stride can be any integer but 0. begin is an integer which represents the index of the first value to select while end represents the index of the last value to select. The number of values selected in each dimension is `end - begin` if `stride > 0` and `begin - end` if `stride < 0`. begin and end can be negative where `-1` is the last element, `-2` is the second to last. begin_mask controls whether to replace the explicitly given begin with an implicit effective value of `0` if `stride > 0` and `-1` if `stride < 0`. end_mask is analogous but produces the number required to create the largest open interval. For example, given a shape `(3,)` tensor `foo[:]`, the effective begin and end are `0` and `3`. Do not assume this is equivalent to `foo[0:-1]` which has an effective begin and end of `0` and `2`. Another example is `foo[-2::-1]` which reverses the first dimension of a tensor while dropping the last two (in the original order elements). For example `foo = [1,2,3,4]; foo[-2::-1]` is `[4,3]`.
A single index. This is used to keep only elements that have a given index. For example (`foo[2, :]` on a shape `(5,6)` tensor produces a shape `(6,)` tensor. This is encoded in begin and end and shrink_axis_mask.

Each conceptual range specification is encoded in the op's argument. This encoding is best understand by considering a non-trivial example. In particular, `foo[1, 2:4, None, ..., :-3:-1, :]` will be encoded as

```prettyprint begin = [1, 2, x, x, 0, x] # x denotes don't care (usually 0) end = [2, 4, x, x, -3, x] strides = [1, 1, x, x, -1, 1] begin_mask = 1<<4 | 1 << 5 = 48 end_mask = 1<<5 = 32 ellipsis_mask = 1<<3 = 8 new_axis_mask = 1<<2 4 shrink_axis_mask = 1<<0 ```

In this case if `foo.shape` is (5, 5, 5, 5, 5, 5) the final shape of the slice becomes (2, 1, 5, 5, 2, 5). Let us walk step by step through each argument specification.

The first argument in the example slice is turned into `begin = 1` and `end = begin + 1 = 2`. To disambiguate from the original spec `2:4` we also set the appropriate bit in shrink_axis_mask.
`2:4` is contributes 2, 4, 1 to begin, end, and stride. All masks have zero bits contributed.
None is a synonym for `tf.newaxis`. This means insert a dimension of size 1 dimension in the final shape. Dummy values are contributed to begin, end and stride, while the new_axis_mask bit is set.
... grab the full ranges from as many dimensions as needed to fully specify a slice for every dimension of the input shape.
`:-3:-1` shows the use of negative indices. A negative index i associated with a dimension that has shape s is converted to a positive index `s + i`. So `-1` becomes `s-1` (i.e. the last element). This conversion is done internally so begin, end and strides receive x, -3, and -1. The appropriate begin_mask bit is set to indicate the start range is the full range (ignoring the x).
: indicates that the entire contents of the corresponding dimension is selected. This is equivalent to `::` or `0::1`. begin, end, and strides receive 0, 0, and 1, respectively. The appropriate bits in begin_mask and end_mask are also set.

Requirements*: `0 != strides[i] for i in [0, m)` `ellipsis_mask must be a power of two (only one ellipsis)`

slice Source

Arguments

:: (TensorType t, TensorType index, OneOf `[Int32, Int64]` index)
=> Tensor v1 t	input
-> Tensor v2 index	begin: begin[i] specifies the offset into the `i`th dimension of `input` to slice from.
-> Tensor v3 index	size: size[i] specifies the number of elements of the `i`th dimension of `input` to slice. If size[i] is -1, all remaining elements in dimension i are included in the slice (i.e. this is equivalent to setting size[i] = input.dim_size(i) - begin[i]).
-> Tensor Value t	output

Return a slice from input.

The output tensor is a tensor with dimensions described by size whose values are extracted from input starting at the offsets in begin.

Requirements*: 0 <= begin[i] <= begin[i] + size[i] <= Di for i in [0, n)

quantizedConv2D Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType tfilter, OneOf `[Int16, Int32, Word16, Word8]` tfilter, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	input
-> Tensor v2 tfilter	filter: filter's input_depth dimension must match input's depth dimensions.
-> Tensor v3 Float	min_input: The float value that the lowest quantized input value represents.
-> Tensor v4 Float	max_input: The float value that the highest quantized input value represents.
-> Tensor v5 Float	min_filter: The float value that the lowest quantized filter value represents.
-> Tensor v6 Float	max_filter: The float value that the highest quantized filter value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(output, min_output, max_output) output min_output: The float value that the lowest quantized output value represents. max_output: The float value that the highest quantized output value represents.

Computes a 2D convolution given quantized 4D input and filter tensors.

The inputs are quantized tensors where the lowest value represents the real number of the associated minimum, and the highest represents the maximum. This means that you can only interpret the quantized output in the same way, by taking the returned minimum and maximum values into account.

relu6Grad Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	gradients: The backpropagated gradients to the corresponding Relu6 operation.
-> Tensor v2 t	features: The features passed as input to the corresponding Relu6 operation.
-> Tensor Value t	backprops: The gradients: `gradients * (features > 0) * (features < 6)`.

Computes rectified linear 6 gradients for a Relu6 operation.

avgPoolGrad Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 Int32	orig_input_shape: 1-D. Shape of the original input to `avg_pool`.
-> Tensor v2 t	grad: 4-D with shape `[batch, height, width, channels]`. Gradients w.r.t. the output of `avg_pool`.
-> Tensor Value t	output: 4-D. Gradients w.r.t. the input of `avg_pool`.

Computes gradients of the average pooling function.

stringSplit Source

Arguments

:: Tensor v1 ByteString

input: 1-D. Strings to split.

-> Tensor v2 ByteString

delimiter: 0-D. Delimiter character, or empty string.

-> (Tensor Value Int64, Tensor Value ByteString, Tensor Value Int64)

(indices, values, shape)

indices: A dense matrix of int64 representing the indices of the sparse tensor.
values: A vector of strings corresponding to the splited values.
shape: a length-2 vector of int64 representing the shape of the sparse tensor, where the first value is N and the second value is the maximum number of tokens in a single input entry.

Split elements of input based on delimiter into a SparseTensor.

Let N be the size of source (typically N will be the batch size). Split each element of input based on delimiter and return a SparseTensor containing the splitted tokens. Empty tokens are ignored.

delimiter can be empty or a single-byte character. If delimiter is an empty string, each element of input is split into individual single-byte character strings, including splitting of UTF-8 multibyte sequences.

For example: N = 2, input[0] is 'hello world' and input[1] is 'a b c', then the output will be

indices = [0, 0; 0, 1; 1, 0; 1, 1; 1, 2] shape = [2, 3] values = [hello, world, a, b, c]

rank Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor Value Int32	output

Returns the rank of a tensor.

This operation returns an integer representing the rank of input.

For example:

```prettyprint # t is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]] # shape of tensor t is [2, 2, 3] rank(t) ==> 3 ```

*Note**: The rank of a tensor is not the same as the rank of a matrix. The rank of a tensor is the number of indices required to uniquely select each element of the tensor. Rank is also known as "order", "degree", or "ndims."

reciprocal Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the reciprocal of x element-wise.

I.e., $y = 1 / x$.

reverseSequence Source

Arguments

:: (TensorType t, TensorType tlen, OneOf `[Int32, Int64]` tlen)
=> Int64	seq_dim: The dimension which is partially reversed.
-> Tensor v1 t	input: The input to reverse.
-> Tensor v2 tlen	seq_lengths: 1-D with length `input.dims(batch_dim)` and `max(seq_lengths) < input.dims(seq_dim)`
-> Tensor Value t	output: The partially reversed input. It has the same shape as `input`.

Reverses variable length slices.

This op first slices input along the dimension batch_dim, and for each slice i, reverses the first `seq_lengths[i]` elements along the dimension seq_dim.

The elements of seq_lengths must obey `seq_lengths[i] < input.dims[seq_dim]`, and seq_lengths must be a vector of length `input.dims[batch_dim]`.

The output slice i along dimension batch_dim is then given by input slice i, with the first `seq_lengths[i]` slices along dimension seq_dim reversed.

For example:

```prettyprint # Given this: batch_dim = 0 seq_dim = 1 input.dims = (4, 8, ...) seq_lengths = [7, 2, 3, 5]

# then slices of input are reversed on seq_dim, but only up to seq_lengths: output[0, 0:7, :, ...] = input[0, 7:0:-1, :, ...] output[1, 0:2, :, ...] = input[1, 2:0:-1, :, ...] output[2, 0:3, :, ...] = input[2, 3:0:-1, :, ...] output[3, 0:5, :, ...] = input[3, 5:0:-1, :, ...]

# while entries past seq_lens are copied through: output[0, 7:, :, ...] = input[0, 7:, :, ...] output[1, 2:, :, ...] = input[1, 2:, :, ...] output[2, 3:, :, ...] = input[2, 3:, :, ...] output[3, 2:, :, ...] = input[3, 2:, :, ...] ```

In contrast, if:

```prettyprint # Given this: batch_dim = 2 seq_dim = 0 input.dims = (8, ?, 4, ...) seq_lengths = [7, 2, 3, 5]

# then slices of input are reversed on seq_dim, but only up to seq_lengths: output[0:7, :, 0, :, ...] = input[7:0:-1, :, 0, :, ...] output[0:2, :, 1, :, ...] = input[2:0:-1, :, 1, :, ...] output[0:3, :, 2, :, ...] = input[3:0:-1, :, 2, :, ...] output[0:5, :, 3, :, ...] = input[5:0:-1, :, 3, :, ...]

# while entries past seq_lens are copied through: output[7:, :, 0, :, ...] = input[7:, :, 0, :, ...] output[2:, :, 1, :, ...] = input[2:, :, 1, :, ...] output[3:, :, 2, :, ...] = input[3:, :, 2, :, ...] output[2:, :, 3, :, ...] = input[2:, :, 3, :, ...] ```

biasAddGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	out_backprop: Any number of dimensions.
-> Tensor Value t	output: 1-D with size the feature dimension of `out_backprop`.

The backward operation for BiasAdd on the "bias" tensor.

It accumulates all the values from out_backprop into the feature dimension. For NHWC data format, the feature dimension is the last. For NCHW data format, the feature dimension is the third-to-last.

addSparseToTensorsMap Source

Arguments

:: TensorType t
=> Tensor v1 Int64	sparse_indices: 2-D. The `indices` of the `SparseTensor`.
-> Tensor v2 t	sparse_values: 1-D. The `values` of the `SparseTensor`.
-> Tensor v3 Int64	sparse_shape: 1-D. The `shape` of the `SparseTensor`.
-> Build (Tensor Value Int64)	sparse_handle: 0-D. The handle of the `SparseTensor` now stored in the `SparseTensorsMap`.

Add a SparseTensor to a SparseTensorsMap return its handle.

A SparseTensor is represented by three tensors: sparse_indices, sparse_values, and sparse_shape.

This operator takes the given SparseTensor and adds it to a container object (a SparseTensorsMap). A unique key within this container is generated in the form of an int64, and this is the value that is returned.

The SparseTensor can then be read out as part of a minibatch by passing the key as a vector element to TakeManySparseFromTensorsMap. To ensure the correct SparseTensorsMap is accessed, ensure that the same container and shared_name are passed to that Op. If no shared_name is provided here, instead use the *name* of the Operation created by calling AddSparseToTensorsMap as the shared_name passed to TakeManySparseFromTensorsMap. Ensure the Operations are colocated.

tan Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes tan of x element-wise.

sparseReduceSumSparse Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int64	input_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, possibly not in canonical ordering.
-> Tensor v2 t	input_values: 1-D. `N` non-empty values corresponding to `input_indices`.
-> Tensor v3 Int64	input_shape: 1-D. Shape of the input SparseTensor.
-> Tensor v4 Int32	reduction_axes: 1-D. Length-`K` vector containing the reduction axes.
-> (Tensor Value Int64, Tensor Value t, Tensor Value Int64)	(output_indices, output_values, output_shape) output_indices output_values output_shape

Computes the sum of elements across dimensions of a SparseTensor.

This Op takes a SparseTensor and is the sparse counterpart to `tf.reduce_sum()`. In contrast to SparseReduceSum, this Op returns a SparseTensor.

Reduces sp_input along the dimensions given in reduction_axes. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_axes. If keep_dims is true, the reduced dimensions are retained with length 1.

If reduction_axes has no entries, all dimensions are reduced, and a tensor with a single element is returned. Additionally, the axes can be negative, which are interpreted according to the indexing rules in Python.

shapeN Source

Arguments

:: (TensorType t, TensorType out_type, OneOf `[Int32, Int64]` out_type)
=> [Tensor v1 t]	input
-> [Tensor Value out_type]	output

Returns shape of tensors.

This operation returns N 1-D integer tensors representing shape of `input[i]s`.

shape Source

Arguments

:: (TensorType t, TensorType out_type, OneOf `[Int32, Int64]` out_type)
=> Tensor v1 t	input
-> Tensor Value out_type	output

Returns the shape of a tensor.

This operation returns a 1-D integer tensor representing the shape of input.

For example:

```prettyprint # t is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]] shape(t) ==> [2, 2, 3] ```

unique Source

Arguments

:: (TensorType t, TensorType out_idx, OneOf `[Int32, Int64]` out_idx)

=> Tensor v1 t

x: 1-D.

-> (Tensor Value t, Tensor Value out_idx)

(y, idx)

y: 1-D.
idx: 1-D.

Finds unique elements in a 1-D tensor.

This operation returns a tensor y containing all of the unique elements of x sorted in the same order that they occur in x. This operation also returns a tensor idx the same size as x that contains the index of each value of x in the unique output y. In other words:

`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`

For example:

```prettyprint # tensor x is [1, 1, 2, 4, 4, 4, 7, 8, 8] y, idx = unique(x) y ==> [1, 2, 4, 7, 8] idx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4] ```

truncatedNormal Source

Arguments

:: (TensorType dtype, OneOf `[Word16, Double, Float]` dtype, TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	shape: The shape of the output tensor.
-> Build (Tensor Value dtype)	output: A tensor of the specified shape filled with random truncated normal values.

Outputs random values from a truncated normal distribution.

The generated values follow a normal distribution with mean 0 and standard deviation 1, except that values whose magnitude is more than 2 standard deviations from the mean are dropped and re-picked.

invertPermutation Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	x: 1-D.
-> Tensor Value t	y: 1-D.

Computes the inverse permutation of a tensor.

This operation computes the inverse of an index permutation. It takes a 1-D integer tensor x, which represents the indices of a zero-based array, and swaps each value with its index position. In other words, for an output tensor y and an input tensor x, this operation computes the following:

`y[x[i]] = i for i in [0, 1, ..., len(x) - 1]`

The values must include 0. There can be no duplicate values or negative values.

For example:

```prettyprint # tensor x is [3, 4, 0, 2, 1] invert_permutation(x) ==> [2, 4, 3, 0, 1] ```

checkNumerics Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	tensor
-> Tensor Value t	output

Checks a tensor for NaN and Inf values.

When run, reports an InvalidArgument error if tensor has any values that are not a number (NaN) or infinity (Inf). Otherwise, passes tensor as-is.

uniformCandidateSampler Source

Arguments

:: Int64	num_sampled: Number of candidates to randomly sample per batch.
-> Int64	num_true: Number of true labels per context.
-> Int64	range_max: The sampler will sample integers from the interval [0, range_max).
-> Bool	unique: If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
-> Tensor v1 Int64	true_classes: A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.
-> (Tensor Value Int64, Tensor Value Float, Tensor Value Float)	(sampled_candidates, true_expected_count, sampled_expected_count) sampled_candidates: A vector of length num_sampled, in which each element is the ID of a sampled candidate. true_expected_count: A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability. sampled_expected_count: A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

Generates labels for candidate sampling with a uniform distribution.

See explanations of candidate sampling and the data formats at go/candidate-sampling.

For each batch, this op picks a single set of sampled candidate labels.

The advantages of sampling candidates per-batch are simplicity and the possibility of efficient dense matrix multiplication. The disadvantage is that the sampled candidates must be chosen independently of the context and of the true labels.

gather Source

Arguments

:: (TensorType tparams, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 tparams	params
-> Tensor v2 tindices	indices
-> Tensor Value tparams	output

Gather slices from params according to indices.

indices must be an integer tensor of any dimension (usually 0-D or 1-D). Produces an output tensor with shape `indices.shape + params.shape[1:]` where:

```python # Scalar indices output[:, ..., :] = params[indices, :, ... :]

# Vector indices output[i, :, ..., :] = params[indices[i], :, ... :]

# Higher rank indices output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :] ```

If indices is a permutation and `len(indices) == params.shape[0]` then this operation will permute params accordingly.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/Gather.png" alt /div

const Source

Arguments

:: TensorType dtype
=> Tensor Value dtype	output

Returns a constant tensor.

fill Source

Arguments

:: TensorType t
=> Tensor v1 Int32	dims: 1-D. Represents the shape of the output tensor.
-> Tensor v2 t	value: 0-D (scalar). Value to fill the returned tensor. `compatibility(numpy) Equivalent to np.full`end_compatibility
-> Tensor Value t	output

Creates a tensor filled with a scalar value.

This operation creates a tensor of shape dims and fills it with value.

For example:

```prettyprint # Output tensor has shape [2, 3]. fill([2, 3], 9) ==> [[9, 9, 9] [9, 9, 9]] ```

editDistance Source

Arguments

:: TensorType t
=> Tensor v1 Int64	hypothesis_indices: The indices of the hypothesis list SparseTensor. This is an N x R int64 matrix.
-> Tensor v2 t	hypothesis_values: The values of the hypothesis list SparseTensor. This is an N-length vector.
-> Tensor v3 Int64	hypothesis_shape: The shape of the hypothesis list SparseTensor. This is an R-length vector.
-> Tensor v4 Int64	truth_indices: The indices of the truth list SparseTensor. This is an M x R int64 matrix.
-> Tensor v5 t	truth_values: The values of the truth list SparseTensor. This is an M-length vector.
-> Tensor v6 Int64	truth_shape: truth indices, vector.
-> Tensor Value Float	output: A dense float tensor with rank R - 1. For the example input: // hypothesis represents a 2x1 matrix with variable-length values: // (0,0) = ["a"] // (1,0) = ["b"] hypothesis_indices = [[0, 0, 0], [1, 0, 0]] hypothesis_values = ["a", "b"] hypothesis_shape = [2, 1, 1] // truth represents a 2x2 matrix with variable-length values: // (0,0) = [] // (0,1) = ["a"] // (1,0) = ["b", "c"] // (1,1) = ["a"] truth_indices = [[0, 1, 0], [1, 0, 0], [1, 0, 1], [1, 1, 0]] truth_values = ["a", "b", "c", "a"] truth_shape = [2, 2, 2] normalize = true The output will be: // output is a 2x2 matrix with edit distances normalized by truth lengths. output = [[inf, 1.0], // (0,0): no truth, (0,1): no hypothesis [0.5, 1.0]] // (1,0): addition, (1,1): no hypothesis

Computes the (possibly normalized) Levenshtein Edit Distance.

The inputs are variable-length sequences provided by SparseTensors (hypothesis_indices, hypothesis_values, hypothesis_shape) and (truth_indices, truth_values, truth_shape).

The inputs are:

reverse Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Bool, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	tensor: Up to 8-D.
-> Tensor v2 Bool	dims: 1-D. The dimensions to reverse.
-> Tensor Value t	output: The same shape as `tensor`.

Reverses specific dimensions of a tensor.

Given a tensor, and a bool tensor dims representing the dimensions of tensor, this operation reverses each dimension i of tensor where `dims[i]` is True.

tensor can have up to 8 dimensions. The number of dimensions of tensor must equal the number of elements in dims. In other words:

`rank(tensor) = size(dims)`

For example:

```prettyprint # tensor t is [[[[ 0, 1, 2, 3], # [ 4, 5, 6, 7], # [ 8, 9, 10, 11]], # [[12, 13, 14, 15], # [16, 17, 18, 19], # [20, 21, 22, 23]]]] # tensor t shape is [1, 2, 3, 4]

# dims is [False, False, False, True] reverse(t, dims) ==> [[[[ 3, 2, 1, 0], [ 7, 6, 5, 4], [ 11, 10, 9, 8]], [[15, 14, 13, 12], [19, 18, 17, 16], [23, 22, 21, 20]]]]

# dims is [False, True, False, False] reverse(t, dims) ==> [[[[12, 13, 14, 15], [16, 17, 18, 19], [20, 21, 22, 23] [[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11]]]]

# dims is [False, False, True, False] reverse(t, dims) ==> [[[[8, 9, 10, 11], [4, 5, 6, 7], [0, 1, 2, 3]] [[20, 21, 22, 23], [16, 17, 18, 19], [12, 13, 14, 15]]]] ```

matrixSetDiag Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Rank `k+1`, where `k >= 1`.
-> Tensor v2 t	diagonal: Rank `k`, where `k >= 1`.
-> Tensor Value t	output: Rank `k+1`, with `output.shape = input.shape`.

Returns a batched matrix tensor with new batched diagonal values.

Given input and diagonal, this operation returns a tensor with the same shape and values as input, except for the main diagonal of the innermost matrices. These will be overwritten by the values in diagonal.

The output is computed as follows:

Assume input has `k+1` dimensions `[I, J, K, ..., M, N]` and diagonal has k dimensions `[I, J, K, ..., min(M, N)]`. Then the output is a tensor of rank `k+1` with dimensions `[I, J, K, ..., M, N]` where:

`output[i, j, k, ..., m, n] = diagonal[i, j, k, ..., n]` for `m == n`.
`output[i, j, k, ..., m, n] = input[i, j, k, ..., m, n]` for `m != n`.

matrixDiag Source

Arguments

:: TensorType t
=> Tensor v1 t	diagonal: Rank `k`, where `k >= 1`.
-> Tensor Value t	output: Rank `k+1`, with `output.shape = diagonal.shape + [diagonal.shape[-1]]`.

Returns a batched diagonal tensor with a given batched diagonal values.

Given a diagonal, this operation returns a tensor with the diagonal and everything else padded with zeros. The diagonal is computed as follows:

Assume diagonal has k dimensions `[I, J, K, ..., N]`, then the output is a tensor of rank `k+1` with dimensions [I, J, K, ..., N, N]` where:

`output[i, j, k, ..., m, n] = 1{m=n} * diagonal[i, j, k, ..., n]`.

For example:

```prettyprint # diagonal is [[1, 2, 3, 4], [5, 6, 7, 8]]

and diagonal.shape = (2, 4)

tf.matrix_diag(diagonal) ==> [[[1, 0, 0, 0] [0, 2, 0, 0] [0, 0, 3, 0] [0, 0, 0, 4]], [[5, 0, 0, 0] [0, 6, 0, 0] [0, 0, 7, 0] [0, 0, 0, 8]]]

which has shape (2, 4, 4) ```

diag Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Double, Float]` t)
=> Tensor v1 t	diagonal: Rank k tensor where k is at most 3.
-> Tensor Value t	output

Returns a diagonal tensor with a given diagonal values.

Given a diagonal, this operation returns a tensor with the diagonal and everything else padded with zeros. The diagonal is computed as follows:

Assume diagonal has dimensions [D1,..., Dk], then the output is a tensor of rank 2k with dimensions [D1,..., Dk, D1,..., Dk] where:

`output[i1,..., ik, i1,..., ik] = diagonal[i1, ..., ik]` and 0 everywhere else.

For example:

```prettyprint # diagonal is [1, 2, 3, 4] tf.diag(diagonal) ==> [[1, 0, 0, 0] [0, 2, 0, 0] [0, 0, 3, 0] [0, 0, 0, 4]] ```

immutableConst Source

Arguments

:: TensorType dtype
=> Shape	shape: Shape of the returned tensor.
-> Tensor Value dtype	tensor

Returns immutable tensor from memory region.

The current implementation memmaps the tensor from a file.

concat Source

Arguments

:: TensorType t
=> Tensor v1 Int32	concat_dim: 0-D. The dimension along which to concatenate. Must be in the range [0, rank(values)).
-> [Tensor v2 t]	values: The `N` Tensors to concatenate. Their ranks and types must match, and their sizes must match in all dimensions except `concat_dim`.
-> Tensor Value t	output: A `Tensor` with the concatenation of values stacked along the `concat_dim` dimension. This tensor's shape matches that of `values` except in `concat_dim` where it has the sum of the sizes.

Concatenates tensors along one dimension.

unpack Source

Arguments

:: TensorType t
=> Int64	num
-> Tensor v1 t	value: 1-D or higher, with `axis` dimension size equal to `num`.
-> [Tensor Value t]	output: The list of tensors unpacked from `value`.

Unpacks a given dimension of a rank-R tensor into num rank-`(R-1)` tensors.

Unpacks num tensors from value by chipping it along the axis dimension. For example, given a tensor of shape `(A, B, C, D)`;

If `axis == 0` then the i'th tensor in output is the slice `value[i, :, :, :]` and each tensor in output will have shape `(B, C, D)`. (Note that the dimension unpacked along is gone, unlike split).

If `axis == 1` then the i'th tensor in output is the slice `value[:, i, :, :]` and each tensor in output will have shape `(A, C, D)`. Etc.

This is the opposite of pack.

fact Source

Arguments

:: Tensor Value ByteString

fact

Output a fact about factorials.

abs Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the absolute value of a tensor.

Given a tensor x, this operation returns a tensor containing the absolute value of each element in x. For example, if x is an input element and y is an output element, this operation computes $y = |x|$.

softmax Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	logits: 2-D with shape `[batch_size, num_classes]`.
-> Tensor Value t	softmax: Same shape as `logits`.

Computes softmax activations.

For each batch i and class j we have

softmax[i, j] = exp(logits[i, j]) / sum_j(exp(logits[i, j]))

reverseV2 Source

Arguments

:: (TensorType tidx, OneOf `[Int32, Int64]` tidx, TensorType t, OneOf `[Complex Double, Complex Float, Bool, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	tensor: Up to 8-D.
-> Tensor v2 tidx	axis: 1-D. The indices of the dimensions to reverse.
-> Tensor Value t	output: The same shape as `tensor`.

Reverses specific dimensions of a tensor.

Given a tensor, and a int32 tensor axis representing the set of dimensions of tensor to reverse. This operation reverses each dimension i for which there exists j s.t. `axis[j] == i`.

tensor can have up to 8 dimensions. The number of dimensions specified in axis may be 0 or more entries. If an index is specified more than once, a InvalidArgument error is raised.

For example:

```prettyprint # tensor t is [[[[ 0, 1, 2, 3], # [ 4, 5, 6, 7], # [ 8, 9, 10, 11]], # [[12, 13, 14, 15], # [16, 17, 18, 19], # [20, 21, 22, 23]]]] # tensor t shape is [1, 2, 3, 4]

# dims is [3] or dims is -1 reverse(t, dims) ==> [[[[ 3, 2, 1, 0], [ 7, 6, 5, 4], [ 11, 10, 9, 8]], [[15, 14, 13, 12], [19, 18, 17, 16], [23, 22, 21, 20]]]]

# dims is '[1]' (or dims is '[-3]') reverse(t, dims) ==> [[[[12, 13, 14, 15], [16, 17, 18, 19], [20, 21, 22, 23] [[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11]]]]

# dims is '[2]' (or dims is '[-2]') reverse(t, dims) ==> [[[[8, 9, 10, 11], [4, 5, 6, 7], [0, 1, 2, 3]] [[20, 21, 22, 23], [16, 17, 18, 19], [12, 13, 14, 15]]]] ```

identity Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor Value t	output

Return a tensor with the same shape and contents as the input tensor or value.

sparseAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType treal, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` treal)
=> Tensor v1 Int64	a_indices: 2-D. The `indices` of the first `SparseTensor`, size `[nnz, ndims]` Matrix.
-> Tensor v2 t	a_values: 1-D. The `values` of the first `SparseTensor`, size `[nnz]` Vector.
-> Tensor v3 Int64	a_shape: 1-D. The `shape` of the first `SparseTensor`, size `[ndims]` Vector.
-> Tensor v4 Int64	b_indices: 2-D. The `indices` of the second `SparseTensor`, size `[nnz, ndims]` Matrix.
-> Tensor v5 t	b_values: 1-D. The `values` of the second `SparseTensor`, size `[nnz]` Vector.
-> Tensor v6 Int64	b_shape: 1-D. The `shape` of the second `SparseTensor`, size `[ndims]` Vector.
-> Tensor v7 treal	thresh: 0-D. The magnitude threshold that determines if an output value/index pair takes space.
-> (Tensor Value Int64, Tensor Value t, Tensor Value Int64)	(sum_indices, sum_values, sum_shape) sum_indices sum_values sum_shape

Adds two SparseTensor objects to produce another SparseTensor.

The input SparseTensor objects' indices are assumed ordered in standard lexicographic order. If this is not the case, before this step run SparseReorder to restore index ordering.

By default, if two values sum to zero at some index, the output SparseTensor would still include that particular location in its index, storing a zero in the corresponding value slot. To override this, callers can specify thresh, indicating that if the sum has a magnitude strictly smaller than thresh, its corresponding value and index would then not be included. In particular, `thresh == 0` (default) means everything is kept and actual thresholding happens only for a positive value.

In the following shapes, nnz is the count after taking thresh into account.

sparseApplyCenteredRMSProp Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	mg: Should be from a Variable().
-> Tensor Ref t	ms: Should be from a Variable().
-> Tensor Ref t	mom: Should be from a Variable().
-> Tensor v5 t	lr: Scaling factor. Must be a scalar.
-> Tensor v6 t	rho: Decay rate. Must be a scalar.
-> Tensor v7 t	momentum
-> Tensor v8 t	epsilon: Ridge term. Must be a scalar.
-> Tensor v9 t	grad: The gradient.
-> Tensor v10 tindices	indices: A vector of indices into the first dimension of var, ms and mom.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the centered RMSProp algorithm.

The centered RMSProp algorithm uses an estimate of the centered second moment (i.e., the variance) for normalization, as opposed to regular RMSProp, which uses the (uncentered) second moment. This often helps with training, but is slightly more expensive in terms of computation and memory.

Note that in dense implementation of this algorithm, mg, ms, and mom will update even if the grad is zero, but in this sparse implementation, mg, ms, and mom will not update in iterations during which the grad is zero.

mean_square = decay * mean_square + (1-decay) * gradient ** 2 mean_grad = decay * mean_grad + (1-decay) * gradient Delta = learning_rate * gradient / sqrt(mean_square + epsilon - mean_grad ** 2)

ms <- rho * ms_{t-1} + (1-rho) * grad * grad mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon) var <- var - mom

addN Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> [Tensor v1 t]	inputs: Must all be the same size and shape.
-> Tensor Value t	sum

Add all input tensors element wise.

concatOffset Source

Arguments

:: Tensor v1 Int32

concat_dim: The dimension along which to concatenate.

-> [Tensor v2 Int32]

shape: The N int32 vectors representing shape of tensors being concatenated.

-> [Tensor Value Int32]

offset: The N int32 vectors representing the starting offset of input tensors within the concatenated output.

This is typically used by gradient computations for a concat operation.

Computes offsets of concat inputs within its output.

For example:

```prettyprint # x is [2, 2, 7] # y is [2, 3, 7] # z is [2, 5, 7] concat_offset(2, [x, y, z]) => [0, 0, 0], [0, 2, 0], [0, 5, 0] ```

concatV2 Source

Arguments

:: (TensorType t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> [Tensor v1 t]	values: List of `N` Tensors to concatenate. Their ranks and types must match, and their sizes must match in all dimensions except `concat_dim`.
-> Tensor v2 tidx	axis: 0-D. The dimension along which to concatenate. Must be in the range [0, rank(values)).
-> Tensor Value t	output: A `Tensor` with the concatenation of values stacked along the `concat_dim` dimension. This tensor's shape matches that of `values` except in `concat_dim` where it has the sum of the sizes.

Concatenates tensors along one dimension.

zerosLike Source

Arguments

:: TensorType t
=> Tensor v1 t	x: a tensor of type T.
-> Tensor Value t	y: a tensor of the same shape and type as x but filled with zeros.

Returns a tensor of zeros with the same shape and type as x.

applyCenteredRMSProp Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	mg: Should be from a Variable().
-> Tensor Ref t	ms: Should be from a Variable().
-> Tensor Ref t	mom: Should be from a Variable().
-> Tensor v5 t	lr: Scaling factor. Must be a scalar.
-> Tensor v6 t	rho: Decay rate. Must be a scalar.
-> Tensor v7 t	momentum
-> Tensor v8 t	epsilon: Ridge term. Must be a scalar.
-> Tensor v9 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the centered RMSProp algorithm.

The centered RMSProp algorithm uses an estimate of the centered second moment (i.e., the variance) for normalization, as opposed to regular RMSProp, which uses the (uncentered) second moment. This often helps with training, but is slightly more expensive in terms of computation and memory.

Note that in dense implementation of this algorithm, mg, ms, and mom will update even if the grad is zero, but in this sparse implementation, mg, ms, and mom will not update in iterations during which the grad is zero.

mean_square = decay * mean_square + (1-decay) * gradient ** 2 mean_grad = decay * mean_grad + (1-decay) * gradient

Delta = learning_rate * gradient / sqrt(mean_square + epsilon - mean_grad ** 2)

mg <- rho * mg_{t-1} + (1-rho) * grad ms <- rho * ms_{t-1} + (1-rho) * grad * grad mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms - mg * mg + epsilon) var <- var - mom

applyRMSProp Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	ms: Should be from a Variable().
-> Tensor Ref t	mom: Should be from a Variable().
-> Tensor v4 t	lr: Scaling factor. Must be a scalar.
-> Tensor v5 t	rho: Decay rate. Must be a scalar.
-> Tensor v6 t	momentum
-> Tensor v7 t	epsilon: Ridge term. Must be a scalar.
-> Tensor v8 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the RMSProp algorithm.

Note that in dense implementation of this algorithm, ms and mom will update even if the grad is zero, but in this sparse implementation, ms and mom will not update in iterations during which the grad is zero.

mean_square = decay * mean_square + (1-decay) * gradient ** 2 Delta = learning_rate * gradient / sqrt(mean_square + epsilon)

ms <- rho * ms_{t-1} + (1-rho) * grad * grad mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon) var <- var - mom

assignAddVariableOp Source

Arguments

:: TensorType dtype
=> ResourceHandle dtype	resource: handle to the resource in which to store the variable.
-> Tensor v2 dtype	value: the value by which the variable will be incremented.
-> Build ControlNode

Adds a value to the current value of a variable.

Any ReadVariableOp which depends directly or indirectly on this assign is guaranteed to see the incremented value or a subsequent newer one.

Outputs the incremented value, which can be used to totally order the increments to this variable.

applyAdam Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	m: Should be from a Variable().
-> Tensor Ref t	v: Should be from a Variable().
-> Tensor v4 t	beta1_power: Must be a scalar.
-> Tensor v5 t	beta2_power: Must be a scalar.
-> Tensor v6 t	lr: Scaling factor. Must be a scalar.
-> Tensor v7 t	beta1: Momentum factor. Must be a scalar.
-> Tensor v8 t	beta2: Momentum factor. Must be a scalar.
-> Tensor v9 t	epsilon: Ridge term. Must be a scalar.
-> Tensor v10 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the Adam algorithm.

lr_t <- learning_rate * sqrt(1 - beta2^t) / (1 - beta1^t) m_t <- beta1 * m_{t-1} + (1 - beta1) * g_t v_t <- beta2 * v_{t-1} + (1 - beta2) * g_t * g_t variable <- variable - lr_t * m_t / (sqrt(v_t) + epsilon)

extractGlimpse Source

Arguments

:: Tensor v1 Float	input: A 4-D float tensor of shape `[batch_size, height, width, channels]`.
-> Tensor v2 Int32	size: A 1-D tensor of 2 elements containing the size of the glimpses to extract. The glimpse height must be specified first, following by the glimpse width.
-> Tensor v3 Float	offsets: A 2-D integer tensor of shape `[batch_size, 2]` containing the x, y locations of the center of each window.
-> Tensor Value Float	glimpse: A tensor representing the glimpses `[batch_size, glimpse_height, glimpse_width, channels]`.

Extracts a glimpse from the input tensor.

Returns a set of windows called glimpses extracted at location offsets from the input tensor. If the windows only partially overlaps the inputs, the non overlapping areas will be filled with random noise.

The result is a 4-D tensor of shape `[batch_size, glimpse_height, glimpse_width, channels]`. The channels and batch dimensions are the same as that of the input tensor. The height and width of the output windows are specified in the size parameter.

The argument normalized and centered controls how the windows are built:

If the coordinates are normalized but not centered, 0.0 and 1.0 correspond to the minimum and maximum of each height and width dimension.
If the coordinates are both normalized and centered, they range from
1.0 to 1.0. The coordinates (-1.0, -1.0) correspond to the upper left corner, the lower right corner is located at (1.0, 1.0) and the center is at (0, 0).
If the coordinates are not normalized they are interpreted as numbers of pixels.

sparseApplyMomentum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Learning rate. Must be a scalar.
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Tensor v6 t	momentum: Momentum. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update relevant entries in '*var' and '*accum' according to the momentum scheme.

Set use_nesterov = True if you want to use Nesterov momentum.

That is for rows we have grad for, we update var and accum as follows:

accum = accum * momentum + grad var -= lr * accum

applyMomentum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Scaling factor. Must be a scalar.
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 t	momentum: Momentum. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the momentum scheme. Set use_nesterov = True if you

want to use Nesterov momentum.

accum = accum * momentum + grad var -= lr * accum

fIFOQueue Source

Arguments

:: Build (Tensor Ref ByteString)

handle: The handle to the queue.

A queue that produces elements in first-in first-out order.

sparseApplyFtrl Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor Ref t	linear: Should be from a Variable().
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Tensor v6 t	lr: Scaling factor. Must be a scalar.
-> Tensor v7 t	l1: L1 regularization. Must be a scalar.
-> Tensor v8 t	l2: L2 regularization. Must be a scalar.
-> Tensor v9 t	lr_power: Scaling factor. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update relevant entries in '*var' according to the Ftrl-proximal scheme.

That is for rows we have grad for, we update var, accum and linear as follows: accum_new = accum + grad * grad linear += grad + (accum_new^(-lr_power) - accum^(-lr_power)) / lr * var quadratic = 1.0 / (accum_new^(lr_power) * lr) + 2 * l2 var = (sign(linear) * l1 - linear) / quadratic if |linear| > l1 else 0.0 accum = accum_new

sparseApplyAdagradDA Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	gradient_accumulator: Should be from a Variable().
-> Tensor Ref t	gradient_squared_accumulator: Should be from a Variable().
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Tensor v6 t	lr: Learning rate. Must be a scalar.
-> Tensor v7 t	l1: L1 regularization. Must be a scalar.
-> Tensor v8 t	l2: L2 regularization. Must be a scalar.
-> Tensor v9 Int64	global_step: Training step number. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update entries in '*var' and '*accum' according to the proximal adagrad scheme.

floorDiv Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x // y element-wise.

NOTE*: FloorDiv supports broadcasting. More about broadcasting here

applyAdagradDA Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	gradient_accumulator: Should be from a Variable().
-> Tensor Ref t	gradient_squared_accumulator: Should be from a Variable().
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 t	lr: Scaling factor. Must be a scalar.
-> Tensor v6 t	l1: L1 regularization. Must be a scalar.
-> Tensor v7 t	l2: L2 regularization. Must be a scalar.
-> Tensor v8 Int64	global_step: Training step number. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the proximal adagrad scheme.

applyAdagrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Scaling factor. Must be a scalar.
-> Tensor v4 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the adagrad scheme.

accum += grad * grad var -= lr * grad * (1 / sqrt(accum))

sigmoidGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the gradient of the sigmoid of x wrt its input.

Specifically, `grad = dy * y * (1 - y)`, where `y = sigmoid(x)`, and dy is the corresponding input gradient.

applyAdadelta Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor Ref t	accum_update: Should be from a Variable().
-> Tensor v4 t	lr: Scaling factor. Must be a scalar.
-> Tensor v5 t	rho: Decay factor. Must be a scalar.
-> Tensor v6 t	epsilon: Constant factor. Must be a scalar.
-> Tensor v7 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the adadelta scheme.

accum = rho() * accum + (1 - rho()) * grad.square(); update = (update_accum + epsilon).sqrt() * (accum + epsilon()).rsqrt() * grad; update_accum = rho() * update_accum + (1 - rho()) * update.square(); var -= update;

sparseApplyProximalGradientDescent Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor v2 t	alpha: Scaling factor. Must be a scalar.
-> Tensor v3 t	l1: L1 regularization. Must be a scalar.
-> Tensor v4 t	l2: L2 regularization. Must be a scalar.
-> Tensor v5 t	grad: The gradient.
-> Tensor v6 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Build (Tensor Ref t)	out: Same as "var".

Sparse update '*var' as FOBOS algorithm with fixed learning rate.

That is for rows we have grad for, we update var as follows: prox_v = var - alpha * grad var = sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0}

applyProximalGradientDescent Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor v2 t	alpha: Scaling factor. Must be a scalar.
-> Tensor v3 t	l1: L1 regularization. Must be a scalar.
-> Tensor v4 t	l2: L2 regularization. Must be a scalar.
-> Tensor v5 t	delta: The change.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' as FOBOS algorithm with fixed learning rate.

prox_v = var - alpha * delta var = sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0}

matrixSolve Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Double, Float]` t)
=> Tensor v1 t	matrix: Shape is `[..., M, M]`.
-> Tensor v2 t	rhs: Shape is `[..., M, K]`.
-> Tensor Value t	output: Shape is `[..., M, K]`.

Solves systems of linear equations.

Matrix is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices. Rhs is a tensor of shape `[..., M, K]`. The output is a tensor shape `[..., M, K]`. If adjoint is False then each output matrix satisfies `matrix[..., :, :] * output[..., :, :] = rhs[..., :, :]`. If adjoint is True then each output matrix satisfies `adjoint(matrix[..., :, :]) * output[..., :, :] = rhs[..., :, :]`.

sparseApplyProximalAdagrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Learning rate. Must be a scalar.
-> Tensor v4 t	l1: L1 regularization. Must be a scalar.
-> Tensor v5 t	l2: L2 regularization. Must be a scalar.
-> Tensor v6 t	grad: The gradient.
-> Tensor v7 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Build (Tensor Ref t)	out: Same as "var".

Sparse update entries in '*var' and '*accum' according to FOBOS algorithm.

That is for rows we have grad for, we update var and accum as follows: accum += grad * grad prox_v = var prox_v -= lr * grad * (1 / sqrt(accum)) var = sign(prox_v)/(1+lr*l2) * max{|prox_v|-lr*l1,0}

applyGradientDescent Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor v2 t	alpha: Scaling factor. Must be a scalar.
-> Tensor v3 t	delta: The change.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' by subtracting alpha * delta from it.

batchNormWithGlobalNormalization Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Bool	scale_after_normalization: A bool indicating whether the resulted tensor needs to be multiplied with gamma.
-> Float	variance_epsilon: A small float number to avoid dividing by 0.
-> Tensor v1 t	t: A 4D input Tensor.
-> Tensor v2 t	m: A 1D mean Tensor with size matching the last dimension of t. This is the first output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v3 t	v: A 1D variance Tensor with size matching the last dimension of t. This is the second output from tf.nn.moments, or a saved moving average thereof.
-> Tensor v4 t	beta: A 1D beta Tensor with size matching the last dimension of t. An offset to be added to the normalized tensor.
-> Tensor v5 t	gamma: A 1D gamma Tensor with size matching the last dimension of t. If "scale_after_normalization" is true, this tensor will be multiplied with the normalized tensor.
-> Tensor Value t	result

Batch normalization.

This op is deprecated. Prefer `tf.nn.batch_normalization`.

encodeBase64 Source

Arguments

:: Tensor v1 ByteString	input: Strings to be encoded.
-> Tensor Value ByteString	output: Input strings encoded in base64.

Encode strings into web-safe base64 format.

Refer to the following article for more information on base64 format: en.wikipedia.orgwikiBase64. Base64 strings may have padding with '=' at the end so that the encoded has length multiple of 4. See Padding section of the link above.

Web-safe means that the encoder uses - and _ instead of + and /.

stringJoin Source

Arguments

:: [Tensor v1 ByteString]	inputs: A list of string tensors. The tensors must all have the same shape, or be scalars. Scalars may be mixed in; these will be broadcast to the shape of non-scalar inputs.
-> Tensor Value ByteString	output

Joins the strings in the given list of string tensors into one tensor;

with the given separator (default is an empty separator).

cropAndResizeGradImage Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 Float	grads: A 4-D tensor of shape `[num_boxes, crop_height, crop_width, depth]`.
-> Tensor v2 Float	boxes: A 2-D tensor of shape `[num_boxes, 4]`. The `i`-th row of the tensor specifies the coordinates of a box in the `box_ind[i]` image and is specified in normalized coordinates `[y1, x1, y2, x2]`. A normalized coordinate value of `y` is mapped to the image coordinate at `y * (image_height - 1)`, so as the `[0, 1]` interval of normalized image height is mapped to `[0, image_height - 1] in image height coordinates. We do allow y1 > y2, in which case the sampled crop is an up-down flipped version of the original image. The width dimension is treated similarly. Normalized coordinates outside the `[0, 1]` range are allowed, in which case we use `extrapolation_value` to extrapolate the input image values.
-> Tensor v3 Int32	box_ind: A 1-D tensor of shape `[num_boxes]` with int32 values in `[0, batch)`. The value of `box_ind[i]` specifies the image that the `i`-th box refers to.
-> Tensor v4 Int32	image_size: A 1-D tensor with value `[batch, image_height, image_width, depth]` containing the original image size. Both `image_height` and `image_width` need to be positive.
-> Tensor Value t	output: A 4-D tensor of shape `[batch, image_height, image_width, depth]`.

Computes the gradient of the crop_and_resize op wrt the input image tensor.

tanh Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes hyperbolic tangent of x element-wise.

asString Source

Arguments

:: (TensorType t, OneOf `[Complex Float, Bool, Int32, Int64, Int8, Double, Float]` t)
=> Tensor v1 t	input
-> Tensor Value ByteString	output

Converts each entry in the given tensor to strings. Supports many numeric

types and boolean.

iFFT2D Source

Arguments

:: Tensor v1 (Complex Float)

input: A complex64 tensor.

-> Tensor Value (Complex Float)

output: A complex64 tensor of the same shape as input. The inner-most 2 dimensions of input are replaced with their inverse 2D Fourier Transform.

compatibility(numpy) Equivalent to np.ifft2end_compatibility

Compute the inverse 2-dimensional discrete Fourier Transform over the inner-most

2 dimensions of input.

sparseConcat Source

Arguments

:: TensorType t
=> Int64	concat_dim: Dimension to concatenate along. Must be in range [-rank, rank), where rank is the number of dimensions in each input `SparseTensor`.
-> [Tensor v1 Int64]	indices: 2-D. Indices of each input `SparseTensor`.
-> [Tensor v2 t]	values: 1-D. Non-empty values of each `SparseTensor`.
-> [Tensor v3 Int64]	shapes: 1-D. Shapes of each `SparseTensor`.
-> (Tensor Value Int64, Tensor Value t, Tensor Value Int64)	(output_indices, output_values, output_shape) output_indices: 2-D. Indices of the concatenated `SparseTensor`. output_values: 1-D. Non-empty values of the concatenated `SparseTensor`. output_shape: 1-D. Shape of the concatenated `SparseTensor`.

Concatenates a list of SparseTensor along the specified dimension.

Concatenation is with respect to the dense versions of these sparse tensors. It is assumed that each input is a SparseTensor whose elements are ordered along increasing dimension number.

All inputs' shapes must match, except for the concat dimension. The indices, values, and shapes lists must have the same length.

The output shape is identical to the inputs', except along the concat dimension, where it is the sum of the inputs' sizes along that dimension.

The output elements will be resorted to preserve the sort order along increasing dimension number.

This op runs in `O(M log M)` time, where M is the total number of non-empty values across all inputs. This is due to the need for an internal sort in order to concatenate efficiently across an arbitrary dimension.

For example, if `concat_dim = 1` and the inputs are

sp_inputs[0]: shape = [2, 3] [0, 2]: "a" [1, 0]: "b" [1, 1]: "c"

sp_inputs[1]: shape = [2, 4] [0, 1]: "d" [0, 2]: "e"

then the output will be

shape = [2, 7] [0, 2]: "a" [0, 4]: "d" [0, 5]: "e" [1, 0]: "b" [1, 1]: "c"

Graphically this is equivalent to doing

a: concat [ d e ] = [ a d e ]
b c: [ ] [b c ]

shardedFilespec Source

Arguments

:: Tensor v1 ByteString	basename
-> Tensor v2 Int32	num_shards
-> Tensor Value ByteString	filename

Generate a glob pattern matching all sharded file names.

transpose Source

Arguments

:: (TensorType t, TensorType tperm, OneOf `[Int32, Int64]` tperm)
=> Tensor v1 t	x
-> Tensor v2 tperm	perm
-> Tensor Value t	y

Shuffle dimensions of x according to a permutation.

The output y has the same rank as x. The shapes of x and y satisfy: `y.shape[i] == x.shape[perm[i]] for i in [0, 1, ..., rank(x) - 1]`

reduceJoin Source

Arguments

:: Tensor v1 ByteString	inputs: The input to be joined. All reduced indices must have non-zero size.
-> Tensor v2 Int32	reduction_indices: The dimensions to reduce over. Dimensions are reduced in the order specified. Omitting `reduction_indices` is equivalent to passing `[n-1, n-2, ..., 0]`. Negative indices from `-n` to `-1` are supported.
-> Tensor Value ByteString	output: Has shape equal to that of the input with reduced dimensions removed or set to `1` depending on `keep_dims`.

Joins a string Tensor across the given dimensions.

Computes the string join across dimensions in the given string Tensor of shape `[d_0, d_1, ..., d_n-1]`. Returns a new Tensor created by joining the input strings with the given separator (default: empty string). Negative indices are counted backwards from the end, with `-1` being equivalent to `n - 1`. Passing an empty reduction_indices joins all strings in linear index order and outputs a scalar string.

For example:

``` # tensor a is [["a", "b"], ["c", "d"]] tf.reduce_join(a, 0) ==> ["ac", "bd"] tf.reduce_join(a, 1) ==> ["ab", "cd"] tf.reduce_join(a, -2) = tf.reduce_join(a, 0) ==> ["ac", "bd"] tf.reduce_join(a, -1) = tf.reduce_join(a, 1) ==> ["ab", "cd"] tf.reduce_join(a, 0, keep_dims=True) ==> [["ac", "bd"]] tf.reduce_join(a, 1, keep_dims=True) ==> [["ab"], ["cd"]] tf.reduce_join(a, 0, separator=".") ==> ["a.c", "b.d"] tf.reduce_join(a, [0, 1]) ==> ["acbd"] tf.reduce_join(a, [1, 0]) ==> ["abcd"] tf.reduce_join(a, []) ==> ["abcd"] ```

stringToHashBucket Source

Arguments

:: Int64	num_buckets: The number of buckets.
-> Tensor v1 ByteString	string_tensor
-> Tensor Value Int64	output: A Tensor of the same shape as the input `string_tensor`.

Converts each string in the input Tensor to its hash mod by a number of buckets.

The hash function is deterministic on the content of the string within the process.

Note that the hash function may change from time to time. This functionality will be deprecated and it's recommended to use `tf.string_to_hash_bucket_fast()` or `tf.string_to_hash_bucket_strong()`.

multinomial Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	logits: 2-D Tensor with shape `[batch_size, num_classes]`. Each slice `[i, :]` represents the unnormalized log probabilities for all classes.
-> Tensor v2 Int32	num_samples: 0-D. Number of independent samples to draw for each row slice.
-> Build (Tensor Value Int64)	output: 2-D Tensor with shape `[batch_size, num_samples]`. Each slice `[i, :]` contains the drawn class labels with range `[0, num_classes)`.

Draws samples from a multinomial distribution.

stringToHashBucketStrong Source

Arguments

:: Int64	num_buckets: The number of buckets.
-> Tensor v1 ByteString	input: The strings to assign a hash bucket.
-> Tensor Value Int64	output: A Tensor of the same shape as the input `string_tensor`.

Converts each string in the input Tensor to its hash mod by a number of buckets.

The hash function is deterministic on the content of the string within the process. The hash function is a keyed hash function, where attribute key defines the key of the hash function. key is an array of 2 elements.

A strong hash is important when inputs may be malicious, e.g. URLs with additional components. Adversaries could try to make their inputs hash to the same bucket for a denial-of-service attack or to skew the results. A strong hash prevents this by making it dificult, if not infeasible, to compute inputs that hash to the same bucket. This comes at a cost of roughly 4x higher compute time than `tf.string_to_hash_bucket_fast`.

scatterNdUpdate Source

Arguments

:: (TensorType t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: A mutable Tensor. Should be from a Variable node.
-> Tensor v2 tindices	indices: A Tensor. Must be one of the following types: int32, int64. A tensor of indices into ref.
-> Tensor v3 t	updates: A Tensor. Must have the same type as ref. A tensor of updated values to add to ref.
-> Build (Tensor Ref t)	output_ref: Same as ref. Returned as a convenience for operations that want to use the updated values after the update is done.

Applies sparse updates to individual values or slices within a given

variable according to indices.

ref is a Tensor with rank P and indices is a Tensor of rank Q.

indices must be integer tensor, containing indices into ref. It must be shape `[d_0, ..., d_{Q-2}, K]` where `0 < K <= P`.

The innermost dimension of indices (with length K) corresponds to indices into elements (if `K = P`) or slices (if `K < P`) along the Kth dimension of ref.

updates is Tensor of rank `Q-1+P-K` with shape:

``` [d_0, ..., d_{Q-2}, ref.shape[K], ..., ref.shape[P-1]]. ```

For example, say we want to update 4 scattered elements to a rank-1 tensor to 8 elements. In Python, that update would look like this:

ref = tf.Variable([1, 2, 3, 4, 5, 6, 7, 8]) indices = tf.constant([[4], [3], [1] ,[7]]) updates = tf.constant([9, 10, 11, 12]) update = tf.scatter_nd_update(ref, indices, updates) with tf.Session() as sess: print sess.run(update)

The resulting update to ref would look like this:

1, 11, 3, 10, 9, 6, 7, 12

See tf.scatter_nd for more details about how to make updates to slices.

fakeQuantWithMinMaxVarsGradient Source

Arguments

:: Tensor v1 Float	gradients: Backpropagated gradients above the FakeQuantWithMinMaxVars operation.
-> Tensor v2 Float	inputs: Values passed as inputs to the FakeQuantWithMinMaxVars operation. min, max: Quantization interval, scalar floats.
-> Tensor v3 Float	min
-> Tensor v4 Float	max
-> (Tensor Value Float, Tensor Value Float, Tensor Value Float)	(backprops_wrt_input, backprop_wrt_min, backprop_wrt_max) backprops_wrt_input: Backpropagated gradients w.r.t. inputs: `gradients * (inputs >= min && inputs <= max)`. backprop_wrt_min: Backpropagated gradients w.r.t. min parameter: `sum(gradients * (inputs < min))`. backprop_wrt_max: Backpropagated gradients w.r.t. max parameter: `sum(gradients * (inputs > max))`.

Compute gradients for a FakeQuantWithMinMaxVars operation.

size Source

Arguments

:: (TensorType t, TensorType out_type, OneOf `[Int32, Int64]` out_type)
=> Tensor v1 t	input
-> Tensor Value out_type	output

Returns the size of a tensor.

This operation returns an integer representing the number of elements in input.

For example:

```prettyprint # t is [[[1, 1,, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]] size(t) ==> 12 ```

scatterDiv Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 tindices	indices: A tensor of indices into the first dimension of `ref`.
-> Tensor v3 t	updates: A tensor of values that `ref` is divided by.
-> Build (Tensor Ref t)	output_ref: = Same as `ref`. Returned as a convenience for operations that want to use the updated values after the update is done.

Divides a variable reference by sparse updates.

This operation computes

# Scalar indices ref[indices, ...] /= updates[...]

# Vector indices (for each i) ref[indices[i], ...] /= updates[i, ...]

# High rank indices (for each i, ..., j) ref[indices[i, ..., j], ...] /= updates[i, ..., j, ...]

This operation outputs ref after the update is done. This makes it easier to chain operations that need to use the reset value.

Duplicate entries are handled correctly: if multiple indices reference the same location, their contributions divide.

Requires `updates.shape = indices.shape + ref.shape[1:]`.

scatterMul Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 tindices	indices: A tensor of indices into the first dimension of `ref`.
-> Tensor v3 t	updates: A tensor of updated values to multiply to `ref`.
-> Build (Tensor Ref t)	output_ref: = Same as `ref`. Returned as a convenience for operations that want to use the updated values after the update is done.

Multiplies sparse updates into a variable reference.

This operation computes

# Scalar indices ref[indices, ...] *= updates[...]

# Vector indices (for each i) ref[indices[i], ...] *= updates[i, ...]

# High rank indices (for each i, ..., j) ref[indices[i, ..., j], ...] *= updates[i, ..., j, ...]

This operation outputs ref after the update is done. This makes it easier to chain operations that need to use the reset value.

Duplicate entries are handled correctly: if multiple indices reference the same location, their contributions multiply.

Requires `updates.shape = indices.shape + ref.shape[1:]`.

copyHost Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Input tensor.
-> Tensor Value t	output: Output tensor, deep-copied from input.

Copy Host Op.

Performs CPU-to-CPU deep-copying of tensor.

Unlike the Copy Op, this op has HostMemory constraint on its input or output.

wholeFileReader Source

Arguments

:: Build (Tensor Ref ByteString)

reader_handle: The handle to reference the Reader.

A Reader that outputs the entire contents of a file as a value.

To use, enqueue filenames in a Queue. The output of ReaderRead will be a filename (key) and the contents of that file (value).

takeManySparseFromTensorsMap Source

Arguments

:: TensorType dtype

=> Tensor v1 Int64

sparse_handles: 1-D, The N serialized SparseTensor objects. Shape: `[N]`.

-> Build (Tensor Value Int64, Tensor Value dtype, Tensor Value Int64)

(sparse_indices, sparse_values, sparse_shape)

sparse_indices: 2-D. The indices of the minibatch SparseTensor.
sparse_values: 1-D. The values of the minibatch SparseTensor.
sparse_shape: 1-D. The shape of the minibatch SparseTensor.

Read SparseTensors from a SparseTensorsMap and concatenate them.

The input sparse_handles must be an int64 matrix of shape `[N, 1]` where N is the minibatch size and the rows correspond to the output handles of AddSparseToTensorsMap or AddManySparseToTensorsMap. The ranks of the original SparseTensor objects that went into the given input ops must all match. When the final SparseTensor is created, it has rank one higher than the ranks of the incoming SparseTensor objects (they have been concatenated along a new row dimension on the left).

The output SparseTensor object's shape values for all dimensions but the first are the max across the input SparseTensor objects' shape values for the corresponding dimensions. Its first shape value is N, the minibatch size.

The input SparseTensor objects' indices are assumed ordered in standard lexicographic order. If this is not the case, after this step run SparseReorder to restore index ordering.

For example, if the handles represent an input, which is a `[2, 3]` matrix representing two original SparseTensor objects:

``` index = [ 0] [10] [20] values = [1, 2, 3] shape = [50] ```

and

``` index = [ 2] [10] values = [4, 5] shape = [30] ```

then the final SparseTensor will be:

``` index = [0 0] [0 10] [0 20] [1 2] [1 10] values = [1, 2, 3, 4, 5] shape = [2 50] ```

destroyTemporaryVariable Source

Arguments

:: TensorType t
=> Tensor Ref t	ref: A reference to the temporary variable tensor.
-> Build (Tensor Value t)	value

Destroys the temporary variable and returns its final value.

Sets output to the value of the Tensor pointed to by ref, then destroys the temporary variable called var_name. All other uses of ref *must* have executed before this op. This is typically achieved by chaining the ref through each assign op, or by using control dependencies.

Outputs the final value of the tensor pointed to by ref.

assignSub Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 t	value: The value to be subtracted to the variable.
-> Build (Tensor Ref t)	output_ref: = Same as "ref". Returned as a convenience for operations that want to use the new value after the variable has been updated.

Update ref by subtracting value from it.

This operation outputs "ref" after the update is done. This makes it easier to chain operations that need to use the reset value.

encodeJpeg Source

Arguments

:: Tensor v1 Word8	image: 3-D with shape `[height, width, channels]`.
-> Tensor Value ByteString	contents: 0-D. JPEG-encoded image.

JPEG-encode an image.

image is a 3-D uint8 Tensor of shape `[height, width, channels]`.

The attr format can be used to override the color format of the encoded output. Values can be:

`''`: Use a default format based on the number of channels in the image.
grayscale: Output a grayscale JPEG image. The channels dimension of image must be 1.
rgb: Output an RGB JPEG image. The channels dimension of image must be 3.

If format is not specified or is the empty string, a default format is picked in function of the number of channels in image:

1: Output a grayscale image.
3: Output an RGB image.

temporaryVariable Source

Arguments

:: TensorType dtype
=> Shape	shape: The shape of the variable tensor.
-> Build (Tensor Ref dtype)	ref: A reference to the variable tensor.

Returns a tensor that may be mutated, but only persists within a single step.

This is an experimental op for internal use only and it is possible to use this op in unsafe ways. DO NOT USE unless you fully understand the risks.

It is the caller's responsibility to ensure that ref is eventually passed to a matching DestroyTemporaryVariable op after all other uses have completed.

Outputs a ref to the tensor state so it may be read or modified.

E.g. var = state_ops._temporary_variable([1, 2], types.float_) var_name = var.op.name var = state_ops.assign(var, [[4.0, 5.0]]) var = state_ops.assign_add(var, [[6.0, 7.0]]) final = state_ops._destroy_temporary_variable(var, var_name=var_name)

isVariableInitialized Source

Arguments

:: TensorType dtype
=> Tensor Ref dtype	ref: Should be from a `Variable` node. May be uninitialized.
-> Build (Tensor Value Bool)	is_initialized

Checks whether a tensor has been initialized.

Outputs boolean scalar indicating whether the tensor has been initialized.

variable Source

Arguments

:: TensorType dtype
=> Shape	shape: The shape of the variable tensor.
-> Build (Tensor Ref dtype)	ref: A reference to the variable tensor.

Holds state in the form of a tensor that persists across steps.

Outputs a ref to the tensor state so it may be read or modified. TODO(zhifengc/mrry): Adds a pointer to a more detail document about sharing states in tensorflow.

sparseSparseMinimum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int64	a_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, in the canonical lexicographic ordering.
-> Tensor v2 t	a_values: 1-D. `N` non-empty values corresponding to `a_indices`.
-> Tensor v3 Int64	a_shape: 1-D. Shape of the input SparseTensor.
-> Tensor v4 Int64	b_indices: counterpart to `a_indices` for the other operand.
-> Tensor v5 t	b_values: counterpart to `a_values` for the other operand; must be of the same dtype.
-> Tensor v6 Int64	b_shape: counterpart to `a_shape` for the other operand; the two shapes must be equal.
-> (Tensor Value Int64, Tensor Value t)	(output_indices, output_values) output_indices: 2-D. The indices of the output SparseTensor. output_values: 1-D. The values of the output SparseTensor.

Returns the element-wise min of two SparseTensors.

Assumes the two SparseTensors have the same shape, i.e., no broadcasting.

betainc Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	a
-> Tensor v2 t	b
-> Tensor v3 t	x
-> Tensor Value t	z

Compute the regularized incomplete beta integral $I_x(a, b)$.

The regularized incomplete beta integral is defined as:

``` I_x(a, b) = frac{B(x; a, b)}{B(a, b)} ``` where

``` B(x; a, b) = int_0^x t^{a-1} (1 - t)^{b-1} dt ```

is the incomplete beta function and $B(a, b)$ is the *complete* beta function.

assign Source

Arguments

:: TensorType t
=> Tensor Ref t	ref: Should be from a `Variable` node. May be uninitialized.
-> Tensor v2 t	value: The value to be assigned to the variable.
-> Build (Tensor Ref t)	output_ref: = Same as "ref". Returned as a convenience for operations that want to use the new value after the variable has been reset.

Update ref by assigning value to it.

This operation outputs "ref" after the assignment is done. This makes it easier to chain operations that need to use the reset value.

sparseSoftmax Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 Int64	sp_indices: 2-D. `NNZ x R` matrix with the indices of non-empty values in a SparseTensor, in canonical ordering.
-> Tensor v2 t	sp_values: 1-D. `NNZ` non-empty values corresponding to `sp_indices`.
-> Tensor v3 Int64	sp_shape: 1-D. Shape of the input SparseTensor.
-> Tensor Value t	output: 1-D. The `NNZ` values for the result `SparseTensor`.

Applies softmax to a batched N-D SparseTensor.

The inputs represent an N-D SparseTensor with logical shape `[..., B, C]` (where `N >= 2`), and with indices sorted in the canonical lexicographic order.

This op is equivalent to applying the normal `tf.nn.softmax()` to each innermost logical submatrix with shape `[B, C]`, but with the catch that *the implicitly zero elements do not participate*. Specifically, the algorithm is equivalent to the following:

Applies `tf.nn.softmax()` to a densified view of each innermost submatrix with shape `[B, C]`, along the size-C dimension;
Masks out the original implicitly-zero locations;
Renormalizes the remaining elements.

Hence, the SparseTensor result has exactly the same non-zero indices and shape.

sparseDenseCwiseAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int64	sp_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, possibly not in canonical ordering.
-> Tensor v2 t	sp_values: 1-D. `N` non-empty values corresponding to `sp_indices`.
-> Tensor v3 Int64	sp_shape: 1-D. Shape of the input SparseTensor.
-> Tensor v4 t	dense: `R`-D. The dense Tensor operand.
-> Tensor Value t	output: 1-D. The `N` values that are operated on.

Adds up a SparseTensor and a dense Tensor, using these special rules:

Broadcasts the dense side to have the same shape as the sparse side, if eligible;
Then, only the dense values pointed to by the indices of the SparseTensor participate in the cwise addition.

By these rules, the result is a logical SparseTensor with exactly the same indices and shape, but possibly with different non-zero values. The output of this Op is the resultant non-zero values.

logicalNot Source

Arguments

:: Tensor v1 Bool	x
-> Tensor Value Bool	y

Returns the truth value of NOT x element-wise.

queueSize Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a queue.
-> Build (Tensor Value Int32)	size: The number of elements in the given queue.

Computes the number of elements in the given queue.

sparseApplyAdagrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Learning rate. Must be a scalar.
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Build (Tensor Ref t)	out: Same as "var".

Update relevant entries in '*var' and '*accum' according to the adagrad scheme.

That is for rows we have grad for, we update var and accum as follows: accum += grad * grad var -= lr * grad * (1 / sqrt(accum))

getSessionHandle Source

Arguments

:: TensorType t
=> Tensor v1 t	value: The tensor to be stored.
-> Tensor Value ByteString	handle: The handle for the tensor stored in the session state.

Store the input tensor in the state of the current session.

sparseDenseCwiseMul Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int64	sp_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, possibly not in canonical ordering.
-> Tensor v2 t	sp_values: 1-D. `N` non-empty values corresponding to `sp_indices`.
-> Tensor v3 Int64	sp_shape: 1-D. Shape of the input SparseTensor.
-> Tensor v4 t	dense: `R`-D. The dense Tensor operand.
-> Tensor Value t	output: 1-D. The `N` values that are operated on.

Component-wise multiplies a SparseTensor by a dense Tensor.

The output locations corresponding to the implicitly zero elements in the sparse tensor will be zero (i.e., will not take up storage space), regardless of the contents of the dense tensor (even if it's +/-INF and that INF*0 == NaN).

Limitation*: this Op only broadcasts the dense side to the sparse side, but not the other direction.

sparseTensorDenseAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 tindices	a_indices: 2-D. The `indices` of the `SparseTensor`, with shape `[nnz, ndims]`.
-> Tensor v2 t	a_values: 1-D. The `values` of the `SparseTensor`, with shape `[nnz]`.
-> Tensor v3 tindices	a_shape: 1-D. The `shape` of the `SparseTensor`, with shape `[ndims]`.
-> Tensor v4 t	b: `ndims`-D Tensor. With shape `a_shape`.
-> Tensor Value t	output

Adds up a SparseTensor and a dense Tensor, producing a dense Tensor.

This Op does not require a_indices be sorted in standard lexicographic order.

getSessionTensor Source

Arguments

:: TensorType dtype
=> Tensor v1 ByteString	handle: The handle for a tensor stored in the session state.
-> Tensor Value dtype	value: The tensor for the given handle.

Get the value of the tensor specified by its handle.

sparseReorder Source

Arguments

:: TensorType t
=> Tensor v1 Int64	input_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, possibly not in canonical ordering.
-> Tensor v2 t	input_values: 1-D. `N` non-empty values corresponding to `input_indices`.
-> Tensor v3 Int64	input_shape: 1-D. Shape of the input SparseTensor.
-> (Tensor Value Int64, Tensor Value t)	(output_indices, output_values) output_indices: 2-D. `N x R` matrix with the same indices as input_indices, but in canonical row-major ordering. output_values: 1-D. `N` non-empty values corresponding to `output_indices`.

Reorders a SparseTensor into the canonical, row-major ordering.

Note that by convention, all sparse ops preserve the canonical ordering along increasing dimension number. The only time ordering can be violated is during manual manipulation of the indices and values vectors to add entries.

Reordering does not affect the shape of the SparseTensor.

If the tensor has rank R and N non-empty values, input_indices has shape `[N, R]`, input_values has length N, and input_shape has length R.

sparseSplit Source

Arguments

:: TensorType t
=> Int64	num_split: The number of ways to split.
-> Tensor v1 Int64	split_dim: 0-D. The dimension along which to split. Must be in the range `[0, rank(shape))`.
-> Tensor v2 Int64	indices: 2-D tensor represents the indices of the sparse tensor.
-> Tensor v3 t	values: 1-D tensor represents the values of the sparse tensor.
-> Tensor v4 Int64	shape: 1-D. tensor represents the shape of the sparse tensor. output indices: A list of 1-D tensors represents the indices of the output sparse tensors.
-> ([Tensor Value Int64], [Tensor Value t], [Tensor Value Int64])	(output_indices, output_values, output_shape) output_indices output_values: A list of 1-D tensors represents the values of the output sparse tensors. output_shape: A list of 1-D tensors represents the shape of the output sparse tensors.

Split a SparseTensor into num_split tensors along one dimension.

If the `shape[split_dim]` is not an integer multiple of num_split. Slices `[0 : shape[split_dim] % num_split]` gets one extra dimension. For example, if `split_dim = 1` and `num_split = 2` and the input is

input_tensor = shape = [2, 7] [ a d e ] [b c ]

Graphically the output tensors are:

output_tensor[0] = shape = [2, 4] [ a ] [b c ]

output_tensor[1] = shape = [2, 3] [ d e ] [ ]

pad Source

Arguments

:: (TensorType t, TensorType tpaddings, OneOf `[Int32, Int64]` tpaddings)
=> Tensor v1 t	input
-> Tensor v2 tpaddings	paddings
-> Tensor Value t	output

Pads a tensor with zeros.

This operation pads a input with zeros according to the paddings you specify. paddings is an integer tensor with shape `[Dn, 2]`, where n is the rank of input. For each dimension D of input, `paddings[D, 0]` indicates how many zeros to add before the contents of input in that dimension, and `paddings[D, 1]` indicates how many zeros to add after the contents of input in that dimension.

The padded size of each dimension D of the output is:

`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`

For example:

```prettyprint # t is [[1, 1], [2, 2]] # paddings is [[1, 1], [2, 2]] # rank of t is 2 pad(t, paddings) ==> [[0, 0, 0, 0, 0, 0] [0, 0, 1, 1, 0, 0] [0, 0, 2, 2, 0, 0] [0, 0, 0, 0, 0, 0]] ```

sparseToDense Source

Arguments

:: (TensorType t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 tindices	sparse_indices: 0-D, 1-D, or 2-D. `sparse_indices[i]` contains the complete index where `sparse_values[i]` will be placed.
-> Tensor v2 tindices	output_shape: 1-D. Shape of the dense output tensor.
-> Tensor v3 t	sparse_values: 1-D. Values corresponding to each row of `sparse_indices`, or a scalar value to be used for all sparse indices.
-> Tensor v4 t	default_value: Scalar value to set for indices not specified in `sparse_indices`.
-> Tensor Value t	dense: Dense output tensor of shape `output_shape`.

Converts a sparse representation into a dense tensor.

Builds an array dense with shape output_shape such that

```prettyprint # If sparse_indices is scalar dense[i] = (i == sparse_indices ? sparse_values : default_value)

# If sparse_indices is a vector, then for each i dense[sparse_indices[i]] = sparse_values[i]

# If sparse_indices is an n by d matrix, then for each i in [0, n) dense[sparse_indices[i][0], ..., sparse_indices[i][d-1]] = sparse_values[i] ```

All other values in dense are set to default_value. If sparse_values is a scalar, all sparse indices are set to this single value.

Indices should be sorted in lexicographic order, and indices must not contain any repeats. If validate_indices is true, these properties are checked during execution.

sparseTensorDenseMatMul Source

Arguments

:: TensorType t
=> Tensor v1 Int64	a_indices: 2-D. The `indices` of the `SparseTensor`, size `[nnz, 2]` Matrix.
-> Tensor v2 t	a_values: 1-D. The `values` of the `SparseTensor`, size `[nnz]` Vector.
-> Tensor v3 Int64	a_shape: 1-D. The `shape` of the `SparseTensor`, size `[2]` Vector.
-> Tensor v4 t	b: 2-D. A dense Matrix.
-> Tensor Value t	product

Multiply SparseTensor (of rank 2) A by dense matrix B.

No validity checking is performed on the indices of A. However, the following input format is recommended for optimal behavior:

if adjoint_a == false: A should be sorted in lexicographically increasing order. Use SparseReorder if you're not sure. if adjoint_a == true: A should be sorted in order of increasing dimension 1 (i.e., "column major" order instead of "row major" order).

mirrorPadGrad Source

Arguments

:: (TensorType t, TensorType tpaddings, OneOf `[Int32, Int64]` tpaddings)
=> Tensor v1 t	input: The input tensor to be folded.
-> Tensor v2 tpaddings	paddings: A two-column matrix specifying the padding sizes. The number of rows must be the same as the rank of `input`.
-> Tensor Value t	output: The folded tensor.

Gradient op for MirrorPad op. This op folds a mirror-padded tensor.

This operation folds the padded areas of input by MirrorPad according to the paddings you specify. paddings must be the same as paddings argument given to the corresponding MirrorPad op.

The folded size of each dimension D of the output is:

`input.dim_size(D) - paddings(D, 0) - paddings(D, 1)`

For example:

```prettyprint # t is [[1, 2, 3], [4, 5, 6], [7, 8, 9]]. # paddings is [[0, 1]], [0, 1]]. # mode is SYMMETRIC. # rank of t is 2. pad(t, paddings) ==> [[ 1, 5] [11, 28]] ```

randomShuffle Source

Arguments

:: TensorType t
=> Tensor v1 t	value: The tensor to be shuffled.
-> Build (Tensor Value t)	output: A tensor of same shape and type as `value`, shuffled along its first dimension.

Randomly shuffles a tensor along its first dimension.

The tensor is shuffled along dimension 0, such that each `value[j]` is mapped to one and only one `output[i]`. For example, a mapping that might occur for a 3x2 tensor is:

```prettyprint [[1, 2], [[5, 6], [3, 4], ==> [1, 2], [5, 6]] [3, 4]] ```

select Source

Arguments

:: TensorType t
=> Tensor v1 Bool	condition
-> Tensor v2 t	t: = A `Tensor` which may have the same shape as `condition`. If `condition` is rank 1, `t` may have higher rank, but its first dimension must match the size of `condition`.
-> Tensor v3 t	e: = A `Tensor` with the same type and shape as `t`.
-> Tensor Value t	output: = A `Tensor` with the same type and shape as `t` and `e`.

Selects elements from t or e, depending on condition.

The t, and e tensors must all have the same shape, and the output will also have that shape.

The condition tensor must be a scalar if t and e are scalars. If t and e are vectors or higher rank, then condition must be either a scalar, a vector with size matching the first dimension of t, or must have the same shape as t.

The condition tensor acts as a mask that chooses, based on the value at each element, whether the corresponding element / row in the output should be taken from t (if true) or e (if false).

If condition is a vector and t and e are higher rank matrices, then it chooses which row (outer dimension) to copy from t and e. If condition has the same shape as t and e, then it chooses which element to copy from t and e.

For example:

```prettyprint # condition tensor is [[True, False] # [False, True]] # t is [[1, 2], # [3, 4]] # e is [[5, 6], # [7, 8]] select(condition, t, e) ==> [[1, 6], [7, 4]]

# condition tensor is [True, False] # t is [[1, 2], # [3, 4]] # e is [[5, 6], # [7, 8]] select(condition, t, e) ==> [[1, 2], [7, 8]]

```

sparseAddGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	backprop_val_grad: 1-D with shape `[nnz(sum)]`. The gradient with respect to the non-empty values of the sum.
-> Tensor v2 Int64	a_indices: 2-D. The `indices` of the `SparseTensor` A, size `[nnz(A), ndims]`.
-> Tensor v3 Int64	b_indices: 2-D. The `indices` of the `SparseTensor` B, size `[nnz(B), ndims]`.
-> Tensor v4 Int64	sum_indices: 2-D. The `indices` of the sum `SparseTensor`, size `[nnz(sum), ndims]`.
-> (Tensor Value t, Tensor Value t)	(a_val_grad, b_val_grad) a_val_grad: 1-D with shape `[nnz(A)]`. The gradient with respect to the non-empty values of A. b_val_grad: 1-D with shape `[nnz(B)]`. The gradient with respect to the non-empty values of B.

The gradient operator for the SparseAdd op.

The SparseAdd op calculates A + B, where A, B, and the sum are all represented as SparseTensor objects. This op takes in the upstream gradient w.r.t. non-empty values of the sum, and outputs the gradients w.r.t. the non-empty values of A and B.

sdcaFprint Source

Arguments

:: Tensor v1 ByteString	input: vector of strings to compute fingerprints on.
-> Tensor Value Int64	output: a (N,2) shaped matrix where N is the number of elements in the input vector. Each row contains the low and high parts of the fingerprint.

Computes fingerprints of the input strings.

tensorArrayUnpack Source

Arguments

:: TensorType t
=> Tensor Ref ByteString	handle
-> Tensor v2 t	value
-> Tensor v3 Float	flow_in
-> Build (Tensor Value Float)	flow_out

quantizedAvgPool Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Word16, Word8]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 Float	min_input: The float value that the lowest quantized input value represents.
-> Tensor v3 Float	max_input: The float value that the highest quantized input value represents.
-> (Tensor Value t, Tensor Value Float, Tensor Value Float)	(output, min_output, max_output) output min_output: The float value that the lowest quantized output value represents. max_output: The float value that the highest quantized output value represents.

Produces the average pool of the input tensor for quantized types.

adjustContrastv2 Source

Arguments

:: Tensor v1 Float	images: Images to adjust. At least 3-D.
-> Tensor v2 Float	contrast_factor: A float multiplier for adjusting contrast.
-> Tensor Value Float	output: The contrast-adjusted image or images.

Adjust the contrast of one or more images.

images is a tensor of at least 3 dimensions. The last 3 dimensions are interpreted as `[height, width, channels]`. The other dimensions only represent a collection of images, such as `[batch, height, width, channels].`

Contrast is adjusted independently for each channel of each image.

For each channel, the Op first computes the mean of the image pixels in the channel and then adjusts each component of each pixel to `(x - mean) * contrast_factor + mean`.

resourceGather Source

Arguments

:: (TensorType dtype, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> ResourceHandle dtype	resource
-> Tensor v2 tindices	indices
-> Build (Tensor Value dtype)	output

Gather slices from the variable pointed to by resource according to indices.

indices must be an integer tensor of any dimension (usually 0-D or 1-D). Produces an output tensor with shape `indices.shape + params.shape[1:]` where:

```python # Scalar indices output[:, ..., :] = params[indices, :, ... :]

# Vector indices output[i, :, ..., :] = params[indices[i], :, ... :]

# Higher rank indices output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :] ```

mergeSummary Source

Arguments

:: [Tensor v1 ByteString]	inputs: Can be of any shape. Each must contain serialized `Summary` protocol buffers.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Merges summaries.

This op creates a `Summary` protocol buffer that contains the union of all the values in the input summaries.

When the Op is run, it reports an InvalidArgument error if multiple values in the summaries to merge use the same tag.

serializeSparse Source

Arguments

:: TensorType t
=> Tensor v1 Int64	sparse_indices: 2-D. The `indices` of the `SparseTensor`.
-> Tensor v2 t	sparse_values: 1-D. The `values` of the `SparseTensor`.
-> Tensor v3 Int64	sparse_shape: 1-D. The `shape` of the `SparseTensor`.
-> Tensor Value ByteString	serialized_sparse

Serialize a SparseTensor into a string 3-vector (1-D Tensor) object.

negTrain Source

Arguments

:: Int64	num_negative_samples: Number of negative samples per example.
-> Tensor Ref Float	w_in: input word embedding.
-> Tensor Ref Float	w_out: output word embedding.
-> Tensor v3 Int32	examples: A vector of word ids.
-> Tensor v4 Int32	labels: A vector of word ids.
-> Tensor v5 Float	lr
-> Build ControlNode

Training via negative sampling.

tensorArrayCloseV2 Source

Arguments

:: Tensor v1 ByteString	handle: The handle to a TensorArray (output of TensorArray or TensorArrayGrad).
-> ControlNode

Delete the TensorArray from its resource container. This enables

the user to close and release the resource in the middle of a step/run.

threadUnsafeUnigramCandidateSampler Source

Arguments

:: Int64	num_sampled: Number of candidates to randomly sample per batch.
-> Int64	num_true: Number of true labels per context.
-> Int64	range_max: The sampler will sample integers from the interval [0, range_max).
-> Bool	unique: If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
-> Tensor v1 Int64	true_classes: A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.
-> (Tensor Value Int64, Tensor Value Float, Tensor Value Float)	(sampled_candidates, true_expected_count, sampled_expected_count) sampled_candidates: A vector of length num_sampled, in which each element is the ID of a sampled candidate. true_expected_count: A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability. sampled_expected_count: A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

Generates labels for candidate sampling with a learned unigram distribution.

See explanations of candidate sampling and the data formats at go/candidate-sampling.

For each batch, this op picks a single set of sampled candidate labels.

The advantages of sampling candidates per-batch are simplicity and the possibility of efficient dense matrix multiplication. The disadvantage is that the sampled candidates must be chosen independently of the context and of the true labels.

stringToNumber Source

Arguments

:: (TensorType out_type, OneOf `[Int32, Float]` out_type)
=> Tensor v1 ByteString	string_tensor
-> Tensor Value out_type	output: A Tensor of the same shape as the input `string_tensor`.

Converts each string in the input Tensor to the specified numeric type.

(Note that int32 overflow results in an error while float overflow results in a rounded value.)

cTCBeamSearchDecoder Source

Arguments

:: Int64	beam_width: A scalar >= 0 (beam search beam width).
-> Int64	top_paths: A scalar >= 0, <= beam_width (controls output size).
-> Tensor v1 Float	inputs: 3-D, shape: `(max_time x batch_size x num_classes)`, the logits.
-> Tensor v2 Int32	sequence_length: A vector containing sequence lengths, size `(batch)`.
-> ([Tensor Value Int64], [Tensor Value Int64], [Tensor Value Int64], Tensor Value Float)	(decoded_indices, decoded_values, decoded_shape, log_probability) decoded_indices: A list (length: top_paths) of indices matrices. Matrix j, size `(total_decoded_outputs[j] x 2)`, has indices of a `SparseTensor2`. The rows store: [batch, time]. decoded_values: A list (length: top_paths) of values vectors. Vector j, size `(length total_decoded_outputs[j])`, has the values of a `SparseTensor2`. The vector stores the decoded classes for beam j. decoded_shape: A list (length: top_paths) of shape vector. Vector j, size `(2)`, stores the shape of the decoded `SparseTensor[j]`. Its values are: `[batch_size, max_decoded_length[j]]`. log_probability: A matrix, shaped: `(batch_size x top_paths)`. The sequence log-probabilities.

Performs beam search decoding on the logits given in input.

A note about the attribute merge_repeated: For the beam search decoder, this means that if consecutive entries in a beam are the same, only the first of these is emitted. That is, when the top path is "A B B B B", "A B" is returned if merge_repeated = True but "A B B B B" is returned if merge_repeated = False.

parseTensor Source

Arguments

:: TensorType out_type
=> Tensor v1 ByteString	serialized: A scalar string containing a serialized TensorProto proto.
-> Tensor Value out_type	output: A Tensor of type `out_type`.

Transforms a serialized tensorflow.TensorProto proto into a Tensor.

imageSummary Source

Arguments

:: (TensorType t, OneOf `[Word16, Word8, Float]` t)
=> Tensor v1 ByteString	tag: Scalar. Used to build the `tag` attribute of the summary values.
-> Tensor v2 t	tensor: 4-D of shape `[batch_size, height, width, channels]` where `channels` is 1, 3, or 4.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Outputs a Summary protocol buffer with images.

The summary has up to max_images summary values containing images. The images are built from tensor which must be 4-D with shape `[batch_size, height, width, channels]` and where channels can be:

1: tensor is interpreted as Grayscale.
3: tensor is interpreted as RGB.
4: tensor is interpreted as RGBA.

The images have the same number of channels as the input tensor. For float input, the values are normalized one image at a time to fit in the range `[0, 255]`. uint8 values are unchanged. The op uses two different normalization algorithms:

If the input values are all positive, they are rescaled so the largest one is 255.
If any input value is negative, the values are shifted so input value 0.0 is at 127. They are then rescaled so that either the smallest value is 0, or the largest one is 255.

The tag argument is a scalar Tensor of type string. It is used to build the tag of the summary values:

If max_images is 1, the summary value tag is '*tag*/image'.
If max_images is greater than 1, the summary value tags are generated sequentially as '*tag*/image/0', '*tag*/image/1', etc.

The bad_color argument is the color to use in the generated images for non-finite input values. It is a unit8 1-D tensor of length channels. Each element must be in the range `[0, 255]` (It represents the value of a pixel in the output image). Non-finite values in the input tensor are replaced by this tensor in the output image. The default value is the color red.

truncateDiv Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x / y element-wise for integer types.

Truncation designates that negative numbers will round fractional quantities toward zero. I.e. -7 / 5 = 1. This matches C semantics but it is different than Python semantics. See FloorDiv for a division function that matches Python Semantics.

NOTE*: TruncateDiv supports broadcasting. More about broadcasting here

cholesky Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input: Shape is `[..., M, M]`.
-> Tensor Value t	output: Shape is `[..., M, M]`.

Computes the Cholesky decomposition of one or more square matrices.

The input is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices, with the same constraints as the single matrix Cholesky decomposition above. The output is a tensor of the same shape as the input containing the Cholesky decompositions for all input submatrices `[..., :, :]`.

batchMatrixSolveLs Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	matrix
-> Tensor v2 t	rhs
-> Tensor v3 Double	l2_regularizer
-> Tensor Value t	output

lookupTableExport Source

Arguments

:: (TensorType tkeys, TensorType tvalues)

=> Tensor Ref ByteString

table_handle: Handle to the table.

-> Build (Tensor Value tkeys, Tensor Value tvalues)

(keys, values)

keys: Vector of all keys present in the table.
values: Tensor of all values in the table. Indexed in parallel with keys.

Outputs all keys and values in the table.

batchSvd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Double, Float]` t)

=> Tensor v1 t

input

-> (Tensor Value t, Tensor Value t, Tensor Value t)

(s, u, v)

s
u
v

resizeBicubic Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	images: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 Int32	size: = A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The new size for the images.
-> Tensor Value Float	resized_images: 4-D with shape `[batch, new_height, new_width, channels]`.

Resize images to size using bicubic interpolation.

Input images can be of different types but output images are always float.

hSVToRGB Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	images: 1-D or higher rank. HSV data to convert. Last dimension must be size 3.
-> Tensor Value t	output: `images` converted to RGB.

Convert one or more images from HSV to RGB.

Outputs a tensor of the same shape as the images tensor, containing the RGB value of the pixels. The output is only well defined if the value in images are in `[0,1]`.

See rgb_to_hsv for a description of the HSV encoding.

avgPool3D Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: Shape `[batch, depth, rows, cols, channels]` tensor to pool over.
-> Tensor Value t	output: The average pooled output tensor.

Performs 3D average pooling on the input.

stackClose Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a stack.
-> Build ControlNode

Delete the stack from its resource container.

assignVariableOp Source

Arguments

:: TensorType dtype
=> ResourceHandle dtype	resource: handle to the resource in which to store the variable.
-> Tensor v2 dtype	value: the value to set the new tensor to use.
-> Build ControlNode

Assigns a new value to a variable.

Any ReadVariableOp with a control dependency on this op is guaranteed to return this value or a subsequent newer value of the variable.

lRN Source

Arguments

:: (TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	input: 4-D.
-> Tensor Value t	output

Local Response Normalization.

The 4-D input tensor is treated as a 3-D array of 1-D vectors (along the last dimension), and each vector is normalized independently. Within a given vector, each component is divided by the weighted, squared sum of inputs within depth_radius. In detail,

sqr_sum[a, b, c, d] = sum(input[a, b, c, d - depth_radius : d + depth_radius + 1] ** 2) output = input / (bias + alpha * sqr_sum) ** beta

For details, see Krizhevsky et al., ImageNet classification with deep convolutional neural networks (NIPS 2012).

zeta Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	q
-> Tensor Value t	z

Compute the Hurwitz zeta function $zeta(x, q)$.

The Hurwitz zeta function is defined as:

``` zeta(x, q) = sum_{n=0}^{infty} (q + n)^{-x} ```

tensorArrayGradV2 Source

Arguments

:: Tensor v1 ByteString	handle: The handle to the forward TensorArray.
-> Tensor v2 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Build (Tensor Value ByteString)	grad_handle

Creates a TensorArray for storing the gradients of values in the given handle.

If the given TensorArray gradient already exists, returns a reference to it.

Locks the size of the original TensorArray by disabling its dynamic size flag.

*A note about the input flow_in:**

The handle flow_in forces the execution of the gradient lookup to occur only after certain other operations have occurred. For example, when the forward TensorArray is dynamically sized, writes to this TensorArray may resize the object. The gradient TensorArray is statically sized based on the size of the forward TensorArray when this operation executes. Furthermore, the size of the forward TensorArray is frozen by this call. As a result, the flow is used to ensure that the call to generate the gradient TensorArray only happens after all writes are executed.

In the case of dynamically sized TensorArrays, gradient computation should only be performed on read operations that have themselves been chained via flow to occur only after all writes have executed. That way the final size of the forward TensorArray is known when this operation is called.

*A note about the source attribute:**

TensorArray gradient calls use an accumulator TensorArray object. If multiple gradients are calculated and run in the same session, the multiple gradient nodes may accidentally flow throuth the same accumulator TensorArray. This double counts and generally breaks the TensorArray gradient flow.

The solution is to identify which gradient call this particular TensorArray gradient is being called in. This is performed by identifying a unique string (e.g. "gradients", "gradients_1", ...) from the input gradient Tensor's name. This string is used as a suffix when creating the TensorArray gradient object here (the attribute source).

The attribute source is added as a suffix to the forward TensorArray's name when performing the creation / lookup, so that each separate gradient calculation gets its own TensorArray accumulator.

cast Source

Arguments

:: (TensorType srcT, TensorType dstT)
=> Tensor v1 srcT	x
-> Tensor Value dstT	y

Cast x of type SrcT to y of DstT.

erf Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value t	y

Computes the Gauss error function of x element-wise.

batchMatrixTriangularSolve Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	matrix
-> Tensor v2 t	rhs
-> Tensor Value t	output

resourceScatterAdd Source

Arguments

:: (TensorType dtype, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` dtype, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> ResourceHandle dtype	resource: Should be from a `Variable` node.
-> Tensor v2 tindices	indices: A tensor of indices into the first dimension of `ref`.
-> Tensor v3 dtype	updates: A tensor of updated values to add to `ref`.
-> Build ControlNode

Adds sparse updates to the variable referenced by resource.

This operation computes

# Scalar indices ref[indices, ...] += updates[...]

# Vector indices (for each i) ref[indices[i], ...] += updates[i, ...]

# High rank indices (for each i, ..., j) ref[indices[i, ..., j], ...] += updates[i, ..., j, ...]

Duplicate entries are handled correctly: if multiple indices reference the same location, their contributions add.

Requires `updates.shape = indices.shape + ref.shape[1:]`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/ScatterAdd.png" alt /div

batchCholeskyGrad Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	l
-> Tensor v2 t	grad
-> Tensor Value t	output

batchMatrixInverse Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input
-> Tensor Value t	output

refIdentity Source

Arguments

:: TensorType t
=> Tensor Ref t	input
-> Build (Tensor Ref t)	output

Return the same ref tensor as the input ref tensor.

svd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Double, Float]` t)

=> Tensor v1 t

input: A tensor of shape `[..., M, N]` whose inner-most 2 dimensions form matrices of size `[M, N]`. Let P be the minimum of M and N.

-> (Tensor Value t, Tensor Value t, Tensor Value t)

(s, u, v)

s: Singular values. Shape is `[..., P]`.
u: Left singular vectors. If full_matrices is False then shape is `[..., M, M]`; if full_matrices is True then shape is `[..., M, P]`. Undefined if compute_uv is False.
v: Left singular vectors. If full_matrices is False then shape is `[..., N, N]`. If full_matrices is True then shape is `[..., N, P]`. Undefined if compute_uv is false.

Computes the singular value decompositions of one or more matrices.

Computes the SVD of each inner matrix in input such that `input[..., :, :] = u[..., :, :] * diag(s[..., :, :]) * transpose(v[..., :, :])`

```prettyprint # a is a tensor containing a batch of matrices. # s is a tensor of singular values for each matrix. # u is the tensor containing of left singular vectors for each matrix. # v is the tensor containing of right singular vectors for each matrix. s, u, v = svd(a) s, _, _ = svd(a, compute_uv=False) ```

matrixSolveLs Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	matrix: Shape is `[..., M, N]`.
-> Tensor v2 t	rhs: Shape is `[..., M, K]`.
-> Tensor v3 Double	l2_regularizer: Scalar tensor. `compatibility(numpy) Equivalent to np.linalg.lstsq`end_compatibility
-> Tensor Value t	output: Shape is `[..., N, K]`.

Solves one or more linear least-squares problems.

matrix is a tensor of shape `[..., M, N]` whose inner-most 2 dimensions form matrices of size `[M, N]`. Rhs is a tensor of shape `[..., M, K]`. The output is a tensor shape `[..., N, K]` where each output matrix solves each of the equations matrix[..., :, :] * output[..., :, :] = rhs[..., :, :] in the least squares sense.

matrix and right-hand sides in the batch:

matrix=$A in Re^{m times n}$, rhs=$B in Re^{m times k}$, output=$X in Re^{n times k}$, l2_regularizer=$lambda$.

If fast is True, then the solution is computed by solving the normal equations using Cholesky decomposition. Specifically, if $m ge n$ then $X = (A^T A + lambda I)^{-1} A^T B$, which solves the least-squares problem $X = mathrm{argmin}_{Z in Re^{n times k}} ||A Z - B||_F^2 + lambda ||Z||_F^2$. If $m lt n$ then output is computed as $X = A^T (A A^T + lambda I)^{-1} B$, which (for $lambda = 0$) is the minimum-norm solution to the under-determined linear system, i.e. $X = mathrm{argmin}_{Z in Re^{n times k}} ||Z||_F^2 $, subject to $A Z = B$. Notice that the fast path is only numerically stable when $A$ is numerically full rank and has a condition number $mathrm{cond}(A) lt frac{1}{sqrt{epsilon_{mach}}}$ or$lambda$ is sufficiently large.

If fast is False an algorithm based on the numerically robust complete orthogonal decomposition is used. This computes the minimum-norm least-squares solution, even when $A$ is rank deficient. This path is typically 6-7 times slower than the fast path. If fast is False then l2_regularizer is ignored.

pack Source

Arguments

:: TensorType t
=> [Tensor v1 t]	values: Must be of same shape and type.
-> Tensor Value t	output: The packed tensor.

Packs a list of N rank-R tensors into one rank-`(R+1)` tensor.

Packs the N tensors in values into a tensor with rank one higher than each tensor in values, by packing them along the axis dimension. Given a list of tensors of shape `(A, B, C)`;

if `axis == 0` then the output tensor will have the shape `(N, A, B, C)`. if `axis == 1` then the output tensor will have the shape `(A, N, B, C)`. Etc.

For example:

```prettyprint # x is [1, 4] # y is [2, 5] # z is [3, 6] pack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim. pack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]] ```

This is the opposite of unpack.

barrierClose Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a barrier.
-> Build ControlNode

Closes the given barrier.

This operation signals that no more new elements will be inserted in the given barrier. Subsequent InsertMany that try to introduce a new key will fail. Subsequent InsertMany operations that just add missing components to already existing elements will continue to succeed. Subsequent TakeMany operations will continue to succeed if sufficient completed elements remain in the barrier. Subsequent TakeMany operations that would block will fail immediately.

selfAdjointEigV2 Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)

=> Tensor v1 t

input: Tensor input of shape `[N, N]`.

-> (Tensor Value t, Tensor Value t)

(e, v)

e: Eigenvalues. Shape is `[N]`.
v: Eigenvectors. Shape is `[N, N]`.

Computes the eigen decomposition of one or more square self-adjoint matrices.

Computes the eigenvalues and (optionally) eigenvectors of each inner matrix in input such that `input[..., :, :] = v[..., :, :] * diag(e[..., :])`.

```prettyprint # a is a tensor. # e is a tensor of eigenvalues. # v is a tensor of eigenvectors. e, v = self_adjoint_eig(a) e = self_adjoint_eig(a, compute_v=False) ```

scatterSub Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 tindices	indices: A tensor of indices into the first dimension of `ref`.
-> Tensor v3 t	updates: A tensor of updated values to subtract from `ref`.
-> Build (Tensor Ref t)	output_ref: = Same as `ref`. Returned as a convenience for operations that want to use the updated values after the update is done.

Subtracts sparse updates to a variable reference.

# Scalar indices ref[indices, ...] -= updates[...]

# Vector indices (for each i) ref[indices[i], ...] -= updates[i, ...]

# High rank indices (for each i, ..., j) ref[indices[i, ..., j], ...] -= updates[i, ..., j, ...]

This operation outputs ref after the update is done. This makes it easier to chain operations that need to use the reset value.

Duplicate entries are handled correctly: if multiple indices reference the same location, their (negated) contributions add.

Requires `updates.shape = indices.shape + ref.shape[1:]`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/ScatterSub.png" alt /div

selfAdjointEig Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input: Shape is `[..., M, M]`.
-> Tensor Value t	output: Shape is `[..., M+1, M]`.

Computes the Eigen Decomposition of a batch of square self-adjoint matrices.

The input is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices, with the same constraints as the single matrix SelfAdjointEig.

The result is a [..., M+1, M] matrix with [..., 0,:] containing the eigenvalues, and subsequent [...,1:, :] containing the eigenvectors.

stopGradient Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor Value t	output

Stops gradient computation.

When executed in a graph, this op outputs its input tensor as-is.

When building ops to compute gradients, this op prevents the contribution of its inputs to be taken into account. Normally, the gradient generator adds ops to a graph to compute the derivatives of a specified loss by recursively finding out inputs that contributed to its computation. If you insert this op in the graph it inputs are masked from the gradient generator. They are not taken into account for computing gradients.

This is useful any time you want to compute a value with TensorFlow but need to pretend that the value was a constant. Some examples include:

The *EM* algorithm where the *M-step* should not involve backpropagation through the output of the *E-step*.
Contrastive divergence training of Boltzmann machines where, when differentiating the energy function, the training must not backpropagate through the graph that generated the samples from the model.
Adversarial training, where no backprop should happen through the adversarial example generation process.

argMax Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	input
-> Tensor v2 tidx	dimension: int32, 0 <= dimension < rank(input). Describes which dimension of the input Tensor to reduce across. For vectors, use dimension = 0.
-> Tensor Value Int64	output

Returns the index with the largest value across dimensions of a tensor.

choleskyGrad Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	l: Output of batch Cholesky algorithm l = cholesky(A). Shape is `[..., M, M]`. Algorithm depends only on lower triangular part of the innermost matrices of this tensor.
-> Tensor v2 t	grad: df/dl where f is some scalar function. Shape is `[..., M, M]`. Algorithm depends only on lower triangular part of the innermost matrices of this tensor.
-> Tensor Value t	output: Symmetrized version of df/dA . Shape is `[..., M, M]`

Computes the reverse mode backpropagated gradient of the Cholesky algorithm.

For an explanation see "Differentiation of the Cholesky algorithm" by Iain Murray http://arxiv.org/abs/1602.07527.

sparseReshape Source

Arguments

:: Tensor v1 Int64	input_indices: 2-D. `N x R_in` matrix with the indices of non-empty values in a SparseTensor.
-> Tensor v2 Int64	input_shape: 1-D. `R_in` vector with the input SparseTensor's dense shape.
-> Tensor v3 Int64	new_shape: 1-D. `R_out` vector with the requested new dense shape.
-> (Tensor Value Int64, Tensor Value Int64)	(output_indices, output_shape) output_indices: 2-D. `N x R_out` matrix with the updated indices of non-empty values in the output SparseTensor. output_shape: 1-D. `R_out` vector with the full dense shape of the output SparseTensor. This is the same as `new_shape` but with any -1 dimensions filled in.

Reshapes a SparseTensor to represent values in a new dense shape.

This operation has the same semantics as reshape on the represented dense tensor. The input_indices are recomputed based on the requested new_shape.

If one component of new_shape is the special value -1, the size of that dimension is computed so that the total dense size remains constant. At most one component of new_shape can be -1. The number of dense elements implied by new_shape must be the same as the number of dense elements originally implied by input_shape.

Reshaping does not affect the order of values in the SparseTensor.

If the input tensor has rank R_in and N non-empty values, and new_shape has length R_out, then input_indices has shape `[N, R_in]`, input_shape has length R_in, output_indices has shape `[N, R_out]`, and output_shape has length R_out.

sparseApplyAdadelta Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor Ref t	accum_update: : Should be from a Variable().
-> Tensor v4 t	lr: Learning rate. Must be a scalar.
-> Tensor v5 t	rho: Decay factor. Must be a scalar.
-> Tensor v6 t	epsilon: Constant factor. Must be a scalar.
-> Tensor v7 t	grad: The gradient.
-> Tensor v8 tindices	indices: A vector of indices into the first dimension of var and accum.
-> Build (Tensor Ref t)	out: Same as "var".

var: Should be from a Variable().

dilation2DBackpropFilter Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, depth]`.
-> Tensor v2 t	filter: 3-D with shape `[filter_height, filter_width, depth]`.
-> Tensor v3 t	out_backprop: 4-D with shape `[batch, out_height, out_width, depth]`.
-> Tensor Value t	filter_backprop: 3-D with shape `[filter_height, filter_width, depth]`.

Computes the gradient of morphological 2-D dilation with respect to the filter.

batchSelfAdjointEigV2 Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)

=> Tensor v1 t

input

-> (Tensor Value t, Tensor Value t)

(e, v)

e
v

barrierIncompleteSize Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a barrier.
-> Build (Tensor Value Int32)	size: The number of incomplete elements (i.e. those with some of their value components not set) in the barrier.

Computes the number of incomplete elements in the given barrier.

fakeQuantWithMinMaxVars Source

Arguments

:: Tensor v1 Float	inputs
-> Tensor v2 Float	min
-> Tensor v3 Float	max
-> Tensor Value Float	outputs

Fake-quantize the inputs tensor of type float and shape `[b, h, w, d]` via

global float scalars min and max to outputs tensor of same shape as inputs.

min; max: is the clamping range for the inputs data. Op divides this range into 255 steps (total of 256 values), then replaces each inputs value with the closest of the quantized step values.

This operation has a gradient and thus allows for training min and max values.

readVariableOp Source

Arguments

:: TensorType dtype
=> ResourceHandle dtype	resource: handle to the resource in which to store the variable.
-> Build (Tensor Value dtype)	value

Reads the value of a variable.

The tensor returned by this operation is immutable.

The value returned by this operation is guaranteed to be influenced by all the writes on which this operation depends directly or indirectly, and to not be influenced by any of the writes which depend directly or indirectly on this operation.

fusedBatchNormGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	y_backprop: A 4D Tensor for the gradient with respect to y.
-> Tensor v2 t	x: A 4D Tensor for input data.
-> Tensor v3 t	scale: A 1D Tensor for scaling factor, to scale the normalized x.
-> Tensor v4 t	reserve_space_1: A 1D Tensor for the computed batch mean, to be reused in the gradient computation.
-> Tensor v5 t	reserve_space_2: A 1D Tensor for the computed batch variance (inverted variance in the cuDNN case), to be used in the gradient computation.
-> (Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t, Tensor Value t)	(x_backprop, scale_backprop, offset_backprop, reserve_space_3, reserve_space_4) x_backprop: A 4D Tensor for the gradient with respect to x. scale_backprop: A 1D Tensor for the gradient with respect to scale. offset_backprop: A 1D Tensor for the gradient with respect to offset. reserve_space_3: Unused placeholder to match the mean input in FusedBatchNorm. reserve_space_4: Unused placeholder to match the variance input in FusedBatchNorm.

Gradient for batch normalization.

Note that the size of 4D Tensors are defined by either NHWC or NCHW. The size of 1D Tensors matches the dimension C of the 4D Tensors.

paddingFIFOQueue Source

Arguments

:: Build (Tensor Ref ByteString)

handle: The handle to the queue.

A queue that produces elements in first-in first-out order.

Variable-size shapes are allowed by setting the corresponding shape dimensions to 0 in the shape attr. In this case DequeueMany will pad up to the maximum size of any given element in the minibatch. See below for details.

matrixInverse Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)

=> Tensor v1 t

input: Shape is `[..., M, M]`.

-> Tensor Value t

output: Shape is `[..., M, M]`.

compatibility(numpy) Equivalent to np.linalg.invend_compatibility

Computes the inverse of one or more square invertible matrices or their

adjoints (conjugate transposes).

The input is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices. The output is a tensor of the same shape as the input containing the inverse for all input submatrices `[..., :, :]`.

The op uses LU decomposition with partial pivoting to compute the inverses.

If a matrix is not invertible there is no guarantee what the op does. It may detect the condition and raise an exception or it may simply return a garbage result.

audioSummaryV2 Source

Arguments

:: Tensor v1 ByteString	tag: Scalar. Used to build the `tag` attribute of the summary values.
-> Tensor v2 Float	tensor: 2-D of shape `[batch_size, frames]`.
-> Tensor v3 Float	sample_rate: The sample rate of the signal in hertz.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Outputs a Summary protocol buffer with audio.

The summary has up to max_outputs summary values containing audio. The audio is built from tensor which must be 3-D with shape `[batch_size, frames, channels]` or 2-D with shape `[batch_size, frames]`. The values are assumed to be in the range of `[-1.0, 1.0]` with a sample rate of sample_rate.

The tag argument is a scalar Tensor of type string. It is used to build the tag of the summary values:

If max_outputs is 1, the summary value tag is '*tag*/audio'.
If max_outputs is greater than 1, the summary value tags are generated sequentially as '*tag*/audio/0', '*tag*/audio/1', etc.

matrixDeterminant Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input: Shape is `[..., M, M]`.
-> Tensor Value t	output: Shape is `[...]`.

Computes the determinant of one ore more square matrices.

The input is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices. The output is a tensor containing the determinants for all input submatrices `[..., :, :]`.

writeFile Source

Arguments

:: Tensor v1 ByteString	filename: scalar. The name of the file to which we write the contents.
-> Tensor v2 ByteString	contents: scalar. The content to be written to the output file.
-> ControlNode

Writes contents to the file at input filename. Creates file if not existing.

quantizedConcat Source

Arguments

:: TensorType t
=> Tensor v1 Int32	concat_dim: 0-D. The dimension along which to concatenate. Must be in the range [0, rank(values)).
-> [Tensor v2 t]	values: The `N` Tensors to concatenate. Their ranks and types must match, and their sizes must match in all dimensions except `concat_dim`.
-> [Tensor v3 Float]	input_mins: The minimum scalar values for each of the input tensors.
-> [Tensor v4 Float]	input_maxes: The maximum scalar values for each of the input tensors.
-> (Tensor Value t, Tensor Value Float, Tensor Value Float)	(output, output_min, output_max) output: A `Tensor` with the concatenation of values stacked along the `concat_dim` dimension. This tensor's shape matches that of `values` except in `concat_dim` where it has the sum of the sizes. output_min: The float value that the minimum quantized output value represents. output_max: The float value that the maximum quantized output value represents.

Concatenates quantized tensors along one dimension.

varHandleOp Source

Arguments

:: TensorType dtype
=> Shape	shape: The (possibly partially specified) shape of this variable.
-> Build (ResourceHandle dtype)	resource

Creates a handle to a Variable resource.

stridedSliceAssign Source

Arguments

:: (TensorType t, TensorType index, OneOf `[Int32, Int64]` index)
=> Tensor Ref t	ref
-> Tensor v2 index	begin
-> Tensor v3 index	end
-> Tensor v4 index	strides
-> Tensor v5 t	value
-> Build (Tensor Ref t)	output_ref

Assign value to the sliced l-value reference of ref.

The values of value are assigned to the positions in the variable ref that are selected by the slice parameters. The slice parameters `begin, end, strides, etc. work exactly as in StridedSlice.

NOTE this op currently does not support broadcasting and so value's shape must be exactly the shape produced by the slice of ref.

varIsInitializedOp Source

Arguments

:: ResourceHandle dtype	resource: the input resource handle.
-> Build (Tensor Value Bool)	is_initialized: a scalar boolean which is true if the variable has been initialized.

Checks whether a resource handle-based variable has been initialized.

sparseApplyRMSProp Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	ms: Should be from a Variable().
-> Tensor Ref t	mom: Should be from a Variable().
-> Tensor v4 t	lr: Scaling factor. Must be a scalar.
-> Tensor v5 t	rho: Decay rate. Must be a scalar.
-> Tensor v6 t	momentum
-> Tensor v7 t	epsilon: Ridge term. Must be a scalar.
-> Tensor v8 t	grad: The gradient.
-> Tensor v9 tindices	indices: A vector of indices into the first dimension of var, ms and mom.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the RMSProp algorithm.

Note that in dense implementation of this algorithm, ms and mom will update even if the grad is zero, but in this sparse implementation, ms and mom will not update in iterations during which the grad is zero.

mean_square = decay * mean_square + (1-decay) * gradient ** 2 Delta = learning_rate * gradient / sqrt(mean_square + epsilon)

ms <- rho * ms_{t-1} + (1-rho) * grad * grad mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon) var <- var - mom

batchCholesky Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input
-> Tensor Value t	output

tensorArrayGather Source

Arguments

:: TensorType dtype
=> Tensor Ref ByteString	handle
-> Tensor v2 Int32	indices
-> Tensor v3 Float	flow_in
-> Build (Tensor Value dtype)	value

readerRestoreState Source

Arguments

:: Tensor Ref ByteString	reader_handle: Handle to a Reader.
-> Tensor v2 ByteString	state: Result of a ReaderSerializeState of a Reader with type matching reader_handle.
-> Build ControlNode

Restore a reader to a previously saved state.

Not all Readers support being restored, so this can produce an Unimplemented error.

sqrtGrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the gradient for the sqrt of x wrt its input.

Specifically, `grad = dy * 0.5 / y`, where `y = sqrt(x)`, and dy is the corresponding input gradient.

split Source

Arguments

:: TensorType t
=> Int64	num_split: The number of ways to split. Must evenly divide `value.shape[split_dim]`.
-> Tensor v1 Int32	split_dim: 0-D. The dimension along which to split. Must be in the range `[0, rank(value))`.
-> Tensor v2 t	value: The tensor to split.
-> [Tensor Value t]	output: They are identically shaped tensors, whose shape matches that of `value` except along `split_dim`, where their sizes are `values.shape[split_dim] / num_split`.

Splits a tensor into num_split tensors along one dimension.

textLineReader Source

Arguments

:: Build (Tensor Ref ByteString)

reader_handle: The handle to reference the Reader.

A Reader that outputs the lines of a file delimited by '\n'.

matrixBandPart Source

Arguments

:: TensorType t
=> Tensor v1 t	input: Rank `k` tensor.
-> Tensor v2 Int64	num_lower: 0-D tensor. Number of subdiagonals to keep. If negative, keep entire lower triangle.
-> Tensor v3 Int64	num_upper: 0-D tensor. Number of superdiagonals to keep. If negative, keep entire upper triangle.
-> Tensor Value t	band: Rank `k` tensor of the same shape as input. The extracted banded tensor.

Copy a tensor setting everything outside a central band in each innermost matrix

to zero.

The band part is computed as follows: Assume input has k dimensions `[I, J, K, ..., M, N]`, then the output is a tensor with the same shape where

`band[i, j, k, ..., m, n] = in_band(m, n) * input[i, j, k, ..., m, n]`.

The indicator function

`in_band(m, n) = (num_lower < 0 || (m-n) <= num_lower)) && (num_upper < 0 || (n-m) <= num_upper)`.

For example:

```prettyprint # if input is [[ 0, 1, 2, 3] [-1, 0, 1, 2] [-2, -1, 0, 1] [-3, -2, -1, 0]],

tf.matrix_band_part(input, 1, -1) ==> [[ 0, 1, 2, 3] [-1, 0, 1, 2] [ 0, -1, 0, 1] [ 0, 0, -1, 0]],

tf.matrix_band_part(input, 2, 1) ==> [[ 0, 1, 0, 0] [-1, 0, 1, 0] [-2, -1, 0, 1] [ 0, -2, -1, 0]] ```

Useful special cases:

```prettyprint tf.matrix_band_part(input, 0, -1) ==> Upper triangular part. tf.matrix_band_part(input, -1, 0) ==> Lower triangular part. tf.matrix_band_part(input, 0, 0) ==> Diagonal. ```

queueClose Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a queue.
-> Build ControlNode

Closes the given queue.

This operation signals that no more elements will be enqueued in the given queue. Subsequent Enqueue(Many) operations will fail. Subsequent Dequeue(Many) operations will continue to succeed if sufficient elements remain in the queue. Subsequent Dequeue(Many) operations that would block will fail immediately.

mergeV2Checkpoints Source

Arguments

:: Tensor v1 ByteString	checkpoint_prefixes: prefixes of V2 checkpoints to merge.
-> Tensor v2 ByteString	destination_prefix: scalar. The desired final prefix. Allowed to be the same as one of the checkpoint_prefixes.
-> ControlNode

V2 format specific: merges the metadata files of sharded checkpoints. The

result is one logical checkpoint, with one physical metadata file and renamed data files.

Intended for "grouping" multiple checkpoints in a sharded checkpoint setup.

If delete_old_dirs is true, attempts to delete recursively the dirname of each path in the input checkpoint_prefixes. This is useful when those paths are non user-facing temporary locations.

barrierReadySize Source

Arguments

:: Tensor Ref ByteString	handle: The handle to a barrier.
-> Build (Tensor Value Int32)	size: The number of complete elements (i.e. those with all of their value components set) in the barrier.

Computes the number of complete elements in the given barrier.

randomShuffleQueue Source

Arguments

:: Build (Tensor Ref ByteString)

handle: The handle to the queue.

A queue that randomizes the order of elements.

notEqual Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Bool, ByteString, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x != y) element-wise.

NOTE*: NotEqual supports broadcasting. More about broadcasting here

nonMaxSuppression Source

Arguments

:: Tensor v1 Float	boxes: A 2-D float tensor of shape `[num_boxes, 4]`.
-> Tensor v2 Float	scores: A 1-D float tensor of shape `[num_boxes]` representing a single score corresponding to each box (each row of boxes).
-> Tensor v3 Int32	max_output_size: A scalar integer tensor representing the maximum number of boxes to be selected by non max suppression.
-> Tensor Value Int32	selected_indices: A 1-D integer tensor of shape `[M]` representing the selected indices from the boxes tensor, where `M <= max_output_size`.

Greedily selects a subset of bounding boxes in descending order of score,

pruning away boxes that have high intersection-over-union (IOU) overlap with previously selected boxes. Bounding boxes are supplied as [y1, x1, y2, x2], where (y1, x1) and (y2, x2) are the coordinates of any diagonal pair of box corners and the coordinates can be provided as normalized (i.e., lying in the interval [0, 1]) or absolute. Note that this algorithm is agnostic to where the origin is in the coordinate system. Note that this algorithm is invariant to orthogonal transformations and translations of the coordinate system; thus translating or reflections of the coordinate system result in the same boxes being selected by the algorithm.

The output of this operation is a set of integers indexing into the input collection of bounding boxes representing the selected boxes. The bounding box coordinates corresponding to the selected indices can then be obtained using the `tf.gather operation`. For example:

selected_indices = tf.image.non_max_suppression( boxes, scores, max_output_size, iou_threshold) selected_boxes = tf.gather(boxes, selected_indices)

tensorArrayWrite Source

Arguments

:: TensorType t
=> Tensor Ref ByteString	handle
-> Tensor v2 Int32	index
-> Tensor v3 t	value
-> Tensor v4 Float	flow_in
-> Build (Tensor Value Float)	flow_out

quantizeAndDequantize Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input: Tensor to quantize and then dequantize.
-> Tensor Value t	output

Quantizes then dequantizes a tensor.

This op simulates the precision loss from the quantized forward pass by: 1. Quantizing the tensor to fixed point numbers, which should match the target quantization method when it is used in inference. 2. Dequantizing it back to floating point numbers for the following ops, most likely matmul.

There are different ways to quantize. This version does not use the full range of the output type, choosing to elide the lowest possible value for symmetry (e.g., output range is -127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to 0.

To perform this op, we first find the range of values in our tensor. The range we use is always centered on 0, so we find m such that

m = max(abs(input_min), abs(input_max)) if range_given is true,
m = max(max(abs(min_elem(input)), abs(max_elem(input))) otherwise.

Our input tensor range is then [-m, m].

Next, we choose our fixed-point quantization buckets, [min_fixed, max_fixed]. If signed_input is true, this is

min_fixed, max_fixed: =
-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1: .

Otherwise, if signed_input is false, the fixed-point range is

min_fixed, max_fixed: = [0, (1 << num_bits) - 1].

From this we compute our scaling factor, s:

s = (max_fixed - min_fixed) / (2 * m).

Now we can quantize and dequantize the elements of our tensor. An element e is transformed into e':

e' = (e * s).round_to_nearest() / s.

Note that we have a different number of buckets in the signed vs. unsigned cases. For example, if num_bits == 8, we get 254 buckets in the signed case vs. 255 in the unsigned case.

For example, suppose num_bits = 8 and m = 1. Then

min_fixed, max_fixed: = [-127, 127], and s = (127 + 127) / 2 = 127.

Given the vector {-1, -0.5, 0, 0.3}, this is quantized to {-127, -63, 0, 38}, and dequantized to {-1, -63.0127, 0, 38.0127}.

readerRead Source

Arguments

:: Tensor Ref ByteString

reader_handle: Handle to a Reader.

-> Tensor Ref ByteString

queue_handle: Handle to a Queue, with string work items.

-> Build (Tensor Value ByteString, Tensor Value ByteString)

(key, value)

key: A scalar.
value: A scalar.

Returns the next record (key, value pair) produced by a Reader.

Will dequeue from the input queue if necessary (e.g. when the Reader needs to start reading from a new file since it has finished with the previous file).

matrixTriangularSolve Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	matrix: Shape is `[..., M, M]`.
-> Tensor v2 t	rhs: Shape is `[..., M, K]`.
-> Tensor Value t	output: Shape is `[..., M, K]`.

Solves systems of linear equations with upper or lower triangular matrices by

backsubstitution.

matrix is a tensor of shape `[..., M, M]` whose inner-most 2 dimensions form square matrices. If lower is True then the strictly upper triangular part of each inner-most matrix is assumed to be zero and not accessed. If lower is False then the strictly lower triangular part of each inner-most matrix is assumed to be zero and not accessed. rhs is a tensor of shape `[..., M, K]`.

The output is a tensor of shape `[..., M, K]`. If adjoint is True then the innermost matrices in output` satisfy matrix equations `matrix[..., :, :] * output[..., :, :] = rhs[..., :, :]`. If adjoint is False then the strictly then the innermost matrices in output satisfy matrix equations `adjoint(matrix[..., i, k]) * output[..., k, j] = rhs[..., i, j]`.

tensorArraySplitV2 Source

Arguments

:: TensorType t
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 t	value: The concatenated tensor to write to the TensorArray.
-> Tensor v3 Int64	lengths: The vector of lengths, how to split the rows of value into the TensorArray.
-> Tensor v4 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value Float	flow_out: A float scalar that enforces proper chaining of operations.

Split the data from the input value into TensorArray elements.

Assuming that lengths takes on values

```(n0, n1, ..., n(T-1))```

and that value has shape

```(n0 + n1 + ... + n(T-1) x d0 x d1 x ...)```,

this splits values into a TensorArray with T tensors.

TensorArray index t will be the subtensor of values with starting position

```(n0 + n1 + ... + n(t-1), 0, 0, ...)```

and having size

```nt x d0 x d1 x ...```

restore Source

Arguments

:: TensorType dt
=> Tensor v1 ByteString	file_pattern: Must have a single element. The pattern of the files from which we read the tensor.
-> Tensor v2 ByteString	tensor_name: Must have a single element. The name of the tensor to be restored.
-> Tensor Value dt	tensor: The restored tensor.

Restores a tensor from checkpoint files.

Reads a tensor stored in one or several files. If there are several files (for instance because a tensor was saved as slices), file_pattern may contain wildcard symbols (* and ?) in the filename portion only, not in the directory portion.

If a file_pattern matches several files, preferred_shard can be used to hint in which file the requested tensor is likely to be found. This op will first open the file at index preferred_shard in the list of matching files and try to restore tensors from that file. Only if some tensors or tensor slices are not found in that first file, then the Op opens all the files. Setting preferred_shard to match the value passed as the shard input of a matching Save Op may speed up Restore. This attribute only affects performance, not correctness. The default value -1 means files are processed in order.

See also RestoreSlice.

quantizedReluX Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	features
-> Tensor v2 Float	max_value
-> Tensor v3 Float	min_features: The float value that the lowest quantized value represents.
-> Tensor v4 Float	max_features: The float value that the highest quantized value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(activations, min_activations, max_activations) activations: Has the same output shape as "features". min_activations: The float value that the lowest quantized value represents. max_activations: The float value that the highest quantized value represents.

Computes Quantized Rectified Linear X: `min(max(features, 0), max_value)`

accumulatorTakeGradient Source

Arguments

:: (TensorType dtype, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` dtype)
=> Tensor Ref ByteString	handle: The handle to an accumulator.
-> Tensor v2 Int32	num_required: Number of gradients required before we return an aggregate.
-> Build (Tensor Value dtype)	average: The average of the accumulated gradients.

Extracts the average gradient in the given ConditionalAccumulator, provided

that sufficient (i.e., more than num_required) gradients have been accumulated. The op blocks until sufficient gradients have been accumulated. If the accumulator has already aggregated more than num_required gradients, it returns the average of the accumulated gradients. Also automatically increments the recorded global_step in the accumulator by 1, and resets the aggregate to 0.

floorMod Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns element-wise remainder of division. When `x < 0` xor `y < 0` is

true, this follows Python semantics in that the result here is consistent with a flooring divide. E.g. `floor(x / y) * y + mod(x, y) = x`.

NOTE*: FloorMod supports broadcasting. More about broadcasting here

matchingFiles Source

Arguments

:: Tensor v1 ByteString	pattern: A (scalar) shell wildcard pattern.
-> Tensor Value ByteString	filenames: A vector of matching filenames.

Returns the set of files matching a pattern.

Note that this routine only supports wildcard characters in the basename portion of the pattern, not in the directory portion.

maxPool Source

Arguments

:: (TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	input: 4-D input to pool over.
-> Tensor Value t	output: The max pooled output tensor.

Performs max pooling on the input.

computeAccidentalHits Source

Arguments

:: Int64	num_true: Number of true labels per context.
-> Tensor v1 Int64	true_classes: The true_classes output of UnpackSparseLabels.
-> Tensor v2 Int64	sampled_candidates: The sampled_candidates output of CandidateSampler.
-> (Tensor Value Int32, Tensor Value Int64, Tensor Value Float)	(indices, ids, weights) indices: A vector of indices corresponding to rows of true_candidates. ids: A vector of IDs of positions in sampled_candidates that match a true_label for the row with the corresponding index in indices. weights: A vector of the same length as indices and ids, in which each element is -FLOAT_MAX.

Computes the ids of the positions in sampled_candidates that match true_labels.

When doing log-odds NCE, the result of this op should be passed through a SparseToDense op, then added to the logits of the sampled candidates. This has the effect of removing the sampled labels that match the true labels by making the classifier sure that they are sampled labels.

deserializeManySparse Source

Arguments

:: TensorType dtype

=> Tensor v1 ByteString

serialized_sparse: 2-D, The N serialized SparseTensor objects. Must have 3 columns.

-> (Tensor Value Int64, Tensor Value dtype, Tensor Value Int64)

(sparse_indices, sparse_values, sparse_shape)

sparse_indices
sparse_values
sparse_shape

Deserialize and concatenate SparseTensors from a serialized minibatch.

The input serialized_sparse must be a string matrix of shape `[N x 3]` where N is the minibatch size and the rows correspond to packed outputs of SerializeSparse. The ranks of the original SparseTensor objects must all match. When the final SparseTensor is created, it has rank one higher than the ranks of the incoming SparseTensor objects (they have been concatenated along a new row dimension).

The output SparseTensor object's shape values for all dimensions but the first are the max across the input SparseTensor objects' shape values for the corresponding dimensions. Its first shape value is N, the minibatch size.

The input SparseTensor objects' indices are assumed ordered in standard lexicographic order. If this is not the case, after this step run SparseReorder to restore index ordering.

For example, if the serialized input is a `[2 x 3]` matrix representing two original SparseTensor objects:

index = [ 0] [10] [20] values = [1, 2, 3] shape = [50]

and

index = [ 2] [10] values = [4, 5] shape = [30]

then the final deserialized SparseTensor will be:

index = [0 0] [0 10] [0 20] [1 2] [1 10] values = [1, 2, 3, 4, 5] shape = [2 50]

cropAndResize Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	image: A 4-D tensor of shape `[batch, image_height, image_width, depth]`. Both `image_height` and `image_width` need to be positive.
-> Tensor v2 Float	boxes: A 2-D tensor of shape `[num_boxes, 4]`. The `i`-th row of the tensor specifies the coordinates of a box in the `box_ind[i]` image and is specified in normalized coordinates `[y1, x1, y2, x2]`. A normalized coordinate value of `y` is mapped to the image coordinate at `y * (image_height - 1)`, so as the `[0, 1]` interval of normalized image height is mapped to `[0, image_height - 1] in image height coordinates. We do allow y1 > y2, in which case the sampled crop is an up-down flipped version of the original image. The width dimension is treated similarly. Normalized coordinates outside the `[0, 1]` range are allowed, in which case we use `extrapolation_value` to extrapolate the input image values.
-> Tensor v3 Int32	box_ind: A 1-D tensor of shape `[num_boxes]` with int32 values in `[0, batch)`. The value of `box_ind[i]` specifies the image that the `i`-th box refers to.
-> Tensor v4 Int32	crop_size: A 1-D tensor of 2 elements, `size = [crop_height, crop_width]`. All cropped image patches are resized to this size. The aspect ratio of the image content is not preserved. Both `crop_height` and `crop_width` need to be positive.
-> Tensor Value Float	crops: A 4-D tensor of shape `[num_boxes, crop_height, crop_width, depth]`.

Extracts crops from the input image tensor and bilinearly resizes them (possibly

with aspect ratio change) to a common output size specified by crop_size. This is more general than the crop_to_bounding_box op which extracts a fixed size slice from the input image and does not allow resizing or aspect ratio change.

Returns a tensor with crops from the input image at positions defined at the bounding box locations in boxes. The cropped boxes are all resized (with bilinear interpolation) to a fixed `size = [crop_height, crop_width]`. The result is a 4-D tensor `[num_boxes, crop_height, crop_width, depth]`.

scatterUpdate Source

Arguments

:: (TensorType t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 tindices	indices: A tensor of indices into the first dimension of `ref`.
-> Tensor v3 t	updates: A tensor of updated values to store in `ref`.
-> Build (Tensor Ref t)	output_ref: = Same as `ref`. Returned as a convenience for operations that want to use the updated values after the update is done.

Applies sparse updates to a variable reference.

This operation computes

# Scalar indices ref[indices, ...] = updates[...]

# Vector indices (for each i) ref[indices[i], ...] = updates[i, ...]

# High rank indices (for each i, ..., j) ref[indices[i, ..., j], ...] = updates[i, ..., j, ...]

This operation outputs ref after the update is done. This makes it easier to chain operations that need to use the reset value.

If values in ref is to be updated more than once, because there are duplicate entires in indices, the order at which the updates happen for each value is undefined.

Requires `updates.shape = indices.shape + ref.shape[1:]`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/ScatterUpdate.png" alt /div

randomGamma Source

Arguments

:: (TensorType s, OneOf `[Int32, Int64]` s, TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 s	shape: 1-D integer tensor. Shape of independent samples to draw from each distribution described by the shape parameters given in alpha.
-> Tensor v2 t	alpha: A tensor in which each scalar is a "shape" parameter describing the associated gamma distribution.
-> Build (Tensor Value t)	output: A tensor with shape `shape + shape(alpha)`. Each slice `[:, ..., :, i0, i1, ...iN]` contains the samples drawn for `alpha[i0, i1, ...iN]`. The dtype of the output matches the dtype of alpha.

Outputs random values from the Gamma distribution(s) described by alpha.

This op uses the algorithm by Marsaglia et al. to acquire samples via transformation-rejection from pairs of uniform and normal random variables. See http://dl.acm.org/citation.cfm?id=358414

batchMatrixSolve Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	matrix
-> Tensor v2 t	rhs
-> Tensor Value t	output

batchMatrixBandPart Source

Arguments

:: TensorType t
=> Tensor v1 t	input
-> Tensor v2 Int64	num_lower
-> Tensor v3 Int64	num_upper
-> Tensor Value t	band

tensorArrayClose Source

Arguments

:: Tensor Ref ByteString	handle
-> Build ControlNode

all Source

Arguments

:: (TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 Bool	input: The tensor to reduce.
-> Tensor v2 tidx	reduction_indices: The dimensions to reduce.
-> Tensor Value Bool	output: The reduced tensor.

Computes the "logical and" of elements across dimensions of a tensor.

Reduces input along the dimensions given in reduction_indices. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1.

readerNumRecordsProduced Source

Arguments

:: Tensor Ref ByteString	reader_handle: Handle to a Reader.
-> Build (Tensor Value Int64)	records_produced

Returns the number of records this Reader has produced.

This is the same as the number of ReaderRead executions that have succeeded.

stackPop Source

Arguments

:: TensorType elem_type
=> Tensor Ref ByteString	handle: The handle to a stack.
-> Build (Tensor Value elem_type)	elem: The tensor that is popped from the top of the stack.

Pop the element at the top of the stack.

tensorArrayScatterV2 Source

Arguments

:: TensorType t
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 Int32	indices: The locations at which to write the tensor elements.
-> Tensor v3 t	value: The concatenated tensor to write to the TensorArray.
-> Tensor v4 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value Float	flow_out: A float scalar that enforces proper chaining of operations.

Scatter the data from the input value into specific TensorArray elements.

indices must be a vector, its length must match the first dim of value.

rGBToHSV Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	images: 1-D or higher rank. RGB data to convert. Last dimension must be size 3.
-> Tensor Value t	output: `images` converted to HSV.

Converts one or more images from RGB to HSV.

Outputs a tensor of the same shape as the images tensor, containing the HSV value of the pixels. The output is only well defined if the value in images are in `[0,1]`.

`output[..., 0]` contains hue, `output[..., 1]` contains saturation, and `output[..., 2]` contains value. All HSV values are in `[0,1]`. A hue of 0 corresponds to pure red, hue 13 is pure green, and 23 is pure blue.

serializeManySparse Source

Arguments

:: TensorType t
=> Tensor v1 Int64	sparse_indices: 2-D. The `indices` of the minibatch `SparseTensor`.
-> Tensor v2 t	sparse_values: 1-D. The `values` of the minibatch `SparseTensor`.
-> Tensor v3 Int64	sparse_shape: 1-D. The `shape` of the minibatch `SparseTensor`.
-> Tensor Value ByteString	serialized_sparse

Serialize an N-minibatch SparseTensor into an `[N, 3]` string Tensor.

The SparseTensor must have rank R greater than 1, and the first dimension is treated as the minibatch dimension. Elements of the SparseTensor must be sorted in increasing order of this first dimension. The serialized SparseTensor objects going into each row of serialized_sparse will have rank `R-1`.

The minibatch size N is extracted from `sparse_shape[0]`.

initializeTableFromTextFile Source

Arguments

:: Int64	key_index: Column index in a line to get the table `key` values from.
-> Int64	value_index: Column index that represents information of a line to get the table `value` values from.
-> Tensor Ref ByteString	table_handle: Handle to a table which will be initialized.
-> Tensor v2 ByteString	filename: Filename of a vocabulary text file.
-> Build ControlNode

Initializes a table from a text file.

It inserts one key-value pair into the table for each line of the file. The key and value is extracted from the whole line content, elements from the split line based on delimiter or the line number (starting from zero). Where to extract the key and value from a line is specified by key_index and value_index.

A value of -1 means use the line number(starting from zero), expects int64.
A value of -2 means use the whole line content, expects string.
A value >= 0 means use the index (starting at zero) of the split line based on delimiter.

decodePng Source

Arguments

:: (TensorType dtype, OneOf `[Word16, Word8]` dtype)
=> Tensor v1 ByteString	contents: 0-D. The PNG-encoded image.
-> Tensor Value dtype	image: 3-D with shape `[height, width, channels]`.

Decode a PNG-encoded image to a uint8 or uint16 tensor.

The attr channels indicates the desired number of color channels for the decoded image.

Accepted values are:

0: Use the number of channels in the PNG-encoded image.
1: output a grayscale image.
3: output an RGB image.
4: output an RGBA image.

If needed, the PNG-encoded image is transformed to match the requested number of color channels.

tensorArraySizeV2 Source

Arguments

:: Tensor v1 ByteString	handle: The handle to a TensorArray (output of TensorArray or TensorArrayGrad).
-> Tensor v2 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value Int32	size: The current size of the TensorArray.

Get the current size of the TensorArray.

div Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x / y element-wise.

NOTE*: Div supports broadcasting. More about broadcasting here

logUniformCandidateSampler Source

Arguments

:: Int64	num_sampled: Number of candidates to randomly sample per batch.
-> Int64	num_true: Number of true labels per context.
-> Int64	range_max: The sampler will sample integers from the interval [0, range_max).
-> Bool	unique: If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
-> Tensor v1 Int64	true_classes: A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.
-> (Tensor Value Int64, Tensor Value Float, Tensor Value Float)	(sampled_candidates, true_expected_count, sampled_expected_count) sampled_candidates: A vector of length num_sampled, in which each element is the ID of a sampled candidate. true_expected_count: A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability. sampled_expected_count: A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

Generates labels for candidate sampling with a log-uniform distribution.

See explanations of candidate sampling and the data formats at go/candidate-sampling.

For each batch, this op picks a single set of sampled candidate labels.

The advantages of sampling candidates per-batch are simplicity and the possibility of efficient dense matrix multiplication. The disadvantage is that the sampled candidates must be chosen independently of the context and of the true labels.

barrier Source

Arguments

:: Build (Tensor Ref ByteString)

handle: The handle to the barrier.

Defines a barrier that persists across different graph executions.

A barrier represents a key-value map, where each key is a string, and each value is a tuple of tensors.

At runtime, the barrier contains complete and incomplete elements. A complete element has defined tensors for all components of its value tuple, and may be accessed using BarrierTakeMany. An incomplete element has some undefined components in its value tuple, and may be updated using BarrierInsertMany.

createVariableOp Source

Arguments

:: TensorType dtype
=> ResourceHandle dtype	resource: handle to the resource in which to store the variable.
-> Tensor v2 dtype	value: the value to set the new tensor to use.
-> Build ControlNode

Creates a variable resource.

accumulatorApplyGradient Source

Arguments

:: (TensorType dtype, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` dtype)
=> Tensor Ref ByteString	handle: The handle to a accumulator.
-> Tensor v2 Int64	local_step: The local_step value at which the gradient was computed.
-> Tensor v3 dtype	gradient: A tensor of the gradient to be accumulated.
-> Build ControlNode

Applies a gradient to a given accumulator. Does not add if local_step is lesser

than the accumulator's global_step.

randomStandardNormal Source

Arguments

:: (TensorType dtype, OneOf `[Word16, Double, Float]` dtype, TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	shape: The shape of the output tensor.
-> Build (Tensor Value dtype)	output: A tensor of the specified shape filled with random normal values.

Outputs random values from a normal distribution.

The generated values will have mean 0 and standard deviation 1.

parameterizedTruncatedNormal Source

Arguments

:: (TensorType dtype, OneOf `[Word16, Double, Float]` dtype, TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	shape: The shape of the output tensor. Batches are indexed by the 0th dimension.
-> Tensor v2 dtype	means: The mean parameter of each batch.
-> Tensor v3 dtype	stdevs: The standard deviation parameter of each batch. Must be greater than 0.
-> Tensor v4 dtype	minvals: The minimum cutoff. May be -infinity.
-> Tensor v5 dtype	maxvals: The maximum cutoff. May be +infinity, and must be more than the minval for each batch.
-> Build (Tensor Value dtype)	output: A matrix of shape num_batches x samples_per_batch, filled with random truncated normal values using the parameters for each row.

Outputs random values from a normal distribution. The parameters may each be a

scalar which applies to the entire output, or a vector of length shape[0] which stores the parameters for each batch.

accumulatorSetGlobalStep Source

Arguments

:: Tensor Ref ByteString	handle: The handle to an accumulator.
-> Tensor v2 Int64	new_global_step: The new global_step value to set.
-> Build ControlNode

Updates the accumulator with a new value for global_step. Logs warning if the

accumulator's value is already higher than new_global_step.

resizeBilinear Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	images: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 Int32	size: = A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The new size for the images.
-> Tensor Value Float	resized_images: 4-D with shape `[batch, new_height, new_width, channels]`.

Resize images to size using bilinear interpolation.

Input images can be of different types but output images are always float.

quantizeV2 Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Word16, Word8]` t)
=> Tensor v1 Float	input
-> Tensor v2 Float	min_range: The minimum scalar value possibly produced for the input.
-> Tensor v3 Float	max_range: The maximum scalar value possibly produced for the input.
-> (Tensor Value t, Tensor Value Float, Tensor Value Float)	(output, output_min, output_max) output: The quantized data produced from the float input. output_min: The actual minimum scalar value used for the output. output_max: The actual maximum scalar value used for the output.

Quantize the input tensor of type float to output tensor of type T.

min_range, max_range: are scalar floats that specify the range for the input data. The mode attribute controls exactly which calculations are used to convert the float values to their quantized equivalents.

In MIN_COMBINED mode, each value of the tensor will undergo the following:

``` out[i] = (in[i] - min_range) * range(T) / (max_range - min_range) if T == qint8, out[i] -= (range(T) + 1) / 2.0 ``` here `range(T) = numeric_limitsT::max() - numeric_limitsT::min()`

MIN_COMBINED Mode Example*

Assume the input is type float and has a possible range of [0.0, 6.0] and the output type is quint8 ([0, 255]). The min_range and max_range values should be specified as 0.0 and 6.0. Quantizing from float to quint8 will multiply each value of the input by 255/6 and cast to quint8.

If the output type was qint8 ([-128, 127]), the operation will additionally subtract each value by 128 prior to casting, so that the range of values aligns with the range of qint8.

If the mode is MIN_FIRST, then this approach is used:

``` number_of_steps = 1 << (# of bits in T) range_adjust = number_of_steps / (number_of_steps - 1) range = (range_max - range_min) * range_adjust range_scale = number_of_steps / range quantized = round(input * range_scale) - round(range_min * range_scale) + numeric_limitsT::min() quantized = max(quantized, numeric_limitsT::min()) quantized = min(quantized, numeric_limitsT::max()) ```

The biggest difference between this and MIN_COMBINED is that the minimum range is rounded first, before it's subtracted from the rounded value. With MIN_COMBINED, a small bias is introduced where repeated iterations of quantizing and dequantizing will introduce a larger and larger error.

One thing to watch out for is that the operator may choose to adjust the requested minimum and maximum values slightly during the quantization process, so you should always use the output ports as the range for further calculations. For example, if the requested minimum and maximum values are close to equal, they will be separated by a small epsilon value to prevent ill-formed quantized buffers from being created. Otherwise, you can end up with buffers where all the quantized values map to the same float value, which causes problems for operations that have to perform further calculations on them.

decodeJpeg Source

Arguments

:: Tensor v1 ByteString	contents: 0-D. The JPEG-encoded image.
-> Tensor Value Word8	image: 3-D with shape `[height, width, channels]`..

Decode a JPEG-encoded image to a uint8 tensor.

The attr channels indicates the desired number of color channels for the decoded image.

Accepted values are:

0: Use the number of channels in the JPEG-encoded image.
1: output a grayscale image.
3: output an RGB image.

If needed, the JPEG-encoded image is transformed to match the requested number of color channels.

The attr ratio allows downscaling the image by an integer factor during decoding. Allowed values are: 1, 2, 4, and 8. This is much faster than downscaling the image later.

pow Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Int64, Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Computes the power of one value to another.

Given a tensor x and a tensor y, this operation computes $x^y$ for corresponding elements in x and y. For example:

``` # tensor x is [[2, 2]], [3, 3]] # tensor y is [[8, 16], [2, 3]] tf.pow(x, y) ==> [[256, 65536], [9, 27]] ```

loopCond Source

Arguments

:: Tensor v1 Bool	input: A boolean scalar, representing the branch predicate of the Switch op.
-> Tensor Value Bool	output: The same tensor as `input`.

Forwards the input to the output.

This operator represents the loop termination condition used by the "pivot" switches of a loop.

readFile Source

Arguments

:: Tensor v1 ByteString	filename
-> Tensor Value ByteString	contents

Reads and outputs the entire contents of the input filename.

imag Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float]` t, TensorType tout, OneOf `[Double, Float]` tout)
=> Tensor v1 t	input
-> Tensor Value tout	output

Returns the imaginary part of a complex number.

Given a tensor input of complex numbers, this operation returns a tensor of type float that is the imaginary part of each element in input. All elements in input must be complex numbers of the form $a + bj$, where *a* is the real part and *b* is the imaginary part returned by this operation.

For example:

``` # tensor input is [-2.25 + 4.75j, 3.25 + 5.75j] tf.imag(input) ==> [4.75, 5.75] ```

tensorArrayGrad Source

Arguments

:: Tensor v1 ByteString	handle
-> Tensor v2 Float	flow_in
-> Build (Tensor Ref ByteString)	grad_handle

histogramSummary Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 ByteString	tag: Scalar. Tag to use for the `Value`.
-> Tensor v2 t	values: Any shape. Values to use to build the histogram.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Outputs a Summary protocol buffer with a histogram.

The generated `Summary` has one summary value containing a histogram for values.

This op reports an InvalidArgument error if any value is not finite.

conv3DBackpropInputV2 Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int32	input_sizes: An integer vector representing the tensor shape of `input`, where `input` is a 5-D `[batch, depth, rows, cols, in_channels]` tensor.
-> Tensor v2 t	filter: Shape `[depth, rows, cols, in_channels, out_channels]`. `in_channels` must match between `input` and `filter`.
-> Tensor v3 t	out_backprop: Backprop signal of shape `[batch, out_depth, out_rows, out_cols, out_channels]`.
-> Tensor Value t	output

Computes the gradients of 3-D convolution with respect to the input.

resizeBilinearGrad Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 Float	grads: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 t	original_image: 4-D with shape `[batch, orig_height, orig_width, channels]`, The image tensor that was resized.
-> Tensor Value t	output: 4-D with shape `[batch, orig_height, orig_width, channels]`. Gradients with respect to the input image. Input image must have been float or double.

Computes the gradient of bilinear interpolation.

addManySparseToTensorsMap Source

Arguments

:: TensorType t
=> Tensor v1 Int64	sparse_indices: 2-D. The `indices` of the minibatch `SparseTensor`. `sparse_indices[:, 0]` must be ordered values in `[0, N)`.
-> Tensor v2 t	sparse_values: 1-D. The `values` of the minibatch `SparseTensor`.
-> Tensor v3 Int64	sparse_shape: 1-D. The `shape` of the minibatch `SparseTensor`. The minibatch size `N == sparse_shape[0]`.
-> Build (Tensor Value Int64)	sparse_handles: 1-D. The handles of the `SparseTensor` now stored in the `SparseTensorsMap`. Shape: `[N]`.

Add an N-minibatch SparseTensor to a SparseTensorsMap, return N handles.

A SparseTensor of rank R is represented by three tensors: sparse_indices, sparse_values, and sparse_shape, where

```sparse_indices.shape[1] == sparse_shape.shape[0] == R```

An N-minibatch of SparseTensor objects is represented as a SparseTensor having a first sparse_indices column taking values between `[0, N)`, where the minibatch size `N == sparse_shape[0]`.

The input SparseTensor must have rank R greater than 1, and the first dimension is treated as the minibatch dimension. Elements of the SparseTensor must be sorted in increasing order of this first dimension. The stored SparseTensor objects pointed to by each row of the output sparse_handles will have rank `R-1`.

The SparseTensor values can then be read out as part of a minibatch by passing the given keys as vector elements to TakeManySparseFromTensorsMap. To ensure the correct SparseTensorsMap is accessed, ensure that the same container and shared_name are passed to that Op. If no shared_name is provided here, instead use the *name* of the Operation created by calling AddManySparseToTensorsMap as the shared_name passed to TakeManySparseFromTensorsMap. Ensure the Operations are colocated.

batchIFFT Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

batchMatrixDeterminant Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t)
=> Tensor v1 t	input
-> Tensor Value t	output

deleteSessionTensor Source

Arguments

:: Tensor v1 ByteString	handle: The handle for a tensor stored in the session state.
-> ControlNode

Delete the tensor specified by its handle in the session.

lookupTableSize Source

Arguments

:: Tensor Ref ByteString	table_handle: Handle to the table.
-> Build (Tensor Value Int64)	size: Scalar that contains number of elements in the table.

Computes the number of elements in the given table.

relu Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	features
-> Tensor Value t	activations

Computes rectified linear: `max(features, 0)`.

dynamicStitch Source

Arguments

:: TensorType t
=> [Tensor v1 Int32]	indices
-> [Tensor v2 t]	data
-> Tensor Value t	merged

Interleave the values from the `data` tensors into a single tensor.

Builds a merged tensor such that

```python merged[indices[m][i, ..., j], ...] = data[m][i, ..., j, ...] ```

For example, if each `indices[m]` is scalar or vector, we have

```python # Scalar indices: merged[indices[m], ...] = data[m][...]

# Vector indices: merged[indices[m][i], ...] = data[m][i, ...] ```

Each `data[i].shape` must start with the corresponding `indices[i].shape`, and the rest of `data[i].shape` must be constant w.r.t. i. That is, we must have `data[i].shape = indices[i].shape + constant`. In terms of this constant, the output shape is

merged.shape = [max(indices)] + constant

Values are merged in order, so if an index appears in both `indices[m][i]` and `indices[n][j]` for `(m,i) < (n,j)` the slice `data[n][j]` will appear in the merged result.

For example:

```python indices[0] = 6 indices[1] = [4, 1] indices[2] = [[5, 2], [0, 3]] data[0] = [61, 62] data[1] = [[41, 42], [11, 12]] data[2] = [[[51, 52], [21, 22]], [[1, 2], [31, 32]]] merged = [[1, 2], [11, 12], [21, 22], [31, 32], [41, 42], [51, 52], [61, 62]] ```

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/DynamicStitch.png" alt /div

lookupTableFind Source

Arguments

:: (TensorType tin, TensorType tout)
=> Tensor Ref ByteString	table_handle: Handle to the table.
-> Tensor v2 tin	keys: Any shape. Keys to look up.
-> Tensor v3 tout	default_value
-> Build (Tensor Value tout)	values: Same shape as `keys`. Values found in the table, or `default_values` for missing keys.

Looks up keys in a table, outputs the corresponding values.

The tensor keys must of the same type as the keys of the table. The output values is of the type of the table values.

The scalar default_value is the value output for keys not present in the table. It must also be of the same type as the table values.

sampleDistortedBoundingBox Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word8]` t)
=> Tensor v1 t	image_size: 1-D, containing `[height, width, channels]`.
-> Tensor v2 Float	bounding_boxes: 3-D with shape `[batch, N, 4]` describing the N bounding boxes associated with the image.
-> Build (Tensor Value t, Tensor Value t, Tensor Value Float)	(begin, size, bboxes) begin: 1-D, containing `[offset_height, offset_width, 0]`. Provide as input to `tf.slice`. size: 1-D, containing `[target_height, target_width, -1]`. Provide as input to `tf.slice`. bboxes: 3-D with shape `[1, 1, 4]` containing the distorted bounding box. Provide as input to `tf.image.draw_bounding_boxes`.

Generate a single randomly distorted bounding box for an image.

Bounding box annotations are often supplied in addition to ground-truth labels in image recognition or object localization tasks. A common technique for training such a system is to randomly distort an image while preserving its content, i.e. *data augmentation*. This Op outputs a randomly distorted localization of an object, i.e. bounding box, given an image_size, bounding_boxes and a series of constraints.

The output of this Op is a single bounding box that may be used to crop the original image. The output is returned as 3 tensors: begin, size and bboxes. The first 2 tensors can be fed directly into `tf.slice` to crop the image. The latter may be supplied to `tf.image.draw_bounding_boxes` to visualize what the bounding box looks like.

Bounding boxes are supplied and returned as `[y_min, x_min, y_max, x_max]`. The bounding box coordinates are floats in `[0.0, 1.0]` relative to the width and height of the underlying image.

For example,

```python # Generate a single distorted bounding box. begin, size, bbox_for_draw = tf.image.sample_distorted_bounding_box( tf.shape(image), bounding_boxes=bounding_boxes)

# Draw the bounding box in an image summary. image_with_box = tf.image.draw_bounding_boxes(tf.expand_dims(image, 0), bbox_for_draw) tf.image_summary(images_with_box, image_with_box)

# Employ the bounding box to distort the image. distorted_image = tf.slice(image, begin, size) ```

Note that if no bounding box information is available, setting `use_image_if_no_bounding_boxes = true` will assume there is a single implicit bounding box covering the whole image. If use_image_if_no_bounding_boxes is false and no bounding boxes are supplied, an error is raised.

splitV Source

Arguments

:: (TensorType t, TensorType tlen, OneOf `[Int32, Int64]` tlen)
=> Int64	num_split
-> Tensor v1 t	value: The tensor to split.
-> Tensor v2 tlen	size_splits: list containing the sizes of each output tensor along the split dimension. Must sum to the dimension of value along split_dim. Can contain one -1 indicating that dimension is to be inferred.
-> Tensor v3 Int32	split_dim: 0-D. The dimension along which to split. Must be in the range `[0, rank(value))`.
-> [Tensor Value t]	output: Tensors whose shape matches that of `value` except along `split_dim`, where their sizes are `size_splits[i]`.

Splits a tensor into num_split tensors along one dimension.

fusedPadConv2D Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	input: 4-D with shape `[batch, in_height, in_width, in_channels]`.
-> Tensor v2 Int32	paddings: A two-column matrix specifying the padding sizes. The number of rows must be the same as the rank of `input`.
-> Tensor v3 t	filter: 4-D with shape `[filter_height, filter_width, in_channels, out_channels]`.
-> Tensor Value t	output

Performs a padding as a preprocess during a convolution.

Similar to FusedResizeAndPadConv2d, this op allows for an optimized implementation where the spatial padding transformation stage is fused with the im2col lookup, but in this case without the bilinear filtering required for resizing. Fusing the padding prevents the need to write out the intermediate results as whole tensors, reducing memory pressure, and we can get some latency gains by merging the transformation calculations. The data_format attribute for Conv2D isn't supported by this op, and NHWC order is used instead. Internally this op uses a single per-graph scratch buffer, which means that it will block if multiple versions are being run in parallel. This is because this operator is primarily an optimization to minimize memory usage.

barrierInsertMany Source

Arguments

:: TensorType t
=> Int64	component_index: The component of the barrier elements that is being assigned.
-> Tensor Ref ByteString	handle: The handle to a barrier.
-> Tensor v2 ByteString	keys: A one-dimensional tensor of keys, with length n.
-> Tensor v3 t	values: An any-dimensional tensor of values, which are associated with the respective keys. The 0th dimension must have length n.
-> Build ControlNode

For each key, assigns the respective value to the specified component.

If a key is not found in the barrier, this operation will create a new incomplete element. If a key is found in the barrier, and the element already has a value at component_index, this operation will fail with INVALID_ARGUMENT, and leave the barrier in an undefined state.

abort :: ControlNode Source

Raise a exception to abort the process when called.

Returns nothing but an exception.

maxPoolWithArgmax Source

Arguments

:: (TensorType targmax, OneOf `[Int32, Int64]` targmax, TensorType t, OneOf `[Word16, Float]` t)

=> Tensor v1 t

input: 4-D with shape `[batch, height, width, channels]`. Input to pool over.

-> (Tensor Value t, Tensor Value targmax)

(output, argmax)

output: The max pooled output tensor.
argmax: 4-D. The flattened indices of the max values chosen for each output.

Performs max pooling on the input and outputs both max values and indices.

The indices in argmax are flattened, so that a maximum value at position `[b, y, x, c]` becomes flattened index `((b * height + y) * width + x) * channels + c`.

refEnter Source

Arguments

:: TensorType t
=> Tensor Ref t	data: The tensor to be made available to the child frame.
-> Build (Tensor Ref t)	output: The same tensor as `data`.

Creates or finds a child frame, and makes `data` available to the child frame.

The unique frame_name is used by the Executor to identify frames. If is_constant is true, output is a constant in the child frame; otherwise it may be changed in the child frame. At most parallel_iterations iterations are run in parallel in the child frame.

dequantize Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Word16, Word8]` t)
=> Tensor v1 t	input
-> Tensor v2 Float	min_range: The minimum scalar value possibly produced for the input.
-> Tensor v3 Float	max_range: The maximum scalar value possibly produced for the input.
-> Tensor Value Float	output

Dequantize the input tensor into a float Tensor.

min_range, max_range: are scalar floats that specify the range for the input data. The mode attribute controls exactly which calculations are used to convert the float values to their quantized equivalents.

In MIN_COMBINED mode, each value of the tensor will undergo the following:

``` if T == qint8, in[i] += (range(T) + 1)/ 2.0 out[i] = min_range + (in[i]* (max_range - min_range) / range(T)) ``` here `range(T) = numeric_limitsT::max() - numeric_limitsT::min()`

MIN_COMBINED Mode Example*

If the input comes from a QuantizedRelu6, the output type is quint8 (range of 0-255) but the possible range of QuantizedRelu6 is 0-6. The min_range and max_range values are therefore 0.0 and 6.0. Dequantize on quint8 will take each value, cast to float, and multiply by 6 / 255. Note that if quantizedtype is qint8, the operation will additionally add each value by 128 prior to casting.

If the mode is MIN_FIRST, then this approach is used:

``` number_of_steps = 1 << (# of bits in T) range_adjust = number_of_steps / (number_of_steps - 1) range = (range_max - range_min) * range_adjust range_scale = range / number_of_steps const double offset_input = static_castdouble(input) - lowest_quantized; result = range_min + ((input - numeric_limitsT::min()) * range_scale) ```

drawBoundingBoxes Source

Arguments

:: (TensorType t, OneOf `[Word16, Float]` t)
=> Tensor v1 t	images: 4-D with shape `[batch, height, width, depth]`. A batch of images.
-> Tensor v2 Float	boxes: 3-D with shape `[batch, num_bounding_boxes, 4]` containing bounding boxes.
-> Tensor Value t	output: 4-D with the same shape as `images`. The batch of input images with bounding boxes drawn on the images.

Draw bounding boxes on a batch of images.

Outputs a copy of images but draws on top of the pixels zero or more bounding boxes specified by the locations in boxes. The coordinates of the each bounding box in boxes are encoded as `[y_min, x_min, y_max, x_max]`. The bounding box coordinates are floats in `[0.0, 1.0]` relative to the width and height of the underlying image.

For example, if an image is 100 x 200 pixels and the bounding box is `[0.1, 0.2, 0.5, 0.9]`, the bottom-left and upper-right coordinates of the bounding box will be `(10, 40)` to `(50, 180)`.

Parts of the bounding box may fall outside the image.

tensorArraySplit Source

Arguments

:: TensorType t
=> Tensor Ref ByteString	handle
-> Tensor v2 t	value
-> Tensor v3 Int64	lengths
-> Tensor v4 Float	flow_in
-> Build (Tensor Value Float)	flow_out

stringToHashBucketFast Source

Arguments

:: Int64	num_buckets: The number of buckets.
-> Tensor v1 ByteString	input: The strings to assign a hash bucket.
-> Tensor Value Int64	output: A Tensor of the same shape as the input `string_tensor`.

Converts each string in the input Tensor to its hash mod by a number of buckets.

The hash function is deterministic on the content of the string within the process and will never change. However, it is not suitable for cryptography. This function may be used when CPU time is scarce and inputs are trusted or unimportant. There is a risk of adversaries constructing inputs that all hash to the same bucket. To prevent this problem, use a strong hash function with `tf.string_to_hash_bucket_strong`.

tensorArrayScatter Source

Arguments

:: TensorType t
=> Tensor Ref ByteString	handle
-> Tensor v2 Int32	indices
-> Tensor v3 t	value
-> Tensor v4 Float	flow_in
-> Build (Tensor Value Float)	flow_out

oneHot Source

Arguments

:: (TensorType t, TensorType tI, OneOf `[Int32, Int64, Word8]` tI)
=> Tensor v1 tI	indices: A tensor of indices.
-> Tensor v2 Int32	depth: A scalar defining the depth of the one hot dimension.
-> Tensor v3 t	on_value: A scalar defining the value to fill in output when `indices[j] = i`.
-> Tensor v4 t	off_value: A scalar defining the value to fill in output when `indices[j] != i`.
-> Tensor Value t	output: The one-hot tensor.

Returns a one-hot tensor.

The locations represented by indices in indices take value on_value, while all other locations take value off_value.

If the input indices is rank N, the output will have rank `N+1`, The new axis is created at dimension axis (default: the new axis is appended at the end).

If indices is a scalar the output shape will be a vector of length depth.

If indices is a vector of length features, the output shape will be: ``` features x depth if axis == -1 depth x features if axis == 0 ```

If indices is a matrix (batch) with shape `[batch, features]`, the output shape will be: ``` batch x features x depth if axis == -1 batch x depth x features if axis == 1 depth x batch x features if axis == 0 ```

Examples =========

Suppose that

``` indices = [0, 2, -1, 1] depth = 3 on_value = 5.0 off_value = 0.0 axis = -1 ```

Then output is `[4 x 3]`:

```output = [5.0 0.0 0.0] // one_hot(0) [0.0 0.0 5.0] // one_hot(2) [0.0 0.0 0.0] // one_hot(-1) [0.0 5.0 0.0] // one_hot(1) ```

Suppose that

``` indices = [0, 2, -1, 1] depth = 3 on_value = 0.0 off_value = 3.0 axis = 0 ```

Then output is `[3 x 4]`:

```output = [0.0 3.0 3.0 3.0] [3.0 3.0 3.0 0.0] [3.0 3.0 3.0 3.0] [3.0 0.0 3.0 3.0] // ^ one_hot(0) // ^ one_hot(2) // ^ one_hot(-1) // ^ one_hot(1) ``` Suppose that

``` indices = [[0, 2], [1, -1]] depth = 3 on_value = 1.0 off_value = 0.0 axis = -1 ```

Then output is `[2 x 2 x 3]`:

```output = [ [1.0, 0.0, 0.0] // one_hot(0) [0.0, 0.0, 1.0] // one_hot(2) ][ [0.0, 1.0, 0.0] // one_hot(1) [0.0, 0.0, 0.0] // one_hot(-1) ]```

batchIFFT3D Source

Arguments

:: Tensor v1 (Complex Float)	input
-> Tensor Value (Complex Float)	output

decodeRaw Source

Arguments

:: (TensorType out_type, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` out_type)
=> Tensor v1 ByteString	bytes: All the elements must have the same length.
-> Tensor Value out_type	output: A Tensor with one more dimension than the input `bytes`. The added dimension will have size equal to the length of the elements of `bytes` divided by the number of bytes to represent `out_type`.

Reinterpret the bytes of a string as a vector of numbers.

tensorArrayPack Source

Arguments

:: TensorType dtype
=> Tensor Ref ByteString	handle
-> Tensor v2 Float	flow_in
-> Build (Tensor Value dtype)	value

applyProximalAdagrad Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor v3 t	lr: Scaling factor. Must be a scalar.
-> Tensor v4 t	l1: L1 regularization. Must be a scalar.
-> Tensor v5 t	l2: L2 regularization. Must be a scalar.
-> Tensor v6 t	grad: The gradient.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' and '*accum' according to FOBOS with Adagrad learning rate.

accum += grad * grad prox_v = var - lr * grad * (1 / sqrt(accum)) var = sign(prox_v)/(1+lr*l2) * max{|prox_v|-lr*l1,0}

sparseAccumulatorApplyGradient Source

Arguments

:: (TensorType dtype, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` dtype)
=> Bool	has_known_shape: Boolean indicating whether gradient_shape is unknown, in which case the input is ignored during validation.
-> Tensor Ref ByteString	handle: The handle to a accumulator.
-> Tensor v2 Int64	local_step: The local_step value at which the sparse gradient was computed.
-> Tensor v3 Int64	gradient_indices: Indices of the sparse gradient to be accumulated. Must be a vector.
-> Tensor v4 dtype	gradient_values: Values are the non-zero slices of the gradient, and must have the same first dimension as indices, i.e., the nnz represented by indices and values must be consistent.
-> Tensor v5 Int64	gradient_shape: Shape of the sparse gradient to be accumulated.
-> Build ControlNode

Applies a sparse gradient to a given accumulator. Does not add if local_step is

lesser than the accumulator's global_step.

add Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, ByteString, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x + y element-wise.

NOTE*: Add supports broadcasting. AddN does not. More about broadcasting here

softsign Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	features
-> Tensor Value t	activations

Computes softsign: `features / (abs(features) + 1)`.

tensorArrayRead Source

Arguments

:: TensorType dtype
=> Tensor Ref ByteString	handle
-> Tensor v2 Int32	index
-> Tensor v3 Float	flow_in
-> Build (Tensor Value dtype)	value

scatterNdSub Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: A mutable Tensor. Should be from a Variable node.
-> Tensor v2 tindices	indices: A Tensor. Must be one of the following types: int32, int64. A tensor of indices into ref.
-> Tensor v3 t	updates: A Tensor. Must have the same type as ref. A tensor of updated values to subtract from ref.
-> Build (Tensor Ref t)	output_ref: Same as ref. Returned as a convenience for operations that want to use the updated values after the update is done.

Applies sparse subtraction between updates and individual values or slices

within a given variable according to indices.

ref is a Tensor with rank P and indices is a Tensor of rank Q.

indices must be integer tensor, containing indices into ref. It must be shape `[d_0, ..., d_{Q-2}, K]` where `0 < K <= P`.

The innermost dimension of indices (with length K) corresponds to indices into elements (if `K = P`) or slices (if `K < P`) along the Kth dimension of ref.

updates is Tensor of rank `Q-1+P-K` with shape:

``` [d_0, ..., d_{Q-2}, ref.shape[K], ..., ref.shape[P-1]]. ```

For example, say we want to subtract 4 scattered elements from a rank-1 tensor with 8 elements. In Python, that subtraction would look like this:

ref = tf.Variable([1, 2, 3, 4, 5, 6, 7, 8]) indices = tf.constant([[4], [3], [1], [7]]) updates = tf.constant([9, 10, 11, 12]) sub = tf.scatter_nd_sub(ref, indices, updates) with tf.Session() as sess: print sess.run(sub)

The resulting update to ref would look like this:

1, -9, 3, -6, -4, 6, 7, -4

See tf.scatter_nd for more details about how to make updates to slices.

restoreSlice Source

Arguments

:: TensorType dt
=> Tensor v1 ByteString	file_pattern: Must have a single element. The pattern of the files from which we read the tensor.
-> Tensor v2 ByteString	tensor_name: Must have a single element. The name of the tensor to be restored.
-> Tensor v3 ByteString	shape_and_slice: Scalar. The shapes and slice specifications to use when restoring a tensors.
-> Tensor Value dt	tensor: The restored tensor.

Restores a tensor from checkpoint files.

This is like Restore except that restored tensor can be listed as filling only a slice of a larger tensor. shape_and_slice specifies the shape of the larger tensor and the slice that the restored tensor covers.

The shape_and_slice input has the same format as the elements of the shapes_and_slices input of the SaveSlices op.

assignAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	ref: Should be from a `Variable` node.
-> Tensor v2 t	value: The value to be added to the variable.
-> Build (Tensor Ref t)	output_ref: = Same as "ref". Returned as a convenience for operations that want to use the new value after the variable has been updated.

Update ref by adding value to it.

This operation outputs "ref" after the update is done. This makes it easier to chain operations that need to use the reset value.

greater Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value Bool	z

Returns the truth value of (x > y) element-wise.

NOTE*: Greater supports broadcasting. More about broadcasting here

readerNumWorkUnitsCompleted Source

Arguments

:: Tensor Ref ByteString	reader_handle: Handle to a Reader.
-> Build (Tensor Value Int64)	units_completed

Returns the number of work units this Reader has finished processing.

tensorArrayGatherV2 Source

Arguments

:: TensorType dtype
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 Int32	indices: The locations in the TensorArray from which to read tensor elements.
-> Tensor v3 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value dtype	value: All of the elements in the TensorArray, concatenated along a new axis (the new dimension 0).

Gather specific elements from the TensorArray into output value.

All elements selected by indices must have the same shape.

tensorArrayReadV2 Source

Arguments

:: TensorType dtype
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 Int32	index
-> Tensor v3 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value dtype	value: The tensor that is read from the TensorArray.

Read an element from the TensorArray into output value.

decodeBase64 Source

Arguments

:: Tensor v1 ByteString	input: Base64 strings to decode.
-> Tensor Value ByteString	output: Decoded strings.

Decode web-safe base64-encoded strings.

Input may or may not have padding at the end. See EncodeBase64 for padding. Web-safe means that input must use - and _ instead of + and /.

tensorArrayWriteV2 Source

Arguments

:: TensorType t
=> Tensor v1 ByteString	handle: The handle to a TensorArray.
-> Tensor v2 Int32	index: The position to write to inside the TensorArray.
-> Tensor v3 t	value: The tensor to write to the TensorArray.
-> Tensor v4 Float	flow_in: A float scalar that enforces proper chaining of operations.
-> Tensor Value Float	flow_out: A float scalar that enforces proper chaining of operations.

Push an element onto the tensor_array.

audioSummary Source

Arguments

:: Float	sample_rate: The sample rate of the signal in hertz.
-> Tensor v1 ByteString	tag: Scalar. Used to build the `tag` attribute of the summary values.
-> Tensor v2 Float	tensor: 2-D of shape `[batch_size, frames]`.
-> Tensor Value ByteString	summary: Scalar. Serialized `Summary` protocol buffer.

Outputs a Summary protocol buffer with audio.

The summary has up to max_outputs summary values containing audio. The audio is built from tensor which must be 3-D with shape `[batch_size, frames, channels]` or 2-D with shape `[batch_size, frames]`. The values are assumed to be in the range of `[-1.0, 1.0]` with a sample rate of sample_rate.

The tag argument is a scalar Tensor of type string. It is used to build the tag of the summary values:

If max_outputs is 1, the summary value tag is '*tag*/audio'.
If max_outputs is greater than 1, the summary value tags are generated sequentially as '*tag*/audio/0', '*tag*/audio/1', etc.

isFinite Source

Arguments

:: (TensorType t, OneOf `[Word16, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor Value Bool	y

Returns which elements of x are finite.

compatibility(numpy) Equivalent to np.isfiniteend_compatibility

tensorArrayConcat Source

Arguments

:: TensorType dtype
=> Tensor Ref ByteString	handle
-> Tensor v2 Float	flow_in
-> Build (Tensor Value dtype, Tensor Value Int64)	(value, lengths) value lengths

sparseReduceSum Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 Int64	input_indices: 2-D. `N x R` matrix with the indices of non-empty values in a SparseTensor, possibly not in canonical ordering.
-> Tensor v2 t	input_values: 1-D. `N` non-empty values corresponding to `input_indices`.
-> Tensor v3 Int64	input_shape: 1-D. Shape of the input SparseTensor.
-> Tensor v4 Int32	reduction_axes: 1-D. Length-`K` vector containing the reduction axes.
-> Tensor Value t	output: `R-K`-D. The reduced Tensor.

Computes the sum of elements across dimensions of a SparseTensor.

This Op takes a SparseTensor and is the sparse counterpart to `tf.reduce_sum()`. In particular, this Op also returns a dense Tensor instead of a sparse one.

Reduces sp_input along the dimensions given in reduction_axes. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_axes. If keep_dims is true, the reduced dimensions are retained with length 1.

If reduction_axes has no entries, all dimensions are reduced, and a tensor with a single element is returned. Additionally, the axes can be negative, which are interpreted according to the indexing rules in Python.

realDiv Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns x / y element-wise for real types.

If x and y are reals, this will return the floating-point division.

NOTE*: Div supports broadcasting. More about broadcasting here

tensorArraySize Source

Arguments

:: Tensor Ref ByteString	handle
-> Tensor v2 Float	flow_in
-> Build (Tensor Value Int32)	size

biasAddV1 Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	value: Any number of dimensions.
-> Tensor v2 t	bias: 1-D with size the last dimension of `value`.
-> Tensor Value t	output: Broadcasted sum of `value` and `bias`.

Adds bias to value.

This is a deprecated version of BiasAdd and will be soon removed.

This is a special case of `tf.add` where bias is restricted to be 1-D. Broadcasting is supported, so value may have any number of dimensions.

logicalOr Source

Arguments

:: Tensor v1 Bool	x
-> Tensor v2 Bool	y
-> Tensor Value Bool	z

Returns the truth value of x OR y element-wise.

NOTE*: LogicalOr supports broadcasting. More about broadcasting here

stackPush Source

Arguments

:: TensorType t
=> Tensor Ref ByteString	handle: The handle to a stack.
-> Tensor v2 t	elem: The tensor to be pushed onto the stack.
-> Build (Tensor Value t)	output: The same tensor as the input `elem`.

Push an element onto the stack.

quantizedRelu Source

Arguments

:: (TensorType tinput, OneOf `[Int16, Int32, Word16, Word8]` tinput, TensorType out_type, OneOf `[Int16, Int32, Word16, Word8]` out_type)
=> Tensor v1 tinput	features
-> Tensor v2 Float	min_features: The float value that the lowest quantized value represents.
-> Tensor v3 Float	max_features: The float value that the highest quantized value represents.
-> (Tensor Value out_type, Tensor Value Float, Tensor Value Float)	(activations, min_activations, max_activations) activations: Has the same output shape as "features". min_activations: The float value that the lowest quantized value represents. max_activations: The float value that the highest quantized value represents.

Computes Quantized Rectified Linear: `max(features, 0)`

broadcastGradientArgs Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 t	s0
-> Tensor v2 t	s1
-> (Tensor Value t, Tensor Value t)	(r0, r1) r0 r1

Return the reduction indices for computing gradients of s0 op s1 with broadcast.

This is typically used by gradient computations for a broadcasting operation.

uniqueWithCounts Source

Arguments

:: (TensorType t, TensorType out_idx, OneOf `[Int32, Int64]` out_idx)

=> Tensor v1 t

x: 1-D.

-> (Tensor Value t, Tensor Value out_idx, Tensor Value out_idx)

(y, idx, count)

y: 1-D.
idx: 1-D.
count: 1-D.

Finds unique elements in a 1-D tensor.

This operation returns a tensor y containing all of the unique elements of x sorted in the same order that they occur in x. This operation also returns a tensor idx the same size as x that contains the index of each value of x in the unique output y. Finally, it returns a third tensor count that contains the count of each element of y in x. In other words:

`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`

For example:

```prettyprint # tensor x is [1, 1, 2, 4, 4, 4, 7, 8, 8] y, idx, count = unique_with_counts(x) y ==> [1, 2, 4, 7, 8] idx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4] count ==> [2, 1, 3, 1, 2] ```

truncateMod Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)
=> Tensor v1 t	x
-> Tensor v2 t	y
-> Tensor Value t	z

Returns element-wise remainder of division. This emulates C semantics where

true, this follows C semantics in that the result here is consistent with a flooring divide. E.g. `floor(x / y) * y + mod(x, y) = x`.

NOTE*: Mod supports broadcasting. More about broadcasting here

stridedSliceGrad Source

Arguments

:: (TensorType t, TensorType index, OneOf `[Int32, Int64]` index)
=> Tensor v1 index	shape
-> Tensor v2 index	begin
-> Tensor v3 index	end
-> Tensor v4 index	strides
-> Tensor v5 t	dy
-> Tensor Value t	output

Returns the gradient of StridedSlice.

Since StridedSlice cuts out pieces of its input which is size shape, its gradient will have the same shape (which is passed here as shape). The gradient will be zero in any element that the slice does not select.

Arguments are the same as StridedSliceGrad with the exception that dy is the input gradient to be propagated and shape is the shape of StridedSlice's input.

fractionalAvgPool Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64, Double, Float]` t)

=> Tensor v1 t

value: 4-D with shape `[batch, height, width, channels]`.

-> (Tensor Value t, Tensor Value Int64, Tensor Value Int64)

(output, row_pooling_sequence, col_pooling_sequence)

output: output tensor after fractional avg pooling.
row_pooling_sequence: row pooling sequence, needed to calculate gradient.
col_pooling_sequence: column pooling sequence, needed to calculate gradient.

Performs fractional average pooling on the input.

Fractional average pooling is similar to Fractional max pooling in the pooling region generation step. The only difference is that after pooling regions are generated, a mean operation is performed instead of a max operation in each pooling region.

sparseAccumulatorTakeGradient Source

Arguments

:: (TensorType dtype, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` dtype)
=> Tensor Ref ByteString	handle: The handle to a SparseConditionalAccumulator.
-> Tensor v2 Int32	num_required: Number of gradients required before we return an aggregate.
-> Build (Tensor Value Int64, Tensor Value dtype, Tensor Value Int64)	(indices, values, shape) indices: Indices of the average of the accumulated sparse gradients. values: Values of the average of the accumulated sparse gradients. shape: Shape of the average of the accumulated sparse gradients.

Extracts the average sparse gradient in the given SparseConditionalAccumulator,

provided that sufficient (i.e., more than num_required) gradients have been accumulated. The op will blocks until sufficient gradients have been accumulated. If the accumulator has already aggregated more than num_required gradients, it will return its average of the accumulated gradients. Also automatically increments the recorded global_step in the accumulator by 1, and resets the aggregate to 0.

decodeJSONExample Source

Arguments

:: Tensor v1 ByteString	json_examples: Each string is a JSON object serialized according to the JSON mapping of the Example proto.
-> Tensor Value ByteString	binary_examples: Each string is a binary Example protocol buffer corresponding to the respective element of `json_examples`.

Convert JSON-encoded Example records to binary protocol buffer strings.

This op translates a tensor containing Example records, encoded using the standard JSON mapping, into a tensor containing the same records encoded as binary protocol buffers. The resulting tensor can then be fed to any of the other Example-parsing ops.

placeholderWithDefault Source

Arguments

:: TensorType dtype
=> Shape	shape: The (possibly partial) shape of the tensor.
-> Tensor v1 dtype	input: The default value to produce when `output` is not fed.
-> Tensor Value dtype	output: A placeholder tensor that defaults to `input` if it is not fed.

A placeholder op that passes though input when its output is not fed.

applyFtrl Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor Ref t	var: Should be from a Variable().
-> Tensor Ref t	accum: Should be from a Variable().
-> Tensor Ref t	linear: Should be from a Variable().
-> Tensor v4 t	grad: The gradient.
-> Tensor v5 t	lr: Scaling factor. Must be a scalar.
-> Tensor v6 t	l1: L1 regulariation. Must be a scalar.
-> Tensor v7 t	l2: L2 regulariation. Must be a scalar.
-> Tensor v8 t	lr_power: Scaling factor. Must be a scalar.
-> Build (Tensor Ref t)	out: Same as "var".

Update '*var' according to the Ftrl-proximal scheme.

accum_new = accum + grad * grad linear += grad + (accum_new^(-lr_power) - accum^(-lr_power)) / lr * var quadratic = 1.0 / (accum_new^(lr_power) * lr) + 2 * l2 var = (sign(linear) * l1 - linear) / quadratic if |linear| > l1 else 0.0 accum = accum_new

sdcaShrinkL1 Source

Arguments

:: Float	l1: Symmetric l1 regularization strength.
-> Float	l2: Symmetric l2 regularization strength. Should be a positive float.
-> [Tensor Ref Float]	weights: a list of vectors where each value is the weight associated with a feature group.
-> Build ControlNode

Applies L1 regularization shrink step on the parameters.

shardedFilename Source

Arguments

:: Tensor v1 ByteString	basename
-> Tensor v2 Int32	shard
-> Tensor v3 Int32	num_shards
-> Tensor Value ByteString	filename

Generate a sharded filename. The filename is printf formatted as

%s-%05d-of-%05d, basename, shard, num_shards.

fakeQuantWithMinMaxArgs Source

Arguments

:: Tensor v1 Float	inputs
-> Tensor Value Float	outputs

Fake-quantize the inputs tensor, type float to outputs tensor of same type.

Attributes [min; max] define the clamping range for the inputs data. Op divides this range into 255 steps (total of 256 values), then replaces each inputs value with the closest of the quantized step values.

Quantization is called fake since the output is still in floating point.

scatterNdAdd Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor Ref t	ref: A mutable Tensor. Should be from a Variable node.
-> Tensor v2 tindices	indices: A Tensor. Must be one of the following types: int32, int64. A tensor of indices into ref.
-> Tensor v3 t	updates: A Tensor. Must have the same type as ref. A tensor of updated values to add to ref.
-> Build (Tensor Ref t)	output_ref: Same as ref. Returned as a convenience for operations that want to use the updated values after the update is done.

Applies sparse addition between updates and individual values or slices

within a given variable according to indices.

ref is a Tensor with rank P and indices is a Tensor of rank Q.

indices must be integer tensor, containing indices into ref. It must be shape `[d_0, ..., d_{Q-2}, K]` where `0 < K <= P`.

The innermost dimension of indices (with length K) corresponds to indices into elements (if `K = P`) or slices (if `K < P`) along the Kth dimension of ref.

updates is Tensor of rank `Q-1+P-K` with shape:

``` [d_0, ..., d_{Q-2}, ref.shape[K], ..., ref.shape[P-1]]. ```

For example, say we want to add 4 scattered elements to a rank-1 tensor to 8 elements. In Python, that addition would look like this:

ref = tf.Variable([1, 2, 3, 4, 5, 6, 7, 8]) indices = tf.constant([[4], [3], [1], [7]]) updates = tf.constant([9, 10, 11, 12]) add = tf.scatter_nd_add(ref, indices, updates) with tf.Session() as sess: print sess.run(add)

The resulting update to ref would look like this:

1, 13, 3, 14, 14, 6, 7, 20

See tf.scatter_nd for more details about how to make updates to slices.

accumulatorNumAccumulated Source

Arguments

:: Tensor Ref ByteString	handle: The handle to an accumulator.
-> Build (Tensor Value Int32)	num_accumulated: The number of gradients aggregated in the given accumulator.

Returns the number of gradients aggregated in the given accumulators.

sparseSegmentSqrtN Source

Arguments

:: (TensorType t, OneOf `[Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	data
-> Tensor v2 tidx	indices: A 1-D tensor. Has same rank as `segment_ids`.
-> Tensor v3 Int32	segment_ids: A 1-D tensor. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the sum along sparse segments of a tensor divided by the sqrt of N.

N is the size of the segment being reduced.

Read the section on Segmentation for an explanation of segments.

depthToSpace Source

Arguments

:: TensorType t
=> Int64	block_size: The size of the spatial block, same as in Space2Depth.
-> Tensor v1 t	input
-> Tensor Value t	output

DepthToSpace for tensors of type T.

Rearranges data from depth into blocks of spatial data. This is the reverse transformation of SpaceToDepth. More specifically, this op outputs a copy of the input tensor where values from the depth dimension are moved in spatial blocks to the height and width dimensions. The attr block_size indicates the input block size and how the data is moved.

Chunks of data of size `block_size * block_size` from depth are rearranged into non-overlapping blocks of size `block_size x block_size`
The width the output tensor is `input_depth * block_size`, whereas the height is `input_height * block_size`.
The depth of the input tensor must be divisible by `block_size * block_size`.

That is, assuming the input is in the shape: `[batch, height, width, depth]`, the shape of the output will be: `[batch, height*block_size, width*block_size, depth/(block_size*block_size)]`

This operation requires that the input tensor be of rank 4, and that block_size be >=1 and that `block_size * block_size` be a divisor of the input depth.

This operation is useful for resizing the activations between convolutions (but keeping all data), e.g. instead of pooling. It is also useful for training purely convolutional models.

For example, given this input of shape `[1, 1, 1, 4]`, and a block size of 2:

```prettyprint x = [[[[1, 2, 3, 4]]]]

```

This operation will output a tensor of shape `[1, 2, 2, 1]`:

```prettyprint [[[[1], [2]], [[3], [4]]]] ```

Here, the input has a batch of 1 and each batch element has shape `[1, 1, 4]`, the corresponding output will have 2x2 elements and will have a depth of 1 channel (1 = `4 / (block_size * block_size)`). The output element shape is `[2, 2, 1]`.

For an input tensor with larger depth, here of shape `[1, 1, 1, 12]`, e.g.

```prettyprint x = [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]] ```

This operation, for block size of 2, will return the following tensor of shape `[1, 2, 2, 3]`

```prettyprint [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]]

```

Similarly, for the following input of shape `[1 2 2 4]`, and a block size of 2:

```prettyprint x = [[[[1, 2, 3, 4], [5, 6, 7, 8]], [[9, 10, 11, 12], [13, 14, 15, 16]]]] ```

the operator will return the following tensor of shape `[1 4 4 1]`:

```prettyprint x = [[ [1], [2], [5], [6]], [ [3], [4], [7], [8]], [ [9], [10], [13], [14]], [ [11], [12], [15], [16]]]

```

allCandidateSampler Source

Arguments

:: Int64	num_sampled: Number of candidates to produce per batch.
-> Int64	num_true: Number of true labels per context.
-> Bool	unique: If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
-> Tensor v1 Int64	true_classes: A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.
-> (Tensor Value Int64, Tensor Value Float, Tensor Value Float)	(sampled_candidates, true_expected_count, sampled_expected_count) sampled_candidates: A vector of length num_sampled, in which each element is the ID of a sampled candidate. true_expected_count: A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability. sampled_expected_count: A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

Generates labels for candidate sampling with a learned unigram distribution.

See explanations of candidate sampling and the data formats at go/candidate-sampling.

For each batch, this op picks a single set of sampled candidate labels.

The advantages of sampling candidates per-batch are simplicity and the possibility of efficient dense matrix multiplication. The disadvantage is that the sampled candidates must be chosen independently of the context and of the true labels.

resizeNearestNeighborGrad Source

Arguments

:: (TensorType t, OneOf `[Int32, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	grads: 4-D with shape `[batch, height, width, channels]`.
-> Tensor v2 Int32	size: = A 1-D int32 Tensor of 2 elements: `orig_height, orig_width`. The original input size.
-> Tensor Value t	output: 4-D with shape `[batch, orig_height, orig_width, channels]`. Gradients with respect to the input image.

Computes the gradient of nearest neighbor interpolation.

cTCGreedyDecoder Source

Arguments

:: Tensor v1 Float

inputs: 3-D, shape: `(max_time x batch_size x num_classes)`, the logits.

-> Tensor v2 Int32

sequence_length: A vector containing sequence lengths, size `(batch_size)`.

-> (Tensor Value Int64, Tensor Value Int64, Tensor Value Int64, Tensor Value Float)

(decoded_indices, decoded_values, decoded_shape, log_probability)

decoded_indices: Indices matrix, size `(total_decoded_outputs x 2)`, of a `SparseTensor2`. The rows store: [batch, time].
decoded_values: Values vector, size: `(total_decoded_outputs)`, of a `SparseTensor2`. The vector stores the decoded classes.
decoded_shape: Shape vector, size `(2)`, of the decoded SparseTensor. Values are: `[batch_size, max_decoded_length]`.
log_probability: Matrix, size `(batch_size x 1)`, containing sequence log-probabilities.

Performs greedy decoding on the logits given in inputs.

A note about the attribute merge_repeated: if enabled, when consecutive logits' maximum indices are the same, only the first of these is emitted. Labeling the blank *, the sequence "A B B * B B" becomes "A B" if merge_repeated = True and "A B B B B" if merge_repeated = False.

Regardless of the value of merge_repeated, if the maximum index of a given time and batch corresponds to the blank, index `(num_classes - 1)`, no new element is emitted.

l2Loss Source

Arguments

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t)
=> Tensor v1 t	t: Typically 2-D, but may have any dimensions.
-> Tensor Value t	output: 0-D.

L2 Loss.

Computes half the L2 norm of a tensor without the sqrt:

output = sum(t ** 2) / 2

segmentMax Source

Arguments

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word16, Word8, Double, Float]` t, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 t	data
-> Tensor v2 tindices	segment_ids: A 1-D tensor whose rank is equal to the rank of `data`'s first dimension. Values should be sorted and can be repeated.
-> Tensor Value t	output: Has same shape as data, except for dimension 0 which has size `k`, the number of segments.

Computes the maximum along segments of a tensor.

Read the section on Segmentation for an explanation of segments.

Computes a tensor such that $output_i = max_j(data_j)$ where max is over j such that `segment_ids[j] == i`.

style="width:70%; margin:auto; margin-bottom:10px; margin-top:20px;" style="width:100%" src="../../images/SegmentMax.png" alt /div

countUpTo Source

Arguments

:: (TensorType t, OneOf `[Int32, Int64]` t)
=> Int64	limit: If incrementing ref would bring it above limit, instead generates an `OutOfRange` error.
-> Tensor Ref t	ref: Should be from a scalar `Variable` node.
-> Build (Tensor Value t)	output: A copy of the input before increment. If nothing else modifies the input, the values produced will all be distinct.

Increments ref until it reaches limit.

tFRecordReader Source

Arguments

:: Build (Tensor Ref ByteString)

reader_handle: The handle to reference the Reader.

A Reader that outputs the records from a TensorFlow Records file.

switch Source

Arguments

:: TensorType t
=> Tensor v1 t	data: The tensor to be forwarded to the appropriate output.
-> Tensor v2 Bool	pred: A scalar that specifies which output port will receive data.
-> (Tensor Value t, Tensor Value t)	(output_false, output_true) output_false: If `pred` is false, data will be forwarded to this output. output_true: If `pred` is true, data will be forwarded to this output.

Forwards `data` to the output port determined by pred.

If pred is true, the `data` input is forwarded to output_true. Otherwise, the data goes to output_false.

:: (TensorType tparams, TensorType tindices, OneOf `[Int32, Int64]` tindices)
=> Tensor v1 tparams	params: `P-D`. The tensor from which to gather values.
-> Tensor v2 tindices	indices: `Q-D`. Index tensor having shape `[d_0, ..., d_{Q-2}, K]`.
-> Tensor Value tparams	output: `(P+Q-K-1)-D`. Values from `params` gathered from indices given by `indices`.

:: TensorType t
=> Tensor v1 t	input: The `input` to squeeze.
-> Tensor Value t	output: Contains the same data as `input`, but has one or more dimensions of size 1 removed.

:: Tensor Ref ByteString	reader_handle: Handle to a `Reader`.
-> Tensor Ref ByteString	queue_handle: Handle to a `Queue`, with string work items.
-> Tensor v3 Int64	num_records: number of records to read from `Reader`.
-> Build (Tensor Value ByteString, Tensor Value ByteString)	(keys, values) keys: A 1-D tensor. values: A 1-D tensor.

:: (TensorType tkey, TensorType tval)
=> Tensor Ref ByteString	table_handle: Handle to a table which will be initialized.
-> Tensor v2 tkey	keys: Keys of type Tkey.
-> Tensor v3 tval	values: Values of type Tval.
-> Build ControlNode

:: TensorType t
=> Int64	num_partitions: The number of partitions to output.
-> Tensor v1 t	data
-> Tensor v2 Int32	partitions: Any shape. Indices in the range `[0, num_partitions)`.
-> [Tensor Value t]	outputs

:: Int64	record_bytes
-> Build (Tensor Ref ByteString)	reader_handle: The handle to reference the Reader.

:: (TensorType t, OneOf `[Double, Float]` t, TensorType tidx, OneOf `[Int32, Int64]` tidx)
=> Tensor v1 t	start: First entry in the range.
-> Tensor v2 t	stop: Last entry in the range.
-> Tensor v3 tidx	num: Number of values to generate.
-> Tensor Value t	output: 1-D. The generated values.

:: TensorType t
=> Tensor v1 t	input: Rank `k` tensor where `k >= 2`.
-> Tensor Value t	diagonal: The extracted diagonal(s) having shape `diagonal.shape = input.shape[:-2] + [min(input.shape[-2:])]`.

:: (TensorType t, OneOf `[Word16, Word8]` t)
=> Tensor v1 t	image: 3-D with shape `[height, width, channels]`.
-> Tensor Value ByteString	contents: 0-D. PNG-encoded image.

:: TensorType t
=> Tensor Ref t	data: The ref tensor to be forwarded to the appropriate output.
-> Tensor v2 Bool	pred: A scalar that specifies which output port will receive data.
-> Build (Tensor Ref t, Tensor Ref t)	(output_false, output_true) output_false: If `pred` is false, data will be forwarded to this output. output_true: If `pred` is true, data will be forwarded to this output.

:: (TensorType t, OneOf `[Complex Double, Complex Float, Int32, Word16, Double, Float]` t)
=> Tensor v1 t	x: 3-D or higher with shape `[..., r_x, c_x]`.
-> Tensor v2 t	y: 3-D or higher with shape `[..., r_y, c_y]`.
-> Tensor Value t	output: 3-D or higher with shape `[..., r_o, c_o]`

:: TensorType t
=> Tensor v1 Int32	index: A scalar that determines the input that gets selected.
-> [Tensor Ref t]	inputs: A list of ref tensors, one of which will be forwarded to `output`.
-> Build (Tensor Ref t)	output: The forwarded tensor.

:: (TensorType t, OneOf `[Int16, Int32, Int64, Int8, Word8, Double, Float]` t)
=> Tensor v1 t	image: 3-D of shape `[height, width, channels]`.
-> Tensor v2 Int64	size: 1-D of length 2 containing: `crop_height`, `crop_width`..
-> Build (Tensor Value t)	output: 3-D of shape `[crop_height, crop_width, channels].`

:: Tensor v1 ByteString	contents: 0-D. The GIF-encoded image.
-> Tensor Value Word8	image: 4-D with shape `[num_frames, height, width, 3]`. RGB order

:: (TensorType t, OneOf `[Int32, Int64]` t)
=> Tensor v1 ByteString	input: Tensor of strings
-> Tensor v2 t	pos: Scalar defining the position of first character in each substring
-> Tensor v3 t	len: Scalar defining the number of characters to include in each substring
-> Tensor Value ByteString	output: Tensor of substrings