mirror of
https://github.com/tensorflow/haskell.git
synced 2024-11-23 11:29:43 +01:00
d62c614695
Distinguish between "rendered" and "unrendered" Tensors.
There are now three types of `Tensor`:
- `Tensor Value a`: rendered value
- `Tensor Ref a`: rendered reference
- `Tensor Build a` : unrendered value
The extra bookkeeping makes it easier to track (and enforce) which tensors are
rendered or not. For examples where this has been confusing in the past, see
With this change, pure ops look similar to before, returning `Tensor Build`
instead of `Tensor Value`. "Stateful" (monadic) ops are unchanged. For
example:
add :: OneOf [..] t => Tensor v'1 t -> Tensor v'2 t -> Tensor Build t
assign :: (MonadBuild m, TensorType t)
=> Tensor Ref t -> Tensor v'2 t -> m (Tensor Ref t)
The `gradients` function now requires that the variables over which it's
differentiating are pre-rendered:
gradients :: (..., Rendered v2) => Tensor v1 a -> [Tensor v2 a]
-> m [Tensor Value a]
(`Rendered v2` means that `v2` is either a `Ref` or a `Value`.)
Additionally, the implementation of `gradients` now takes care to render every
intermediate value when performing the reverse accumulation. I suspect this
fixes an exponential blowup for complicated expressions.
103 lines
3.6 KiB
Haskell
103 lines
3.6 KiB
Haskell
-- Copyright 2016 TensorFlow authors.
|
|
--
|
|
-- Licensed under the Apache License, Version 2.0 (the "License");
|
|
-- you may not use this file except in compliance with the License.
|
|
-- You may obtain a copy of the License at
|
|
--
|
|
-- http://www.apache.org/licenses/LICENSE-2.0
|
|
--
|
|
-- Unless required by applicable law or agreed to in writing, software
|
|
-- distributed under the License is distributed on an "AS IS" BASIS,
|
|
-- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
-- See the License for the specific language governing permissions and
|
|
-- limitations under the License.
|
|
|
|
{-# LANGUAGE FlexibleContexts #-}
|
|
{-# LANGUAGE OverloadedLists #-}
|
|
|
|
module Main where
|
|
|
|
import Google.Test (googleTest)
|
|
import TensorFlow.Test (assertAllClose)
|
|
import Test.Framework (Test)
|
|
import Test.Framework.Providers.HUnit (testCase)
|
|
import qualified Data.Vector as V
|
|
import qualified TensorFlow.Gradient as TF
|
|
import qualified TensorFlow.NN as TF
|
|
import qualified TensorFlow.Ops as TF
|
|
import qualified TensorFlow.Core as TF
|
|
|
|
-- | These tests are ported from:
|
|
--
|
|
-- <tensorflow>/tensorflow/python/ops/nn_xent_tests.py
|
|
--
|
|
-- This is the implementation we use to check the implementation we
|
|
-- wrote in `TensorFlow.NN.sigmoidCrossEntropyWithLogits`.
|
|
--
|
|
sigmoidXentWithLogits :: Floating a => Ord a => [a] -> [a] -> [a]
|
|
sigmoidXentWithLogits logits' targets' =
|
|
let sig = map (\x -> 1 / (1 + exp (-x))) logits'
|
|
eps = 0.0001
|
|
predictions = map (\p -> min (max p eps) (1 - eps)) sig
|
|
xent y z = (-z) * (log y) - (1 - z) * log (1 - y)
|
|
in zipWith xent predictions targets'
|
|
|
|
|
|
data Inputs = Inputs {
|
|
logits :: [Float]
|
|
, targets :: [Float]
|
|
}
|
|
|
|
|
|
defInputs :: Inputs
|
|
defInputs = Inputs {
|
|
logits = [-100, -2, -2, 0, 2, 2, 2, 100]
|
|
, targets = [ 0, 0, 1, 0, 0, 1, 0.5, 1]
|
|
}
|
|
|
|
|
|
testLogisticOutput :: Test
|
|
testLogisticOutput = testCase "testLogisticOutput" $ do
|
|
let inputs = defInputs
|
|
r <- run $ do
|
|
vLogits <- TF.render $ TF.vector $ logits inputs
|
|
vTargets <- TF.render $ TF.vector $ targets inputs
|
|
TF.sigmoidCrossEntropyWithLogits vLogits vTargets
|
|
let ourLoss = V.fromList $ sigmoidXentWithLogits (logits inputs) (targets inputs)
|
|
assertAllClose r ourLoss
|
|
|
|
|
|
testLogisticOutputMultipleDim :: Test
|
|
testLogisticOutputMultipleDim =
|
|
testCase "testLogisticOutputMultipleDim" $ do
|
|
let inputs = defInputs
|
|
shape = [2, 2, 2]
|
|
r <- run $ do
|
|
vLogits <- TF.render $ TF.constant shape (logits inputs)
|
|
vTargets <- TF.render $ TF.constant shape (targets inputs)
|
|
TF.sigmoidCrossEntropyWithLogits vLogits vTargets
|
|
let ourLoss = V.fromList $ sigmoidXentWithLogits (logits inputs) (targets inputs)
|
|
assertAllClose r ourLoss
|
|
|
|
|
|
testGradientAtZero :: Test
|
|
testGradientAtZero = testCase "testGradientAtZero" $ do
|
|
r <- run $ do
|
|
let inputs = defInputs { logits = [0, 0], targets = [0, 1] }
|
|
vTargets <- TF.render $ TF.vector $ targets inputs
|
|
vLogits <- TF.render $ TF.vector $ logits inputs
|
|
let tfLoss = TF.sigmoidCrossEntropyWithLogits vLogits vTargets
|
|
|
|
l <- tfLoss
|
|
TF.gradients l [vLogits]
|
|
|
|
assertAllClose (head r) (V.fromList [0.5, -0.5])
|
|
|
|
run :: TF.Fetchable t a => TF.Session t -> IO a
|
|
run = TF.runSession . (>>= TF.run)
|
|
|
|
main :: IO ()
|
|
main = googleTest [ testGradientAtZero
|
|
, testLogisticOutput
|
|
, testLogisticOutputMultipleDim
|
|
]
|