1
0
Fork 0
mirror of https://github.com/tensorflow/haskell.git synced 2024-12-26 03:29:45 +01:00
Commit graph

49 commits

Author SHA1 Message Date
Bart Schuurmans
4eac338d0f Fix for ghc 9.2 2023-01-26 09:21:10 -08:00
fkm3
199f1c7663 Switch to new TF_STRING format for TF 2.10
See https://github.com/tensorflow/community/blob/master/rfcs/20190411-string-unification.md
2023-01-23 17:17:47 -08:00
Bart Schuurmans
b0e832809c Consistent whitespace usage 2023-01-23 17:17:47 -08:00
Bart Schuurmans
e68a414999 Set cabal-version to 2.0 everywhere 2023-01-23 17:17:47 -08:00
jcmartin
4d9d315fbd
Update Package Version Numbers (#268) 2020-11-11 09:21:35 -08:00
jcmartin
c66c912c32
Tensorflow 2.3.0 Support (#267)
* Tensorflow 2.3.0 building and passing tests.
* Added einsum and test.
* Added ByteString as a possible argument to a function.
* Support more data types for Adam.
* Move to later version of LTS on stackage.
* Added a wrapper module for convolution functions.
* Update ci build to use a later version of stack.
* Removed a deprecated import in GradientTest.
2020-11-06 11:32:21 -08:00
Mike Sperber
568c9b6f03
Update to current proto-lens packages. (#258) 2020-05-21 13:36:52 -07:00
Greg Steuck
580917a731
Switch to LTS 14 and bump versions to 0.2.0.1 (#248)
* Use newer stack and protoc in Dockerfiles
2019-09-08 12:45:24 -07:00
Yorick
7e1d92c6c2 Bump version constraints for proto-lens-* (#247) 2019-09-02 18:19:07 -07:00
Daniel YU
7316062c10 upgrade to ghc 8.6.4 (#237) 2019-04-11 19:27:15 -07:00
Christian Berentsen
61e58fd33f Use proto-lens* == 0.3.* (#212)
* Include more *_Fields modules
2018-09-04 10:44:52 -07:00
fkm3
85bf0bb12c
Bump all of the versions to 0.2.0.0 (#202) 2018-08-02 22:07:30 -04:00
fkm3
1e2dca8701
Update to tensorflow 1.7 (#185)
All of the non-s/1.3/1.7/ changes are because

* There are new tensorflow datatypes
* Some ops have looser types (e.g. fill now accepts both int64 and int32)
* There are more ops of type "func"
2018-04-17 12:24:31 -04:00
fkm3
e35211d49b Fix initialized variables for tensorflow 1.7 (#184)
* Fix initialized variables for tensorflow 1.7

This is needed to support tensorflow 1.7. The trick of initializing a
variable with `Shape []` and then overriding the shape by assigning an
initial value no longer works. It seems that we need to explicitly flip
the unknown_rank bit in the shape proto.

I thought about switching opgen to use `Maybe Shape` when an op requires
a shape attribute, but that will cause a lot of api churn, so I chose to
hold off for now and just do a spot fix to unblock 1.7.
2018-04-16 07:48:05 -07:00
fkm3
7720af0afd Update to tensorflow 1.3 (#161)
* Tested on linux without Docker.
* Couldn't get nix build to work, so I just updated the URL and hash.
* Did not test macos build.

The mnist change was necessary because the argmax output type is now polmorphic.
2017-10-19 13:41:55 -04:00
Jeroen Bransen
d8bf349962 Create monad transformer version of Session (closes #153) (#154) 2017-10-02 16:33:49 -04:00
Judah Jacobson
7328cb277f Fix the build with ghc-8.2.1. (#147)
- Avoid using a deprecated Cabal function
- Use newer versions of proto-lens packages in stack.yaml
- Work around a new type-level warning that affects `OneOf/TensorTypes`.
2017-08-08 09:48:59 -07:00
Judah Jacobson
56038ba27e Depend on a newer proto-lens and remove orphan Ord instances. (#146)
proto-lens-0.2.2.0 generates Ord instances for all message types,
so we can remove the orphan instances we previously added.

Dependends on proto-lens-protoc-0.2.2.1 or newer due to google/proto-lens#113.
2017-08-02 22:47:55 -07:00
fkm3
a86d424cac Add initializedValue function for Variable (#124) 2017-05-20 21:42:45 -07:00
fkm3
ddb4fe4f90 Add ToTensor class 2017-05-16 23:05:33 -07:00
Judah Jacobson
d98e5d637c Add extra-lib-dirs back into .cabal for Mac builds. (#122)
Also bump the version to 0.1.0.2.

Originally we had `extra-lib-dirs: /usr/local/lib` in `stack.yaml`.
I removed it because it wasn't necessary on my Mac.  However,
it turns out that it is necessary for machines with the default installation
of XCode, which *doesn't* search that path by default.

(On my machine, it wasn't necessary because I had run `xcode-select --install`
which adds that path permanently to your search path.  For more context, see
https://github.com/Homebrew/brew/issues/556.)

I'm adding the setting back to `tensorflow.cabal` as well as `stack.yaml` so
that the Hackage release also contains this fix.  Changing `stack.yaml` is
still necessary in order to fix linkage in the `snappy` package (which
`tensorflow-records` depends on).  Hopefully that will go away once we remove
the dependency (#118).
2017-05-16 15:18:01 -07:00
fkm3
0f04e5a50d Expand Rendered class to support ResourceHandle wrappers like Variable
This allows functions like `feed`, `colocateWith`, and (in a later commit)
`gradients` to work with `Variable`.
2017-05-15 19:55:34 -07:00
Judah Jacobson
a7a2be2110 Include c_api.h in the tensorflow sdist. (#114)
Also bumps up the version of the 'tensorflow' package to 0.1.0.1.
2017-05-10 16:50:35 -07:00
Judah Jacobson
64971c876a Consolidate some packages. (#111)
- Merge tensorflow-nn and tensorflow-queue into tensorflow-ops.
  They don't add extra dependencies and each contain a single module, so I
  don't think it's worth separating them at the package level.
- Remove google-shim in favor of direct use of test-framework.
2017-05-10 15:26:03 -07:00
Judah Jacobson
0fa719b701 Fix .cabal files so 'stack check' passes. (#110)
- Add LICENSE files for all packages.
- Add descriptions for packages that were missing one.
- Work around google/proto-lens#69 by symlinking third_party into
  tensorflow-proto.
2017-05-10 11:37:00 -07:00
Judah Jacobson
a64af5076a Work around #92 by always copying TensorData when fetching.
It would be better to avoid the copy when it's not necessary, but
that will require more involved changes to the internal API.  (For example,
Fetchable might need to allow IO or ST actions.)
2017-05-09 00:10:29 -07:00
Judah Jacobson
42f4fc647e Add resource-based variable ops. (#98)
The main difference between these and the `Ref`-bases ops is the explicit
`readValue` op.  I'm not sure how this should interact with gradients
and save/restore, so I'm keeping it as a separate module for now.  Once we
figure out the details, we can merge it into `TensorFlow.Ops` and replace
all uses of the old `Ref`-based ops.  (That would also fix #92.)

Also replaces our special case newtype `ResourceHandle` to
`Tensor Value ResourceHandle`, where `ResourceHandle` is the TF proto
corresponding to `DT_RESOURCE`.
2017-04-16 09:24:02 -07:00
Christian Berentsen
21b723d542 Adapt to lts-8.6 and use proto-lens-0.2.0.1 (#97) 2017-04-11 14:09:01 -07:00
Judah Jacobson
16d660c3bc Support a couple more ops by allowing larger tuples. (#93) 2017-04-06 19:00:18 -07:00
Judah Jacobson
d62c614695 Distinguish between "rendered" and "unrendered" Tensors. (#88)
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.
2017-04-06 15:10:33 -07:00
Judah Jacobson
c99a23b6a7 Add versions of each op that take optional params as an extra arg. (#84)
Each op `foo :: ...` now has a corresponding `foo' :: OpParams -> ...`
which lets you set optional attributes.  `OpParams` is currently a type alias for
`OpDef -> OpDef`.  In the future we should consider more type safety, e.g.,
using type-level strings and OverloadedLabels for optional attributes.

I used it to replace a few manual `buildOp`s in our code with the codegenerated
ops, now that it's easier to set attributes.  I also removed `tensorAttr` and
`named` since it's now possible to set those op attributes directly.

Although this clutters up the API a bit, I think it's simpler than using type
classes to implement optional arguments (as in, for example, `Text.Printf`) --
especially in terms of type inference with the rest of the library.
2017-03-20 18:16:38 -07:00
Judah Jacobson
2c5c879037 Introduce a MonadBuild class, and remove buildAnd. (#83)
This change adds a class that both `Build` and `Session` are instances of:

    class MonadBuild m where
        build :: Build a -> m a

All stateful ops (generated and manually written) now have a signature that returns
an instance of `MonadBuild` (rather than just `Build`).  For example:

    assign_ :: (MonadBuild m, TensorType t)
            => Tensor Ref t -> Tensor v t -> m (Tensor Ref t)

This lets us remove a bunch of spurious calls to `build` in user code.  It also
lets us replace the pattern `buildAnd run foo` with the simpler pattern `foo >>= run`
(or `run =<< foo`, which is sometimes nicer when foo is a complicated expression).

I went ahead and deleted `buildAnd` altogether since it seems to lead to
confusion; in particular a few tests had `buildAnd run . pure` which is
actually equivalent to just `run`.
2017-03-18 12:08:53 -07:00
Judah Jacobson
9209dfc4c4 Support lists of tensors in ops. (#79)
Adds a new type `ListOf` which wraps a heterogeneous list; for example,
`ListOf (Tensor Value) '[Int32, Float]` represents a list of two
elements: a tensor of int32s and a tensor of floats.

Also changes the `Queue2` type (which suppored pairs of tensors) to
`Queue` (which supports arbitrary lists).
2017-03-17 13:53:19 -07:00
Judah Jacobson
0c8d41250a Remove the type parameter from ResourceHandle. (#76)
This change allows us to reenable the rest of the ResourceHandle ops, and
future-proofs us against more being added.  It removes the custom logic that
assumed there was a "dtype" attribute to guess what the type parameter is
(which wasn't true in general.)

When we switch to ResourceHandle (e.g., for queues and variables) we can add
parameters to the wrapper types like "Queue" on a case-by-case basis.
2017-02-21 19:38:26 -08:00
fkm3
b3c0997a8c Add support for logging to tensorboard (#74)
Add support for logging to tensorboard

Based on @gnezdo's internal version with some differences:

* Uses a pure haskell implementation of EventWriter instead of FFI.
* Special `buildAnd*` functions were dropped in favor of using
  `mergeAllSummaries :: Build SummaryTensor` with the normal
  `build` function.
2017-02-20 19:16:42 -08:00
Greg Steuck
72631cb9f3 Uprev to TF 1.0rc1. (#69)
* Download protoc and libtensorflow instead of running bazel.
* Explicitly set permissions of protoc.
2017-02-09 14:20:43 -08:00
fkm3
f170df9d13 Support fetching storable vectors + use them in benchmark (#50)
In addition, you can now fetch TensorData directly. This might be useful in
scenarios where you feed the result of a computation back in, like RNN.

Before:

benchmarking feedFetch/4 byte
time                 83.31 μs   (81.88 μs .. 84.75 μs)
                     0.997 R²   (0.994 R² .. 0.998 R²)
mean                 87.32 μs   (86.06 μs .. 88.83 μs)
std dev              4.580 μs   (3.698 μs .. 5.567 μs)
variance introduced by outliers: 55% (severely inflated)

benchmarking feedFetch/4 KiB
time                 114.9 μs   (111.5 μs .. 118.2 μs)
                     0.996 R²   (0.994 R² .. 0.998 R²)
mean                 117.3 μs   (116.2 μs .. 118.6 μs)
std dev              3.877 μs   (3.058 μs .. 5.565 μs)
variance introduced by outliers: 31% (moderately inflated)

benchmarking feedFetch/4 MiB
time                 109.0 ms   (107.9 ms .. 110.7 ms)
                     1.000 R²   (0.999 R² .. 1.000 R²)
mean                 108.6 ms   (108.2 ms .. 109.2 ms)
std dev              740.2 μs   (353.2 μs .. 1.186 ms)

After:

benchmarking feedFetch/4 byte
time                 82.92 μs   (80.55 μs .. 85.24 μs)
                     0.996 R²   (0.993 R² .. 0.998 R²)
mean                 83.58 μs   (82.34 μs .. 84.89 μs)
std dev              4.327 μs   (3.664 μs .. 5.375 μs)
variance introduced by outliers: 54% (severely inflated)

benchmarking feedFetch/4 KiB
time                 85.69 μs   (83.81 μs .. 87.30 μs)
                     0.997 R²   (0.996 R² .. 0.999 R²)
mean                 86.99 μs   (86.11 μs .. 88.15 μs)
std dev              3.608 μs   (2.854 μs .. 5.273 μs)
variance introduced by outliers: 43% (moderately inflated)

benchmarking feedFetch/4 MiB
time                 1.582 ms   (1.509 ms .. 1.677 ms)
                     0.970 R²   (0.936 R² .. 0.993 R²)
mean                 1.645 ms   (1.554 ms .. 1.981 ms)
std dev              490.6 μs   (138.9 μs .. 1.067 ms)
variance introduced by outliers: 97% (severely inflated)
2016-12-14 18:53:06 -08:00
fkm3
91f508eb5c Fix TensorData encode and decode for Bool (#49) 2016-12-12 19:40:32 -08:00
Judah Jacobson
1539783ee5 Update type constraints to work around a ghc-8 bug. (#47)
Also removes all the ghc-8-specific logic in the .cabal files.

ghc-8 has issues with deeply nested tuples of constraints.  We can
work around it by:
- Changing TensorTypes to a regular class.  This required FlexibleContexts.
  (But we'll probably need it anyway when we support heterogeneous tensor
  lists.)
- Specializing NoneOf for long type lists.

For more details, see: https://ghc.haskell.org/trac/ghc/ticket/12175.

Also added 'directory' to tensorflow-core-ops' dependencies since it's used
in the Setup script.

One more step towards fixing #38.
2016-11-28 21:15:09 -08:00
Judah Jacobson
5b4017e31b Fix the build on ghc-8.0.1 (#38). (#40)
Two issues:
- The definition of `\\` was missing parentheses.  It was probably a bug
  that this used to worked in ghc-7.10.
- Set `-fconstraint-solver-iterations=0` to work around
  https://ghc.haskell.org/trac/ghc/ticket/12175.  It looks like we can
  trigger that bug when defining a significantly complicated op.  Specifically,
  our type shenanigans ("OneOf") along with lens setters (for OpDef) seem
  to confuse GHC.

Still TODO: automate testing of different ghc versions to prevent a regression.
2016-11-21 22:20:08 -08:00
Greg Steuck
2b5e41ffeb Make code --pedantic (#35)
* Enforce pedantic build mode in CI.
* Our imports drifted really far from where they should be.
2016-11-18 10:42:02 -08:00
fkm3
fc3d398ca9 Optimize fetching (#27)
* Add MNIST data to gitignore
* Add simple tensor round-trip benchmark
* Use deepseq + cleaner imports
* Use safe version of fromIntegral in FFI code
* Don't copy data when fetching tensors

BEFORE

benchmarking feedFetch/4 byte
time                 55.79 μs   (54.88 μs .. 56.62 μs)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 55.61 μs   (55.09 μs .. 56.11 μs)
std dev              1.828 μs   (1.424 μs .. 2.518 μs)
variance introduced by outliers: 34% (moderately inflated)

benchmarking feedFetch/4 KiB
time                 231.4 μs   (221.9 μs .. 247.3 μs)
                     0.988 R²   (0.974 R² .. 1.000 R²)
mean                 226.6 μs   (224.1 μs .. 236.2 μs)
std dev              13.45 μs   (7.115 μs .. 27.14 μs)
variance introduced by outliers: 57% (severely inflated)

benchmarking feedFetch/4 MiB
time                 485.8 ms   (424.6 ms .. 526.7 ms)
                     0.998 R²   (0.994 R² .. 1.000 R²)
mean                 515.7 ms   (512.5 ms .. 517.9 ms)
std dev              3.320 ms   (0.0 s .. 3.822 ms)
variance introduced by outliers: 19% (moderately inflated)

AFTER

benchmarking feedFetch/4 byte
time                 53.11 μs   (52.12 μs .. 54.22 μs)
                     0.996 R²   (0.995 R² .. 0.998 R²)
mean                 54.64 μs   (53.59 μs .. 56.18 μs)
std dev              4.249 μs   (2.910 μs .. 6.076 μs)
variance introduced by outliers: 75% (severely inflated)

benchmarking feedFetch/4 KiB
time                 83.83 μs   (82.72 μs .. 84.92 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 83.82 μs   (83.20 μs .. 84.35 μs)
std dev              1.943 μs   (1.557 μs .. 2.614 μs)
variance introduced by outliers: 20% (moderately inflated)

benchmarking feedFetch/4 MiB
time                 95.54 ms   (93.62 ms .. 97.82 ms)
                     0.999 R²   (0.998 R² .. 1.000 R²)
mean                 96.61 ms   (95.76 ms .. 97.51 ms)
std dev              1.408 ms   (1.005 ms .. 1.889 ms)
2016-11-17 10:41:49 -08:00
Greg Steuck
c430e54c3c Uprev tensorflow. (#33)
* No longer need to hide ResourceHandle ops
* Blacklisted not supported TensorArrayV2
* Ownership of feed tensors changed (1f0c5119a0230c5160d45496175b9256f097e144)
2016-11-16 21:16:20 -08:00
Greg Steuck
93e27a12c6 Uprev tensorflow. (#29)
Includes temporary blacklisting for a couple of ops that will be
supported once my fix lands in the main tensorflow repo.
2016-11-14 17:04:44 -08:00
Greg Steuck
0d4f5a9628 Added sessionTracer to log graph operations. (#26)
* Added TracingTest.
2016-11-14 15:14:51 -08:00
fkm3
630850c2d2 Add TensorFlow.Core module to start formalizing the exposed API (#17)
* Add TensorFlow.Core module to start formalizing the exposed API

* Refer to ops packages instead of modules
2016-11-10 09:47:41 -08:00
Greg Steuck
8db944578a Support ResourceHandle. (#18)
Exposed by moving to newer TF.
2016-11-08 16:48:41 -08:00
Judah Jacobson
cdd4a0a747 Compile on platforms where int64_t == long long. (#6)
In particular, this helps fix the build on Mac OS X.
2016-10-25 22:26:42 -07:00
Greg Steuck
67690d1499 Initial commit 2016-10-24 19:26:42 +00:00