| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 192342172
|
|
|
|
| |
PiperOrigin-RevId: 192341561
|
|
|
|
| |
PiperOrigin-RevId: 192341065
|
|
|
|
|
|
|
|
|
|
|
|
| |
- tf.layers layers now subclasses tf.keras.layers layers.
- tf.keras.layers is now agnostic to variable scopes and global collections (future-proof). It also uses ResourceVariable everywhere by default.
- As a result tf.keras.layers is in general lower-complexity, with fewer hacks and workarounds. However some of current code is temporary (variable creation should be moved to Checkpointable, arguably, and there are some dependency issues that will require later refactors).
- The legacy tf.layers layers behavior is kept, with references to variable scopes and global collections injected in the subclassed tf.layers.base.Layer class (the content of tf.layers.base.Layer is the complexity differential between the old implementation and the new one).
Note: this refactor does slightly change the behavior of tf.layers.base.Layer, by disabling extreme edge-case behavior that either has long been invalid, or is dangerous and should most definitely be disabled. This will not affect any users since such behaviors only existed in the base Layer unit tests. The behaviors disabled are:
- Option to create reusable variables in `call` (already invalid for some time).
- Option to use a variable scope to create layer variables outside of the layer while not having the layer track such variables locally.
PiperOrigin-RevId: 192339798
|
|
|
|
|
|
| |
guide
PiperOrigin-RevId: 192336966
|
|
|
|
|
|
|
|
| |
This should make restore() work with defun-wrapped code, when variables are
created inside the function. Just lifts the restore code into the outer
context. Adds a test for it.
PiperOrigin-RevId: 192331065
|
|
|
|
| |
PiperOrigin-RevId: 192327464
|
|
|
|
|
|
| |
This bumps the gRPC version used in OSS Tensorflow to pick up grpc/grpc#14541, which exposes gRPC serialization classes which were previously hidden in an internal namespace. Using these files eliminates files duplicated from gRPC repo
PiperOrigin-RevId: 192327358
|
|
|
|
|
|
| |
where we can apply the involution and mul->conv fusion optimizations.
PiperOrigin-RevId: 192323712
|
|
|
|
|
|
| |
Thanks to the team at Alibaba, who contributed the original version of this code.
PiperOrigin-RevId: 192322484
|
|
|
|
| |
PiperOrigin-RevId: 192318426
|
|
|
|
| |
PiperOrigin-RevId: 192315888
|
|
|
|
| |
PiperOrigin-RevId: 192314177
|
|
|
|
| |
PiperOrigin-RevId: 192310749
|
|
|
|
|
|
|
| |
- Get rid of unnecessary options
- Update various resource variable tests so that they correctly exercise the cases where the variables are placed on GPU (these "with tf.device('/cpu:0')" blocks that were added for eager execution are no longer necessary)
PiperOrigin-RevId: 192309109
|
|
|
|
| |
PiperOrigin-RevId: 192302833
|
|
|
|
| |
PiperOrigin-RevId: 192299871
|
|
|
|
|
|
| |
that is being removed from bazel.
PiperOrigin-RevId: 192298746
|
|
|
|
| |
PiperOrigin-RevId: 192298563
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
utilization.
We often have simple kernels that do very little actual work, duplicating that
can increase the used bandwidth.
This change introduces flags and infrastructure for unrolling kernels, it
doesn't include any cost heuristics and is disabled by default. Based on code
written by Bixia Zheng.
PiperOrigin-RevId: 192296781
|
|
|
|
| |
PiperOrigin-RevId: 192292693
|
|
|
|
| |
PiperOrigin-RevId: 192292160
|
|
|
|
|
|
| |
revisions.
PiperOrigin-RevId: 192284946
|
|
|
|
| |
PiperOrigin-RevId: 192241311
|
|
|
|
|
|
|
|
|
| |
XlaBuidler::BuildConstantSubGraph, and Client::ComputeConstant(XlaComputation...).
- Since the builder no longer holds a client, we moved the ComputeConstant to the client side so that it can communicate with the service side. Now we add XlaBuilder::BuildConstantSubGraph, which is only responsible for building a subgraph that is compile-time constant.
- Before this change, every XlaBuilder has a unique id. Now since it also builds constant subgraph, we give every XlaComputation being built a global unique id, and uniquify instruction names when actually building the XlaComputation.
PiperOrigin-RevId: 192236997
|
|
|
|
| |
PiperOrigin-RevId: 192236845
|
|
|
|
|
|
| |
check EvaluateNodes returns the same output for the original and optimized graph
PiperOrigin-RevId: 192235310
|
|
|
|
|
|
|
|
| |
The original badges I added were redirected/authenticated links (I think). This
resulted in broken images on the README. These new links should not be
redirected and should just link to the badge images.
PiperOrigin-RevId: 192230689
|
|
|
|
| |
PiperOrigin-RevId: 192227995
|
|
|
|
| |
PiperOrigin-RevId: 192226063
|
|
|
|
|
|
|
|
|
| |
implement ComputeAndCompare for XlaBuilder.
- In ComputeAndCompare(XlaBuilder..), compute the reference result by executing on the interpreter backend.
- Also migrate some tests to use the new ComputeAndCompare(XlaBuilder..).
PiperOrigin-RevId: 192225152
|
|
|
|
| |
PiperOrigin-RevId: 192224104
|
|
|
|
| |
PiperOrigin-RevId: 192223560
|
|
|
|
| |
PiperOrigin-RevId: 192220070
|
|
|
|
|
|
| |
only a subset of CSV columns; added same parameter to make_csv_dataset function in tf.contrib.data.
PiperOrigin-RevId: 192219509
|
|
|
|
|
| |
by reading tensor values in a separate session after train_op run.
PiperOrigin-RevId: 192217338
|
|
|
|
|
|
|
| |
It breaks some builds with
error: "EIGEN_USE_THREADS" redefined [-Werror]
PiperOrigin-RevId: 192216796
|
|
|
|
|
|
| |
more control over what to do if checkpoints_iterator times out.
PiperOrigin-RevId: 192216302
|
|
|
|
| |
PiperOrigin-RevId: 192215493
|
|
|
|
| |
PiperOrigin-RevId: 192210794
|
|
|
|
| |
PiperOrigin-RevId: 192209093
|
|
|
|
|
|
|
| |
- XlaBuilder::Infeed is basically ComputationBuilder::Infeed + UserComputation::AddInfeedInstruction + ComputationLowerer::Visit + HloInstruction::CreateInfeed.
- Similar for Outfeed.
PiperOrigin-RevId: 192206502
|
|
|
|
| |
PiperOrigin-RevId: 192204209
|
|
|
|
| |
PiperOrigin-RevId: 192197163
|
|
|
|
|
|
| |
ThreadPool joins its threads when it is destroyed, and there's no way to explicitly join. This means passing a ThreadPool and then scheduling in the callee is risky.
PiperOrigin-RevId: 192193752
|
|
|
|
| |
PiperOrigin-RevId: 192193350
|
|
|
|
| |
PiperOrigin-RevId: 192192757
|
|
|
|
|
|
|
| |
Previously breaking out into Graphs created with eager execution enabled would
enter the graph but not re-enable eager execution.
PiperOrigin-RevId: 192192109
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
TF_Graph creates a ShapeRefiner, which in
turn creates a GraphRunner, which prior to this change would eventually create a
LocalDevice that initialized the global eigen threadpool. This prevents
users from specifying a custom number of threads for the pool via a
ConfigProto.
This change introduces a new device class, SingleThreadedCpuDevice, that can
be used for light-weight computations without initializing the threadpool.
Addresses #18300.
PiperOrigin-RevId: 192188031
|
|
|
|
|
|
|
|
| |
1. Eliminate the checks when the pointer is nullptr.
2. Fix nan array size. Some application can produce a zero size which would
cause out-of-range access error.
PiperOrigin-RevId: 192186224
|