| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
properly display on the same row.
Change: 148806312
|
|
|
|
| |
Change: 148806189
|
|
|
|
| |
Change: 148805879
|
|
|
|
|
| |
(All known callers already assumed that this was the case before.)
Change: 148805391
|
|
|
|
|
| |
under benchmarks repo (https://github.com/tensorflow/benchmarks).
Change: 148802196
|
|
|
|
| |
Change: 148797512
|
|
|
|
| |
Change: 148795638
|
|
|
|
| |
Change: 148794476
|
|
|
|
| |
Change: 148791273
|
|
|
|
| |
Change: 148788753
|
|
|
|
|
|
| |
Removes restrictive requirement that replacing an operand requires
compatible shapes which is not true with broadcast optimizations.
Change: 148786513
|
|
|
|
| |
Change: 148786108
|
|
|
|
| |
Change: 148778464
|
|
|
|
|
| |
into different string literals compared to GCC.
Change: 148767816
|
|
|
|
| |
Change: 148765793
|
|
|
|
|
|
|
|
| |
A Session is configured using the ConfigProto protocol buffer.
For now, continuing with attempts to keep the 'tensorflow' go package
free of any protocol buffer dependencies, SessionOptions uses a serialized
representation of this message. This choice might make sense to revisit.
Change: 148750535
|
|
|
|
| |
Change: 148740836
|
|
|
|
|
| |
into different string literals compared to GCC.
Change: 148735423
|
|
|
|
|
| |
And param order in nce_loss in the code sample
Change: 148732811
|
|
|
|
|
|
| |
compute some per class metrics. The per class AUC had a bug where we were using logits rather than probabilities. This CL fixes this bug.
Change: 148727736
|
|
|
|
| |
Change: 148709245
|
|
|
|
| |
Change: 148707426
|
|
|
|
|
|
|
| |
behaves like any other cast op now.
Micro-optimization: Don't call InferAllocAttr on control edges, since the result is ignored anyway.
Change: 148705319
|
|
|
|
| |
Change: 148702459
|
|
|
|
| |
Change: 148700711
|
|
|
|
| |
Change: 148699787
|
|
|
|
| |
Change: 148698589
|
|
|
|
|
| |
It is disabled in bazel already.
Change: 148691841
|
|
|
|
|
|
|
| |
- Delete memory allocated for exception error messages.
- Use vsnprintf instead of vasprintf for allocating exception error messages
since the latter doesn't seem to be easily available on Windows.
Change: 148687623
|
|
|
|
| |
Change: 148684190
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* enable compilation of VarIsInitializedOp.
* fix deprecated variable initializer in variable_ops_test.py
* simplify variable logic in XlaContext, move intelligence into XlaOpKernelContext.
* add resource variable support in the contrib layers library.
Cleanups and refactorings:
* merge XlaCompiler::CompileSubComputation with XlaCompiler::CompileFunction.
* pass XlaCompiler arguments consistently via XlaCompiler::Options.
* split the two roles of XlaCompiler::CompilationResult::input_shapes into input_mapping and xla_input_shapes.
* initialize the numpy and Python seeds to a constant for XLA test cases.
Change: 148683645
|
|
|
|
|
|
|
|
| |
* Moves the core forwarding logic to a new function forward_input that is not restricted to forwarding to an output slot, but also allows, e.g., reusing input buffers as temporaries or variables.
* Gets rid of forward_input_to_output_with_same_shape that is now unused.
* Adds convenience methods input_memory_type(), output_memory_type, and (private) input_is_ref() to OpKernelContext.
* Misc. small cleanups.
Change: 148683484
|
|
|
|
|
| |
get additional processing before being published.
Change: 148682342
|
|
|
|
| |
Change: 148682151
|
|
|
|
| |
Change: 148681977
|
|
|
|
| |
Change: 148681573
|
|
|
|
| |
Change: 148678769
|
|
|
|
|
|
|
|
|
| |
instead of depending on variable initialization logic. This avoids deadlocks that can
happen if QueueRunners are used to generate data. Deadlocks would happen since
initialization would wait for QueueRunner to generate data, but QueueRunners
were not started till variable initialization was complete.
Also some refactoring of the code to move Hooks and model function out of the Estimator class.
Change: 148678573
|
|
|
|
| |
Change: 148678164
|
|
|
|
| |
Change: 148676738
|
|
|
|
|
|
|
| |
more traditional ops/ and kernels/ directories.
No functional changes.
Change: 148673534
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Applied LLVM optimization patch to libxsmm
(https://github.com/hfp/libxsmm/commit/0e412d5d2769a8754cace64e56e26e14093f887d.patch).
- Limited outstanding libxsmm sparse matrix multiply handle counts to limit
memory usage for temporary space.
- Added extra logging to libxsmm handle management in TensorFlow.
- Added support for running multiple sparse matrix multiplies simultaneously in
performance benchmark to match some practical use cases.
- Added more size combinations to sparse matrix multiply benchmark.
- Fixed dependencies for xsmm_conv2d_test.
Change: 148672973
|
|
|
|
| |
Change: 148671742
|
|
|
|
|
|
|
| |
repeatedly issued by code in headers.
Fixes #7919
Change: 148670735
|
|
|
|
| |
Change: 148669949
|
|
|
|
|
|
| |
the session master.
Change: 148669552
|
|
|
|
|
|
|
|
| |
Add an HLO call graph class. The object includes a node for each computation
and forward and backwards links between them. Node include the calling context:
either parallel (eg, kMap) or sequential (eg, kCall). The class is not used
anywhere yet, but there are numerous potential uses throughout the HLO level.
Change: 148669295
|
|
|
|
| |
Change: 148668422
|
|
|
|
| |
Change: 148668420
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This was identical to libtensorflow.so for all but
a few days in the life of libtensorflow_c.so's existence.
Remove libtensorflow_c.so to remove any confusion about
which target to use for the TensorFlow C library.
(https://github.com/tensorflow/tensorflow/commit/783c52e
and #3830
were the changes that resulted in two targets doing the same thing)
I have made attempts to remove use of libtensorflow_c.so in
other projects such as:
https://github.com/LaurentMazare/tensorflow-ocaml/pull/19
https://github.com/tensorflow/rust/pull/56
https://github.com/malmaud/TensorFlow.jl/pull/157
Change: 148665838
|