提交 · 6e72b20bb51b980ef0e8cfc5ebd2af8ae24619a8 · xxadev / tensorflow

13 6月, 2019 1 次提交
- I
  
  revert line 1 · 6e72b20b
  由 I-Hong 提交于 6月 12, 2019
  
  6e72b20b
07 6月, 2019 1 次提交
- I
  
  LICENSE: Restore original Apache license · 43cacae8
  由 I-Hong 提交于 6月 07, 2019
  
  43cacae8
06 6月, 2019 5 次提交
- G
  Make default Keras ConfigProto use tf.config · 26d3fe87
  由 Gaurav Jain 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251659257
```
  26d3fe87
- A
  [Grappler] Migrate FrameView to use utils::GraphView/utils::MutableGraphView. · d15c612f
  由 Andy Ly 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251659253
```
  d15c612f
- K
  Add serialization support for feature columns in DenseFeatures layers in Keras. · cd09510f
  由 Karmel Allison 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251656649
```
  cd09510f
- B
  Add alternative in high level TPU embedding API to not use feature columns,... · 14d492d7
  由 Bruce Fontaine 提交于 6月 05, 2019
```
Add alternative in high level TPU embedding API to not use feature columns, but to use the mid level API FeatureConfig and TableConfig instead.

PiperOrigin-RevId: 251656574
```
  14d492d7
- J
  [XLA:GPU] Handle AllReduce replica-groups. · c6244844
  由 Justin Lebar 提交于 6月 05, 2019
```
The AllReduce HLO has the ability to say, do an all-reduce on (say) GPUs [0,1]
and separately on GPUs [2,3].  Previously this was not implemented (and in fact
we incorrectly just ignored this and would do an all-reduce across all four
GPUs).

PiperOrigin-RevId: 251654069
```
  c6244844
05 6月, 2019 33 次提交
- A
  Add tile op support to the NN API delegate · 4978b66a
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251651006
```
  4978b66a
- B
  [XLA] Add support for F64 random number generation to the XLA client library. · 1d8ca9e1
  由 Bixia Zheng 提交于 6月 05, 2019
```
Extend the Philox bit generator and the Three Fry bit generator to support F64.

Change UniformF32Distribution to UniformFloatingpointDistribution to add
support for F64. Similarly, change NormalF32Distribution to
NormalFloatingpointDistribution.

Modify the tf2xla bridge to support F64 stateless random ops and stateful
random ops.

Modify the stateless random op tests and the stateful random op tests to test
F64.

PiperOrigin-RevId: 251649864
```
  1d8ca9e1
- J
  [XLA:GPU] Make TupleThunk faster by avoiding synchronous host/device callbacks. · 0bdfaf28
  由 Justin Lebar 提交于 6月 05, 2019
```
TupleThunk registers a callback to free the host buffer (i.e. the CPU memory
that holds the pointers) after the H2D memcpy completes.  Unfortunately these
callbacks are slow.

We add a new routine to StreamExecutor which runs a callback on the next
BlockHostUntilDone, and we use this to free the buffer.  This means that the
buffer will stay live for a bit longer, but in practice these buffers are tiny,
less than 32 bytes.

We also use this in ConvolutionThunk and CudnnBatchnormThunk.  Previously these
were unsoundly relying on an optimization in the GPU driver (?) in which it
seemed to eagerly copy the host memory.

PiperOrigin-RevId: 251646672
```
  0bdfaf28
- A
  Add Makefile for micro_vision tests. · d8e739dd
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251646660
```
  d8e739dd
- J
  [XLA:GPU] Change Thunk::ExecuteOnStream to take its parameters "by struct". · ae64752c
  由 Justin Lebar 提交于 6月 05, 2019
```
This makes it far easier to add a new parameter, which I will do in a later
change.

PiperOrigin-RevId: 251643760
```
  ae64752c
- E
  Update CompositeTensor API to use TypeSpec · 23425e62
  由 Edward Loper 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251640087
```
  23425e62
- A
  Add tile op tests with const input · 770481fb
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251637948
```
  770481fb
- A
  [XLA] Create helper for using reductions as slice-like operations. · 8b7052c4
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251630341
```
  8b7052c4
- A
  [TF:XLA] Broadcast NextAfter arguments when needed. · 58c796df
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251613307
```
  58c796df
- A
  [TF] Increase tolerance of elu tests on half precision numbers. · c47ad082
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
Previously more precision was required than is supported by 16 bit floats.

PiperOrigin-RevId: 251612304
```
  c47ad082
- A
  [TF:XLA] Create helper for broadcasting ops to same dimensions. · 6db4b700
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251603109
```
  6db4b700
- A
  Go: Update generated wrapper functions for TensorFlow ops. · 57c0440a
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251596458
```
  57c0440a
- P
  Make MirroredStrategy throw an error when creating a trainable ReplicaLocalVariable. · d5f641cb
  由 Peter Buchlovsky 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251596345
```
  d5f641cb
- A
  [TF:XLA] Support LeakyRelu for alpha values outside of [0,1] · f57f1ce4
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
PiperOrigin-RevId: 251594608
```
  f57f1ce4
- A
  [TF:XLA] Disable constant folding in tests. · 9eb67b17
  由 A. Unique TensorFlower 提交于 6月 05, 2019
```
Tests were passing when we don't actually support all the functionality.

PiperOrigin-RevId: 251592402
```
  9eb67b17
- S
  Add custom op support to calibrator. · c8fd0ce7
  由 Suharsh Sivakumar 提交于 6月 04, 2019
```
Also fix incorrectly set version while i am here.

PiperOrigin-RevId: 251583288
```
  c8fd0ce7
- A
  Automated rollback of commit f5e3a2cf · 19665c63
  由 Anna R 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251581726
```
  19665c63
- A
  Add API in RecursiveCompilabilityChecker to access un-compilable node if node... · 027714d4
  由 A. Unique TensorFlower 提交于 6月 04, 2019
```
Add API in RecursiveCompilabilityChecker to access un-compilable node if node is inside a potentially recursive function body node.

PiperOrigin-RevId: 251581339
```
  027714d4
- A
  Consume stderr from all processes in tests that fork multiple processes. · 2e66ef0e
  由 Ayush Dubey 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251580137
```
  2e66ef0e
- V
  Fix bug in partial & tf.function. · 51f8f449
  由 Vojtech Bardiovsky 提交于 6月 04, 2019
```
This was caused by wrong alignment of arguments and default values.

Also add one more assert to improve debuggability.

PiperOrigin-RevId: 251580104
```
  51f8f449
- P
  Deserializing loss class in hdf5 format. · 868f22ad
  由 Pavithra Vijay 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251579891
```
  868f22ad
- S
  Fix container names in TRTEngineOp · 69b121d7
  由 Smit Hinsu 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251578768
```
  69b121d7
- A
  Materialize output values deduced during shape inference, rather than folding the nodes again. · 51bc52d6
  由 A. Unique TensorFlower 提交于 6月 04, 2019
```
Save more temporary memory in remapper and memory optimizers.

Get rid of some code duplication in constant_folding.cc.

PiperOrigin-RevId: 251578727
```
  51bc52d6
- A
  Remove unnecessary configuration fields from optimization parameters · 07b77a51
  由 A. Unique TensorFlower 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251577571
```
  07b77a51
- A
  Automated rollback of commit 3125d75c · c64f92f6
  由 Anna R 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251572054
```
  c64f92f6
- T
  Don't `build` a Layer twice if `compute_output_shape` is called manually. · be57b1a6
  由 Thomas O'Malley 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251571946
```
  be57b1a6
- S
  Use tflite::SafeCast when casting float to int32. · a495211a
  由 Suharsh Sivakumar 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251571788
```
  a495211a
- T
  Make the default for `run_eagerly` the value of · 0054b6aa
  由 Thomas O'Malley 提交于 6月 04, 2019
```
`tf.config.experimental_run_functions_eagerly`

PiperOrigin-RevId: 251570860
```
  0054b6aa
- S
  Support model.fit and evaluate in 2.0 with TPUStrategy using the... · a1699237
  由 Sourabh Bajaj 提交于 6月 04, 2019
```
Support model.fit and evaluate in 2.0 with TPUStrategy using the experimental_run + train_on_batch API.

PiperOrigin-RevId: 251570029
```
  a1699237
- Z
  Fix description of input specs in assert_input_compatibility · 0ba31190
  由 Zhenyu Tan 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251569439
```
  0ba31190
- Y
  Change "toco error" to "converter error" in generated example tests · f79678ad
  由 Yu-Cheng Ling 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251568444
```
  f79678ad
- R
  [tf.data] Add experimental dataset ops as passthrough ops in prefetch with slack rewrite. · a9e61868
  由 Rachel Lim 提交于 6月 04, 2019
```
PiperOrigin-RevId: 251566961
```
  a9e61868
- A
  Exposes the already-implemented variable `scatter_div`, `scatter_mul`,... · 76171c3c
  由 A. Unique TensorFlower 提交于 6月 04, 2019
```
Exposes the already-implemented variable `scatter_div`, `scatter_mul`, `scatter_min`, and `scatter_max` operations in the variables API. Also fixes some documentation, and adds tests that directly check the var.scatter_* methods.

PiperOrigin-RevId: 251565839
```
  76171c3c

xxadev / tensorflow 与 Fork 源项目一致

xxadev / tensorflow
与 Fork 源项目一致