GitLab

* Improve performance for Cifar ResNet benchmarks

* Revert batch size changes to benchmarks

* Fix unit tests failures.

* 96% of TF 2.0 tests on GPU are passing.

* Currently all passing GPU and CPU TF 2.0

* Address code comments.

* use tf 2.0 cast.

* Comment about working on TF 2.0 CPU

* Uses contrib turn off for TF 2.0.

* Fix wide_deep and add keras_common_tests.

* use context to get num_gpus.

* Switch to tf.keras.metrics

* cpu benchmark and accuracy tests.

* add docstrings to fix lint.

* XLA FP32 and first test

* More XLA benchmarks FP32.

* Add eager to NCF and refactor resnet.

* fix v2_0 calls and more flag refactor.

* Remove extra flag args.

* 90 epoch default

* add return

* remove xla not used by estimator.

* Remove duplicate run_eagerly.

* fix flag defaults.

* Remove fp16_implementation flag option.

* Remove stop early on mlperf test.

* remove unneeded args.

* load flags from keras mains.

* Do not set learning phase when skipping eval

* Do not set learning phase in no dist strat case

* Added device placement, tweaked benchmarks

* Added tweaked benchmarks for Cifar

* Fix device scope

* Fix lint

* Add explicit GPU placement flag

* Also run accuracy test with explicit GPU placement

* Added doc string

* set default steps to 300K.

* Log flags to perfzero.

* Add XLA support to transformer

- Moved config logic to keras_utils
- Added enable_xla flag to _performance flags
- Did not refactor enable_xla flag from keras resnet due to
  reliance on calling FLAGs in estimator keras and that is
  a needed refactor for another time.

* fix g3 lint complaint.

* Refactor set config into keras_utils.

* Move flags out of main.

* pipe through enable_xla

* Update official/transformer/v2/misc.py
Co-Authored-By: Reed <reedwm@google.com>

* layout off for some tests and channels last.

* 8 gpu tests channels_last

* more layout off tests.

* Add 1 gpu force_eager benchmark

* Add accuracy for no dist strat eager

* remvove return.

Before, there was a global default loss scale for all models. Currently, only resnet uses loss scaling, but this will be useful once more models support it.

* Modify tweaked tests for better performance in no cloning mode

* Tweak trivial models

Because we run warmup tests in all real data benchmarks, XLA bugs will cause non-XLA tests to fail as well.

* Add mlperf like test.

* Final comments.

* docstring wording tweak.

* non-tweaked version

* Support pure eager execution in ResNet50

* Use smaller batch size

* Run different numbers of steps on different platforms

* Add new tests for delayed performance measurement

* Add a graph optional_next Reset benchmark.

* Fix lint error.

* Merged commit includes the following changes:
249776315  by tianlin<tianlin@google.com>:

    Internal change

249763206  by tianlin<tianlin@google.com>:

    For TF 2.0 (related to Beam Search), expand cond dims in tf.where(cond, x, y) to make all parameters broadcastable.

--
249392724  by hongkuny<hongkuny@google.com>:

    Internal change

PiperOrigin-RevId: 249776315

* Merged commit includes the following changes:
249823043  by tianlin<tianlin@google.com>:

    Bring back v2 test for predict and eval.

--

PiperOrigin-RevId: 249823043

* Add a test enabling get_next_as_optional behavior.

* Remove repeated flag.

* Remove trailing space.

* Make the name shorter.

* Fix lint error.

* Refine the benchmark name.

* Add enable_get_next_as_optional flag.

* Set enable_get_next_as_optional to strategy.

* Add comments to explain the flag.

* Remove trailing whitespace.

* Remove trailing space.

* Set the --clone_model_in_keras_dist_strat to None.  Remove the separate no_cloning benchmarks and add a couple of cloning ones.  Fixes the learning rate schedule to cache its ops per graph.

* Added 'tfdata_exp' version of all benchmarks which set
FLAGS.tf_data_experimental_slack = True. Renamed
`data_prefetch_with_slack` to `data_delay_prefetch` (haoyu's change)
to make the names more distinct.

* Add flag to resnet input pipeline and surface through
keras_imagenet_main.py

* Fix trivial model to work properly with fp16

* Add comment on manual casting

Previously we had one dense layer in trivial model. The weight was [224*224*3, num_classes].
Using two dense layers, the weights are [224*224*3, 1] and [1, num_classes].

* Do not report metrics in performance benchmarks

* Rename flag

* Modified tweaked tests to use tensor learning rate