GitLab

* borrowing a tf1.x optimization which converts gradients from sparse to dense for better perf

* cleanup after code review

* 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.

* adding a new perf test for ncf, and changing some names

* Added change to make ncf use the data from the gcp bucket, and removed the need to re-download data >1day old. Reorganized the perf-zero tests

* Add CTL benchmark

* Divide train loss by number of train steps

* increase num epochs to 10

* add benchmark for early stopping with CTL

* remove whitespace

* Fix various lint errors

* Fix logging format

* Add flag to use custom training loop for keras NCF model.

* Add error check to NCF model for custom training loop + tf1.0.

* Add a custom training loop for NCF model with TF2.0

* Fix long line in ncf_keras_main.py

* Remove dataset repeat when using custom training loop.

Add early stopping logic to ncf keras when desired threshold is met. Also change the default batch size to match the tuned hyperparams

* 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

Adding validation every epoch allows us to view the progress during training instead of having to wait until the last eval. Mostly useful for manual runs.

Current batch size 160000 does not converge to the desired HR. So we decrease to 99k which is known to converge. Tested locally and got to 63.5 at epoch 7. Also decreasing number of epochs as I don't see any improvement after epoch 7-8.

* 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.

Replace per_device with per_replica and PerDevice with PerReplica, because the PerDevice concept was renamed and doesn't exist anymore. (#6693)

* Replace per_device with per_replica and PerDevice with PerReplica, because the PerReplica concept was renamed and doesn't exist anymore.

* Add accuracy check.

* Avoid double flag init, move data_dir to real data.

* Comment on lower accuracy target.