GitLab

All models which support loss scaling support dynamic loss scaling, so the argument has no purpose. It used to be that some models scaled the loss manually instead of using a LossScaleOptimizer, and so did not support dynamic loss scaling.

PiperOrigin-RevId: 367719521

The function `tf.train.experimental.enable_mixed_precision_graph_rewrite` will be removed from the TF2 namespace soon, at which point it will only be accessible under tf.compat.v1.

PiperOrigin-RevId: 367046393

PiperOrigin-RevId: 359994674

PiperOrigin-RevId: 359990341

PiperOrigin-RevId: 326286926

PiperOrigin-RevId: 326286926

PiperOrigin-RevId: 299160422

PiperOrigin-RevId: 297002741

PiperOrigin-RevId: 277082247

To test, I did 50 fp32 runs and 50 fp16 runs. I used the following command:

python ncf_keras_main.py --dataset=ml-20m --num_gpus=1 --train_epochs=10 --clean --batch_size=99000 --learning_rate=0.00382059 --beta1=0.783529 --beta2=0.909003 --epsilon=1.45439e-7 --layers=256,256,128,64 --num_factors=64 --hr_threshold=0.635 --ml_perf --nouse_synthetic_data --data_dir ~/ncf_data_dir_python3 --model_dir ~/tmp_model_dir --keras_use_ctl

For the fp16 runs, I added --dtype=fp16. The average hit-rate for both fp16 and fp32 was 0.6365. I also did 50 runs with the mixed precision graph rewrite, and the average hit-rate was 0.6363. The difference is likely due to noise.

PiperOrigin-RevId: 275059871

PiperOrigin-RevId: 273371605

PiperOrigin-RevId: 266376708

--synthetic_data, --dtype, --all_reduce_alg, and --num_packs have been unexposed from models which do not use them

PiperOrigin-RevId: 265483564

--num_parallel_calls, --inter_op_parallelism_threads, and --intra_op_parallelism_threads have been unexposed from models which do not use them

PiperOrigin-RevId: 264965788

Only the V1 resnet model uses --max_train_steps. This unexposes the flag in the keras_application_models, mnist, keras resnet, CTL resnet Models. Before this change, such models allowed the flag to be specified, but ignored it.

I also removed the "max_train" argument from the run_synthetic function, since this only had any meaning for the V1 resnet model. Instead, the V1 resnet model now directly passes --max_train_steps=1 to run_synthetic.

PiperOrigin-RevId: 264269836

* force_v2_in_keras_compile FLAG default to None and added seperate temp path.

* switch to force testing 1v path not force v2 path.

* Rename function force_v1_path.

* Add force_run_distributed tests.

* Added enable_eager

* r/force_run_distributed/force_v2_in_keras_compile

* Adding force_v2 tests and FLAGs.

* Rename method to avoid conflict.

* Add cpu force_v2 tests.

* fix lint, wrap line.

* change to force_v2_in_keras_compile

* Update method name.

* Lower mlperf target to 0.736.

* 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>

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.

* 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

* Add FP16 and benchmarks.

* add missing run and report.

* Add loss_scale as option not included with dtype.

* move loss_scale validation under dtype conditional.

* add loss_scale to flags tested.

This options allows the new tf.train.experimental.enable_mixed_precision_graph_rewrite() function to be used for fp16, instead of manual casts.

* Add num_packs flag for MirroredStrategy's cross device ops.

* fix parens

* Fix lint errors and make all_reduce_alg more robust.

* Set default num_packs to 1

* s/CollectiveAllReduceStrategy/MultiWorkerMirroredStrategy

* More s/contrib.distribute/distribute.experimental

* Collective communication options in MultiWorkerMirroredStrategy.

* Minor fixes

* No checkpointing if multi worker.

* turn off checkpointing

* fix lint

* Add DistributionStrategy to transformer model

* add num_gpu flag

* Calculate per device batch size for transformer

* remove reference to flags_core

* Add synthetic data option to transformer

* fix typo

* add import back in

* Use hierarchical copy

* address PR comments

* lint

* fix spaces

* group train op together to fix single GPU error

* Fix translate bug (sorted_keys is a dict, not a list)

* Change params to a default dict (translate.py was throwing errors because params didn't have the TPU parameters.)

* Address PR comments. Removed multi gpu flag + more

* fix lint

* fix more lints

* add todo for Synthetic dataset

* Update docs

* squash of modular absl usage commits

* delint

* address PR comments

* change hooks to comma separated list, as absl behavior for space separated lists is not as expected