GitLab

Also, do Transformer inference in fp16, as well as training, when --dtype=fp16. In TF 2, layers now cannot run in multiple different dtypes, so we must use the same dtype for training and inference.

* trying fake merge call

* make metrics optional

* Remove extra print

* Fixes that make transformer run.

* Remove debug print statements.

* Changed the permissions to 644.

* Fix the rest of the permissions.

* enable static batch in all benchmarks

* Restrict dist strat hack to training mode

For now we will do predict/eval without dist strat, so remove that hack in non training cases.

* Use `inputs` instead of `x` as arg name for call

Keras has different behavior based on whether the inputs are called `inputs` or not. Using `inputs` gives expected behaviors.

* Avoid extra map fn on input in dist strat case

* Update how we handle custom metrics

This new approach works with and without dist strat. The previous one didn't work with dist strat. We need to fix that but this is reasonable in meantime (b/133724664).

* Update benchmarks

* typo in metrics code

* Revert metrics change

Didn't actually work in distributed case..

* 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

* Merged commit includes the following changes:
249218656  by tianlin<tianlin@google.com>:

    Deal with imports, fix a typo and make unit tests fast.

--
249198645  by tianlin<tianlin@google.com>:

    Trivial: Remove one empty line before "import tensorflow"

--
249195490  by tianlin<tianlin@google.com>:

    Initialize Transformer TF V2 Model with Keras subclassing implementation. (Compatible with TF V1)

--
249195008  by tianlin<tianlin@google.com>:

    Internal change

249173564  by hongkuny<hongkuny@google.com>:

    Internal change

249079258  by hongkuny<hongkuny@google.com>:

    Internal change

247691534  by haoyuzhang<haoyuzhang@google.com>:

    Internal change

247533725  by haoyuzhang<haoyuzhang@google.com>:

    Internal change

247509295  by haoyuzhang<haoyuzhang@google.com>:

    Internal change

247311355  by wangtz<wangtz@google.com>:

    Internal change

247303127  by wangtz<wangtz@google.com>:

  ...

* port changes from previous branch now that transformer util changes are in master

fix incorrect count

correct (hopefully) treatment of batch_size

set eval_metrics to a dummy function for now

add some comments

start bringing metrics to transformer TPU

resolve logits shape

metrics are now working except for tf.py_func metrics

increase batch_size for tpu, and create summary host call

fix host call

reduce tpu default batch size

further tune batch sizes

add minibatch loss to summary

handle case of single_iteration_train_steps > number points in an epoch

begin to incorporate hooks

add sleep workarounds

disable hooks altogether

generalize host call function and move to newly created tpu utils module

remove all traces of params as an object

switch from  to

address some PR comments, and change the number of data points.

minor tweaks

add tpu dry run for testing, and use matmul for TPU embedding

infeed/outfeed queue issue is fixed. Sleeps are no longer necessary

add some documentation.

cleanup and address PR comments

delint

add accelerator __init__

fix embedding

missed PR comment

address PR comments

fix validator bug

rewrite cloud storage validator, and add oauth dependency to requirements.txt

* delint