Commits · 31ca3b97ebc1ca37b1d4db6ff3bf062fcbf16b5d · ModelZoo / ResNet50_tensorflow

08 Jul, 2020 1 commit
- Remove TODOs that will never fulfill. · 3b56ba8d
  Hongkun Yu authored Jul 07, 2020
```
PiperOrigin-RevId: 320124801
```
  3b56ba8d
04 Jul, 2020 1 commit
- Update deprecation message to be more clearer. · 523f5e05
  Hongkun Yu authored Jul 03, 2020
```
PiperOrigin-RevId: 319557230
```
  523f5e05
17 Jun, 2020 1 commit
- Update model_training_utils for BERT. · d0ef3913
  A. Unique TensorFlower authored Jun 16, 2020
```
PiperOrigin-RevId: 316801831
```
  d0ef3913
16 Jun, 2020 1 commit
- Internal cleanup. · 85cfe94d
  Tianqi Liu authored Jun 16, 2020
```
PiperOrigin-RevId: 316734574
```
  85cfe94d
02 Jun, 2020 1 commit
- A minor change to model_training_utils of BERT. · fc724b3e
  A. Unique TensorFlower authored Jun 02, 2020
```
PiperOrigin-RevId: 314417394
```
  fc724b3e
26 May, 2020 1 commit
- Add learning_rate to the summary. · ee760015
  Maxim Neumann authored May 26, 2020
```
PiperOrigin-RevId: 313148142
```
  ee760015
19 May, 2020 2 commits

When not specified by user, make steps_per_loop be dependent on the devices available. · 866176ea

André Susano Pinto authored May 19, 2020

A default of 1 all the times is bad for TPU users which end up not using the device effectively.
A larger default all the times is bad for GPU users.

So compromise and make this dependent on the devices available.

PiperOrigin-RevId: 312230371

866176ea

Internal change · 2a3971dd
A. Unique TensorFlower authored May 18, 2020
```
PiperOrigin-RevId: 312209760
```
2a3971dd

14 May, 2020 1 commit
- Deprecate run_customized_training_loop() method. · 7cdb82e3
  Chen Chen authored May 14, 2020
```
PiperOrigin-RevId: 311577663
```
  7cdb82e3
12 May, 2020 1 commit
- Internal change · adb61343
  Chen Chen authored May 11, 2020
```
PiperOrigin-RevId: 311069693
```
  adb61343
22 Apr, 2020 1 commit
- Add model metrics to the training summary. · 22f10c60
  Chen Chen authored Apr 22, 2020
```
PiperOrigin-RevId: 307914609
```
  22f10c60
16 Apr, 2020 1 commit

Add calling of custom callback on_epoch_begin() and on_epoch_end() methods. · 31da2245

Jeremiah Harmsen authored Apr 16, 2020

Minor refactoring to (1) encapsulate callbacks in a CallbackList and (2) allow evaluation results to be threaded to on_epoch_end callback.

PiperOrigin-RevId: 306827406

31da2245

15 Apr, 2020 2 commits

Internal change · 89599a23
Le Hou authored Apr 15, 2020
```
PiperOrigin-RevId: 306664983
```
89599a23

Remove extraneous evaluation & export which happens on the final epoch. · 8931d298

Jeremiah Harmsen authored Apr 15, 2020

Preserves current behavior for submodels of exporting duplicates (e.g., for 100 total steps, my_submodel_step_100.ckpt and my_submodel.ckpt are equivalent and both written after the final epoch).

PiperOrigin-RevId: 306629202

8931d298

26 Mar, 2020 1 commit
- Move modeling_training_utils.py to bert folder · 6ef71d49
  Hongkun Yu authored Mar 25, 2020
```
PiperOrigin-RevId: 303050809
```
  6ef71d49
25 Mar, 2020 1 commit

Check if there are GPUs instead of if TF is built with CUDA support. · ad09cf49

Reed Wanderman-Milne authored Mar 25, 2020

The TF pip packages are always built with CUDA support, so tf.test.is_built_with_cuda() would return True even if the user had no GPU.

PiperOrigin-RevId: 302928378

ad09cf49

17 Mar, 2020 2 commits
- Adds scale_loss to run_customized_training_loop, which should be the correct treatment. · bf6d6f6f
  Hongkun Yu authored Mar 17, 2020
```
PiperOrigin-RevId: 301441181
```
  bf6d6f6f
- Internal change · 2d7223ad
  Ken Franko authored Mar 16, 2020
```
PiperOrigin-RevId: 301277545
```
  2d7223ad
15 Mar, 2020 1 commit
- Correct grammar. · adc01cd7
  A. Unique TensorFlower authored Mar 15, 2020
```
PiperOrigin-RevId: 301041987
```
  adc01cd7
13 Mar, 2020 1 commit
- Remove confusing legacy comments. · 5a423a2d
  Hongkun Yu authored Mar 12, 2020
```
PiperOrigin-RevId: 300693870
```
  5a423a2d
11 Mar, 2020 1 commit
- Add TimeHistory callback to BERT. · b39958cf
  Will Cromar authored Mar 11, 2020
```
PiperOrigin-RevId: 300433601
```
  b39958cf
10 Mar, 2020 1 commit

Save to tmp directory on non-chief workers in model_training_utils · 682d36b5

Ran Chen authored Mar 10, 2020

In a multi worker set up saving is done on each worker. If they're saving to the same location, e.g. GCS, there will be conflicts. With this change we save to temporary directory on non-chief workers.

Note that, there may be synchronization in saving that needs all workers to participate, so we cannot only save on one worker.

PiperOrigin-RevId: 300141152

682d36b5

07 Mar, 2020 1 commit
- Add TimeHistory callback to BERT. · 7d86c317
  Hongkun Yu authored Mar 07, 2020
```
PiperOrigin-RevId: 299594839
```
  7d86c317
06 Mar, 2020 1 commit

Temporarily disable explicit allreduce in BERT SQuAD · 11ccb99e

Zongwei Zhou authored Mar 05, 2020

In BERT SQuAD, disable explicit allreduce for now to keep the original clip_by_global_norm math. With explicit allreduce, the gradients before allreduce are scaled so even if we move clip_by_global_norm before allreduce (as in TF1 and pre-TF 2.2) it will operate on scaled gradients, the math will be changed. So with explicit allreduce, it is better to move clip_by_global_norm to after allreduce.

PiperOrigin-RevId: 299278082

11ccb99e

05 Mar, 2020 1 commit
- Internal change · cf01596c
  Zongwei Zhou authored Mar 04, 2020
```
PiperOrigin-RevId: 299007295
```
  cf01596c
02 Mar, 2020 2 commits
- Add TimeHistory callback to BERT. · 533d1e6b
  Will Cromar authored Mar 02, 2020
```
PiperOrigin-RevId: 298466825
```
  533d1e6b
- Remove an assert. · 98abe4b8
  Hongkun Yu authored Mar 02, 2020
```
PiperOrigin-RevId: 298402269
```
  98abe4b8
26 Feb, 2020 1 commit
- Internal change · ce83a9db
  Hongkun Yu authored Feb 26, 2020
```
PiperOrigin-RevId: 297383836
```
  ce83a9db
25 Feb, 2020 1 commit
- Internal change · 2781377d
  Zongwei Zhou authored Feb 25, 2020
```
PiperOrigin-RevId: 297222995
```
  2781377d
20 Feb, 2020 1 commit
- Remove some tpu_lib not used at all · 58df0dc6
  Hongkun Yu authored Feb 19, 2020
```
PiperOrigin-RevId: 296140699
```
  58df0dc6
19 Feb, 2020 1 commit
- Use host python training loop in GPU BERT tests · 1ca9e3e4
  Zongwei Zhou authored Feb 18, 2020
```
PiperOrigin-RevId: 295869937
```
  1ca9e3e4
13 Feb, 2020 1 commit
- Revert log passing change since it might hurt performance. · 91c681af
  Haitang Hu authored Feb 13, 2020
```
PiperOrigin-RevId: 294922828
```
  91c681af
07 Feb, 2020 1 commit

Pass training_loss in logs dict to customized training loop. · 67f6015a

Haitang Hu authored Feb 07, 2020

This would match the behavior described in on_batch_end() functions in Keras callback.
See https://www.tensorflow.org/api_docs/python/tf/keras/callbacks/Callback.

PiperOrigin-RevId: 293887583

67f6015a

23 Nov, 2019 1 commit
- Internal change · 83f0a576
  Chen Chen authored Nov 22, 2019
```
PiperOrigin-RevId: 282096004
```
  83f0a576
18 Nov, 2019 1 commit
- Internal change · 81d031d0
  Hongkun Yu authored Nov 18, 2019
```
PiperOrigin-RevId: 281117886
```
  81d031d0
28 Oct, 2019 1 commit
- remove use_remote_tpu as it is deprecated. · 4457c1a8
  Hongkun Yu authored Oct 28, 2019
```
PiperOrigin-RevId: 277087491
```
  4457c1a8
18 Oct, 2019 1 commit
- Fix xlnet save_steps and steps_per_epoch conflicts · e6750c5d
  Hongkun Yu authored Oct 18, 2019
```
Remove train_data_size flag

PiperOrigin-RevId: 275545035
```
  e6750c5d
11 Oct, 2019 1 commit

Change summary directory and model checkpoint directory so that training via... · e67a2064

A. Unique TensorFlower authored Oct 11, 2019

Change summary directory and model checkpoint directory so that training via Keras Compile/Fit() and custom training loop is consistent.

PiperOrigin-RevId: 274202793

e67a2064

07 Oct, 2019 1 commit
- Internal change · 482784ab
  Jing Li authored Oct 07, 2019
```
PiperOrigin-RevId: 273233857
```
  482784ab
24 Sep, 2019 1 commit
- Use experimental_connect_to_cluster API in TPU lib to support training on a slice of a TPU pod. · 497989e0
  Bruce Fontaine authored Sep 24, 2019
```
PiperOrigin-RevId: 270926016
```
  497989e0