0d464ff | Lukasz Kaiser | 12 February 2018, 03:18:28 UTC | Merge pull request #576 from rsepassi/push v1.5.0 | 12 February 2018, 03:18:28 UTC |
88a3c9b | Lukasz Kaiser | 11 February 2018, 23:47:10 UTC | Extend Cloud TPU docs. PiperOrigin-RevId: 185318516 | 12 February 2018, 02:59:25 UTC |
30a79e8 | Ryan Sepassi | 11 February 2018, 21:40:48 UTC | Add models/research/__init__.py PiperOrigin-RevId: 185314217 | 11 February 2018, 21:48:08 UTC |
08f0742 | Lukasz Kaiser | 11 February 2018, 20:40:29 UTC | Add a quick MNIST model for README, comments and documentation, bump back version. PiperOrigin-RevId: 185311943 | 11 February 2018, 20:50:30 UTC |
70a0464 | Lukasz Kaiser | 11 February 2018, 20:11:19 UTC | Move more researchy models under models/research. PiperOrigin-RevId: 185311067 | 11 February 2018, 20:17:17 UTC |
57a27fe | Lukasz Kaiser | 10 February 2018, 22:37:09 UTC | Increase release number. PiperOrigin-RevId: 185267399 | 10 February 2018, 23:21:46 UTC |
d23d241 | Lukasz Kaiser | 10 February 2018, 22:34:39 UTC | internal merge PiperOrigin-RevId: 185267339 | 10 February 2018, 23:21:40 UTC |
9d4c00f | T2T Team | 10 February 2018, 00:15:46 UTC | Shorten the feature name for expected_attentions, too verbose. PiperOrigin-RevId: 185208468 | 10 February 2018, 21:20:58 UTC |
61906cc | Ryan Sepassi | 09 February 2018, 22:09:50 UTC | Fix hyperparameter prefixing for ML Engine autotuning PiperOrigin-RevId: 185190665 | 10 February 2018, 21:20:51 UTC |
4cdbc09 | Ryan Sepassi | 09 February 2018, 19:56:43 UTC | Small var name changes PiperOrigin-RevId: 185172558 | 10 February 2018, 21:20:40 UTC |
944cfbe | Ryan Sepassi | 09 February 2018, 18:47:51 UTC | Update jobId for Cloud ML Engine PiperOrigin-RevId: 185162214 | 10 February 2018, 21:20:27 UTC |
85c60db | Lukasz Kaiser | 09 February 2018, 01:38:29 UTC | Correct latent predicton input in transformer ae. PiperOrigin-RevId: 185079318 | 10 February 2018, 21:20:21 UTC |
a7fbd2e | Ryan Sepassi | 09 February 2018, 01:33:43 UTC | Fix demo VanillaGan model PiperOrigin-RevId: 185078802 | 10 February 2018, 21:20:16 UTC |
52509fb | Ryan Sepassi | 09 February 2018, 00:22:12 UTC | Fix call order for rouge (issue #541) PiperOrigin-RevId: 185070260 | 10 February 2018, 21:20:10 UTC |
73f034d | Ryan Sepassi | 08 February 2018, 23:32:46 UTC | Update License year PiperOrigin-RevId: 185063168 | 10 February 2018, 21:20:03 UTC |
eaefc32 | Lukasz Kaiser | 10 February 2018, 04:39:40 UTC | Merge pull request #568 from deepsense-ai/gae_support RL Improvements | 10 February 2018, 04:39:40 UTC |
28d1841 | Lukasz Kaiser | 10 February 2018, 04:24:12 UTC | Merge pull request #572 from nadavb/master Pretty print and sort Hparams | 10 February 2018, 04:24:12 UTC |
ad30518 | Błażej O | 09 February 2018, 23:20:16 UTC | Cleaning old imports. | 09 February 2018, 23:20:16 UTC |
1bb6fb4 | Błażej O | 09 February 2018, 23:16:20 UTC | Renaming config for hparams. | 09 February 2018, 23:17:07 UTC |
5eb0f96 | Błażej O | 09 February 2018, 23:10:08 UTC | Integrating networks into models/rl. | 09 February 2018, 23:10:08 UTC |
32783a2 | Błażej O | 09 February 2018, 16:22:24 UTC | Masking unused argument. | 09 February 2018, 16:22:24 UTC |
f1c837f | Błażej O | 09 February 2018, 16:15:23 UTC | Update location of t2t_rl_trainer, everywhere. | 09 February 2018, 16:15:23 UTC |
16b3d2b | Błażej O | 09 February 2018, 16:13:03 UTC | Update location of t2t_rl_trainer. | 09 February 2018, 16:13:03 UTC |
7b59450 | Błażej O | 09 February 2018, 16:01:22 UTC | Adding Cartpole environment with discrete actions. | 09 February 2018, 16:01:22 UTC |
93abff7 | Błażej O | 09 February 2018, 15:08:51 UTC | Reintroducing t2t-rl-trainer. | 09 February 2018, 15:13:17 UTC |
4a1ffac | Nadav Benedek | 09 February 2018, 13:29:30 UTC | Pretty print and sort Hparams | 09 February 2018, 13:29:30 UTC |
80ed41e | Błażej O | 09 February 2018, 12:55:04 UTC | RL test should be run separately on Travis. | 09 February 2018, 12:55:04 UTC |
58915a4 | Błażej O | 09 February 2018, 12:11:08 UTC | Fixing merge. | 09 February 2018, 12:11:08 UTC |
cae75b7 | Błażej O | 09 February 2018, 12:08:06 UTC | Merge branch 'master' into gae_support | 09 February 2018, 12:08:06 UTC |
8f76880 | Błażej O | 09 February 2018, 11:52:59 UTC | Fixing rl test. | 09 February 2018, 11:52:59 UTC |
80b2f73 | Lukasz Kaiser | 08 February 2018, 23:02:37 UTC | Merge pull request #570 from rsepassi/push v1.4.4 | 08 February 2018, 23:02:37 UTC |
290a12a | Ryan Sepassi | 08 February 2018, 21:57:00 UTC | v1.4.4 PiperOrigin-RevId: 185046879 | 08 February 2018, 22:27:47 UTC |
7778149 | Noam Shazeer | 08 February 2018, 21:30:45 UTC | Revert "noam" learning-rate-scheme to use linear warmup. Add learning_rate_schedule hparam to specify a schedule that does not have separate warmup and decay phases. PiperOrigin-RevId: 185042750 | 08 February 2018, 21:38:26 UTC |
5e8bc75 | Lukasz Kaiser | 08 February 2018, 19:41:48 UTC | Shake-shake config for TPU. PiperOrigin-RevId: 185026739 | 08 February 2018, 21:38:19 UTC |
c03fd19 | Ryan Sepassi | 08 February 2018, 19:16:13 UTC | Support hyperparameter tuning on Cloud ML Engine (reapply) PiperOrigin-RevId: 185022560 | 08 February 2018, 21:37:51 UTC |
ec364a1 | Lukasz Kaiser | 08 February 2018, 19:15:20 UTC | internal merge PiperOrigin-RevId: 185022378 | 08 February 2018, 21:37:39 UTC |
bbee177 | T2T Team | 08 February 2018, 17:31:26 UTC | Changes to make Librispeech TPU friendly. PiperOrigin-RevId: 185006656 | 08 February 2018, 21:37:31 UTC |
6ae03f7 | T2T Team | 08 February 2018, 16:54:31 UTC | Support hyperparameter tuning on Cloud ML Engine PiperOrigin-RevId: 185001360 | 08 February 2018, 21:37:19 UTC |
424d4ac | Ryan Sepassi | 08 February 2018, 02:44:53 UTC | Add documentation for Cloud ML Engine PiperOrigin-RevId: 184932319 | 08 February 2018, 21:37:12 UTC |
3b1f70f | Ryan Sepassi | 08 February 2018, 01:53:21 UTC | Fix hello_t2t.ipynb PiperOrigin-RevId: 184926720 | 08 February 2018, 21:37:01 UTC |
dd3322c | Ryan Sepassi | 08 February 2018, 01:21:04 UTC | Add Travis test for export and serve PiperOrigin-RevId: 184922812 | 08 February 2018, 21:36:53 UTC |
b86ed58 | Ryan Sepassi | 08 February 2018, 01:11:53 UTC | Support hyperparameter tuning on Cloud ML Engine PiperOrigin-RevId: 184921645 | 08 February 2018, 21:36:44 UTC |
0c3218d | T2T Team | 07 February 2018, 23:27:28 UTC | The most basic version of the "supervised" attention experiment. PiperOrigin-RevId: 184906398 | 08 February 2018, 21:36:37 UTC |
62bc2e7 | Ryan Sepassi | 07 February 2018, 20:23:29 UTC | Add a test for the t2t_trainer script PiperOrigin-RevId: 184877423 | 08 February 2018, 21:36:23 UTC |
963f44c | Ryan Sepassi | 07 February 2018, 19:44:48 UTC | Support --t2t_usr_dir on Cloud ML Engine PiperOrigin-RevId: 184871178 | 08 February 2018, 21:36:17 UTC |
2cde9c2 | Lukasz Kaiser | 07 February 2018, 19:02:02 UTC | Clean up obsolete ae cifar hparams, increase default batch size. PiperOrigin-RevId: 184863791 | 08 February 2018, 21:36:04 UTC |
0b2c8c2 | Ryan Sepassi | 07 February 2018, 18:22:43 UTC | Add dep on google-api-python-client PiperOrigin-RevId: 184857261 | 08 February 2018, 21:35:45 UTC |
416f784 | Ryan Sepassi | 07 February 2018, 04:15:50 UTC | Support launching T2T on --cloud_mlengine with CPU, GPU, and multi-GPU machines PiperOrigin-RevId: 184782692 | 08 February 2018, 21:35:37 UTC |
361f5d1 | Lukasz Kaiser | 07 February 2018, 02:25:12 UTC | Use beam search for latent prediction in transformer ae (only for num_blocks=1 for now). PiperOrigin-RevId: 184774591 | 08 February 2018, 21:35:22 UTC |
f044756 | Ryan Sepassi | 06 February 2018, 23:07:36 UTC | T2TModel.has_input checks for "inputs" PiperOrigin-RevId: 184747531 | 08 February 2018, 21:34:55 UTC |
21c9fed | T2T Team | 06 February 2018, 22:09:08 UTC | Add unmasked version of memory reduced attention PiperOrigin-RevId: 184737000 | 08 February 2018, 21:34:37 UTC |
887c410 | T2T Team | 06 February 2018, 20:35:56 UTC | Rename hierarchical vq-vae to decomposed vq-vae as in paper. PiperOrigin-RevId: 184722477 | 08 February 2018, 21:33:59 UTC |
798c406 | Ashish Vaswani | 06 February 2018, 01:22:47 UTC | Adding common_image_attention, a set of functions that support generating image. Jointly programmed by nikp and avaswani PiperOrigin-RevId: 184612638 | 08 February 2018, 21:33:39 UTC |
9e7a4d1 | T2T Team | 05 February 2018, 23:59:44 UTC | Fix bug with instability in higher z_sizes caused by math.log PiperOrigin-RevId: 184600621 | 08 February 2018, 21:33:29 UTC |
8057abd | T2T Team | 05 February 2018, 21:09:03 UTC | Use z_size instead of v_size for VQ-VAE PiperOrigin-RevId: 184574565 | 08 February 2018, 21:33:17 UTC |
2031610 | Ryan Sepassi | 05 February 2018, 20:44:50 UTC | Reset TPU input partitioning when switching modes PiperOrigin-RevId: 184570928 | 08 February 2018, 21:32:23 UTC |
3843dd2 | Lukasz Kaiser | 05 February 2018, 05:09:44 UTC | Revert to appending latents in transformer ae as it works better. PiperOrigin-RevId: 184480516 | 08 February 2018, 21:32:16 UTC |
522d222 | T2T Team | 04 February 2018, 15:55:04 UTC | Omit newline characters form librispeech targets. Add option to preprocess waveforms in the SpeechModality.bottom PiperOrigin-RevId: 184448336 | 08 February 2018, 21:32:05 UTC |
4bc354c | Ashish Vaswani | 03 February 2018, 23:24:51 UTC | Internal. PiperOrigin-RevId: 184412973 | 08 February 2018, 21:32:00 UTC |
17ddd3e | T2T Team | 02 February 2018, 18:30:13 UTC | make config argument optional in problem.input_fn PiperOrigin-RevId: 184301328 | 08 February 2018, 21:29:35 UTC |
1203853 | Noam Shazeer | 02 February 2018, 08:03:52 UTC | Implement fast decoding of transformer for language modeling problems (no inputs, partial targets). PiperOrigin-RevId: 184248180 | 08 February 2018, 21:29:23 UTC |
b9900e5 | Lukasz Kaiser | 02 February 2018, 02:01:11 UTC | Tune transformer ae hparams. PiperOrigin-RevId: 184224758 | 08 February 2018, 21:28:57 UTC |
41ef3e7 | Błażej O | 08 February 2018, 17:22:23 UTC | Improved version of saving summaries. | 08 February 2018, 17:22:49 UTC |
0a4a96c | Błażej O | 08 February 2018, 16:12:32 UTC | First version of generating summaries. | 08 February 2018, 16:49:20 UTC |
1498f52 | Błażej O | 07 February 2018, 16:16:13 UTC | Implementing Generalized Advantage Estimator. | 08 February 2018, 11:15:47 UTC |
1c98b8e | Lukasz Kaiser | 07 February 2018, 22:27:14 UTC | Merge pull request #566 from deepsense-ai/rl_init Initial commit of reinforcement learning module. | 07 February 2018, 22:27:14 UTC |
3707499 | Piotr Milos | 05 February 2018, 14:04:22 UTC | Initial commit of reinforcement learning module. | 07 February 2018, 15:52:21 UTC |
103d057 | Lukasz Kaiser | 02 February 2018, 01:45:46 UTC | Merge pull request #552 from rsepassi/push v1.4.3 | 02 February 2018, 01:45:46 UTC |
1cd3f25 | Ryan Sepassi | 02 February 2018, 00:52:20 UTC | Revert hparam change to versioned set PiperOrigin-RevId: 184216942 | 02 February 2018, 01:31:06 UTC |
07d8308 | Ryan Sepassi | 02 February 2018, 00:37:10 UTC | Travis test fixes PiperOrigin-RevId: 184214706 | 02 February 2018, 01:30:58 UTC |
d90a0c4 | Ryan Sepassi | 02 February 2018, 00:09:59 UTC | v1.4.3 PiperOrigin-RevId: 184210799 | 02 February 2018, 00:24:42 UTC |
da0bc49 | Lukasz Kaiser | 02 February 2018, 00:09:42 UTC | internal merge PiperOrigin-RevId: 184210759 | 02 February 2018, 00:24:33 UTC |
f9c859a | T2T Team | 01 February 2018, 18:50:25 UTC | Return scores from inference. PiperOrigin-RevId: 184162086 | 01 February 2018, 23:35:21 UTC |
413da43 | T2T Team | 01 February 2018, 13:39:48 UTC | Add metrics to ASR models,optionally remove waveforms from the data dict to prevent allocating space for them. PiperOrigin-RevId: 184126212 | 01 February 2018, 23:35:14 UTC |
c41864e | Lukasz Kaiser | 01 February 2018, 01:07:43 UTC | Make mask startup steps settable, run Adafactor by default. PiperOrigin-RevId: 184071297 | 01 February 2018, 23:35:03 UTC |
5855054 | T2T Team | 01 February 2018, 00:49:43 UTC | Fix block_v_size PiperOrigin-RevId: 184068805 | 01 February 2018, 23:34:57 UTC |
f1b68a9 | T2T Team | 31 January 2018, 22:49:54 UTC | Fix the embed function for hierarchical VQ-VAE and fix block_v_dim. PiperOrigin-RevId: 184051260 | 01 February 2018, 23:34:50 UTC |
6ce992f | Lukasz Kaiser | 31 January 2018, 22:48:48 UTC | Correct MNIST and add a basic 2-layer FC+ReLU model. PiperOrigin-RevId: 184051082 | 01 February 2018, 23:34:44 UTC |
8574057 | Noam Shazeer | 31 January 2018, 22:23:49 UTC | Partition the training data between the multiple hosts. This eliminates the model-quality bugs associated with reading the same data file in near-lockstep on two hosts. Remove the parallel_interleave hparam. Switch to per_host_input_for_training even for multi-tpu setup, since the bug preventing that has been resolved. PiperOrigin-RevId: 184046720 | 01 February 2018, 23:34:38 UTC |
6c324ea | Ryan Sepassi | 31 January 2018, 20:52:28 UTC | Add batch_size_means_tokens to IMDB PiperOrigin-RevId: 184032903 | 01 February 2018, 23:34:18 UTC |
8f32880 | T2T Team | 31 January 2018, 20:02:06 UTC | Write out decoded input file as well. PiperOrigin-RevId: 184025955 | 01 February 2018, 23:34:11 UTC |
5ed2092 | T2T Team | 31 January 2018, 19:39:35 UTC | Share relative position encodings across attention heads. PiperOrigin-RevId: 184022765 | 01 February 2018, 23:34:01 UTC |
51c6946 | T2T Team | 31 January 2018, 19:38:43 UTC | Fix visualization notebook. PiperOrigin-RevId: 184022642 | 01 February 2018, 23:33:54 UTC |
9b52264 | Ryan Sepassi | 31 January 2018, 18:38:15 UTC | Updated learning rate decay and warmup PiperOrigin-RevId: 184012344 | 01 February 2018, 23:33:46 UTC |
0562794 | T2T Team | 31 January 2018, 09:00:10 UTC | Fix typo `ctivation` => `activation` PiperOrigin-RevId: 183954098 | 01 February 2018, 23:33:38 UTC |
fa28c98 | T2T Team | 30 January 2018, 23:18:15 UTC | Remove a few optimizers, as they are already on by default in PiperOrigin-RevId: 183901344 | 01 February 2018, 23:33:28 UTC |
7f6ab2b | Ryan Sepassi | 30 January 2018, 19:14:50 UTC | resnet_50 achieving 75%+ accuracy on ImageNet PiperOrigin-RevId: 183861175 | 01 February 2018, 23:33:15 UTC |
e077e19 | Noam Shazeer | 30 January 2018, 17:39:24 UTC | Don't pass hparams.optimizer_adam_epsilon to AdaFactor optimizer - the PiperOrigin-RevId: 183844594 | 01 February 2018, 23:33:05 UTC |
c90b263 | Noam Shazeer | 30 January 2018, 10:40:04 UTC | Added config transformer_clean_big_tpu. Changed configs PiperOrigin-RevId: 183803333 | 01 February 2018, 23:32:52 UTC |
6eb2544 | Ryan Sepassi | 30 January 2018, 02:07:21 UTC | TF 1.5 requirement; faster input pipeline (esp good for image problems) PiperOrigin-RevId: 183762006 | 01 February 2018, 23:32:44 UTC |
0b13d20 | Ashish Vaswani | 30 January 2018, 00:50:08 UTC | Adding a gated linear unit layer and cifar10 generation problem. PiperOrigin-RevId: 183751275 | 01 February 2018, 23:32:37 UTC |
20b2f7c | Lukasz Kaiser | 29 January 2018, 23:55:51 UTC | Move keep_dims to keepdims as asked by TF, remove obsolete bluenet. PiperOrigin-RevId: 183742593 | 01 February 2018, 23:32:31 UTC |
6f54e43 | T2T Team | 29 January 2018, 23:53:20 UTC | remove the second config that overwrites the config created by PiperOrigin-RevId: 183742192 | 01 February 2018, 23:32:15 UTC |
7eb90b8 | Noam Shazeer | 29 January 2018, 23:43:00 UTC | Improvement to Adafactor - add anomaly resisatnce (like gradient PiperOrigin-RevId: 183740516 | 01 February 2018, 23:32:09 UTC |
70bed80 | Etienne Pot | 29 January 2018, 22:20:17 UTC | Update get_standardized_layers PiperOrigin-RevId: 183726473 | 01 February 2018, 23:32:00 UTC |
7f8e296 | T2T Team | 29 January 2018, 22:10:06 UTC | Small bug in my printing PiperOrigin-RevId: 183724624 | 01 February 2018, 23:31:50 UTC |
a682572 | T2T Team | 29 January 2018, 20:40:03 UTC | Fix API warning for deprecated arg. PiperOrigin-RevId: 183709272 | 01 February 2018, 23:31:44 UTC |
eb71537 | Niki Parmar | 29 January 2018, 20:02:04 UTC | Register Img2imgCeleba problem PiperOrigin-RevId: 183703634 | 01 February 2018, 23:31:38 UTC |
468a8f3 | Lukasz Kaiser | 29 January 2018, 19:45:57 UTC | Make shake-shake model work on CIFAR to some extent. PiperOrigin-RevId: 183700831 | 01 February 2018, 23:31:30 UTC |
a9092f2 | T2T Team | 29 January 2018, 19:10:20 UTC | Add option to use slicing as an alternative to projection. PiperOrigin-RevId: 183694546 | 01 February 2018, 23:31:13 UTC |