Commits

Branch selector

 master

User selector

All users

All time

Commit History

Commits on Jan 25, 2018

• t2t-decoder with --checkpoint_path (#524)

  Show description for b6a57e7
  

  martinpopel

  authored and

  rsepassi

  committed
  b6a57e7

  Copy full SHA for b6a57e7

Commits on Jan 24, 2018

• Show t2t-bleu usage (#522)

  Show description for 02b903b
  

  martinpopel

  authored and

  rsepassi

  committed
  02b903b

  Copy full SHA for 02b903b

Commits on Jan 23, 2018

• make save_checkpoints_secs work again (#521)

  Show description for f0e638e
  

  martinpopel

  authored and

  rsepassi

  committed
  f0e638e

  Copy full SHA for f0e638e

• from six.moves import xrange (#533)

  Show description for afba9dc
  

  cclauss

  authored and

  rsepassi

  committed
  afba9dc

  Copy full SHA for afba9dc

Commits on Jan 13, 2018

• Merge pull request #506 from rsepassi/push

  Show description for a1d7ed7
  lukaszkaiser

  authored
  a1d7ed7

  Copy full SHA for a1d7ed7

• Internal merge

  Show description for c7f24da
  

  Ryan Sepassi

  committed
  c7f24da

  Copy full SHA for c7f24da

• fix when attention matrices have batch dimension also

  Show description for 164d175
  

  Niki Parmar

  authored and

  Ryan Sepassi

  committed
  164d175

  Copy full SHA for 164d175

Commits on Jan 12, 2018

• Internal merges

  Show description for 3d820df
  

  Ryan Sepassi

  committed
  3d820df

  Copy full SHA for 3d820df

• Add support for multivalent hparams, fix for device variable scope.

  Show description for e0fcdbc
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  e0fcdbc

  Copy full SHA for e0fcdbc

• New "Adafactor" optimizer - more memory efficient than Adam.

  Show description for 6bc1efb
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  6bc1efb

  Copy full SHA for 6bc1efb

• Fix behavior of max chars in ChoppedTextProblem

  Show description for 63135e0
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  63135e0

  Copy full SHA for 63135e0

• Add hparams.split_to_length option for chopping long fixed-length language modeling examples on data read. Make logic more explicit for whether to interpret batch_size as examples or as tokens. S…

  Show description for 20d7919
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  20d7919

  Copy full SHA for 20d7919

• Tweak transformer packed hparams on TPU

  Show description for 17ce194
  

  Ryan Sepassi

  committed
  17ce194

  Copy full SHA for 17ce194

• Allow preprocess_example to return multiple examples as a Dataset; fix example skipping

  Show description for f45909f
  

  Ryan Sepassi

  committed
  f45909f

  Copy full SHA for f45909f

• Rename _features_to_nonpadding to features_to_nonpadding.

  Show description for 0f35272
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  0f35272

  Copy full SHA for 0f35272

• Udpating the insights server to configure itself from simple JSON based

  Show description for 2b9e203
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  2b9e203

  Copy full SHA for 2b9e203

• Use tf.data API for file processing.

  Show description for b11b425
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  b11b425

  Copy full SHA for b11b425

• Updating the insights frontend to be configured via a JSON protocol

  Show description for 3f554d2
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  3f554d2

  Copy full SHA for 3f554d2

• Py3 test fix

  Show description for a5bb51a
  

  Ryan Sepassi

  committed
  a5bb51a

  Copy full SHA for a5bb51a

• Internal merges

  Show description for 0d2a059
  

  Ryan Sepassi

  committed
  0d2a059

  Copy full SHA for 0d2a059

• Some tweaks for TPU training

  Show description for 0beb37f
  

  Ryan Sepassi

  committed
  0beb37f

  Copy full SHA for 0beb37f

• Update model not found err msg

  Show description for b890479
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  b890479

  Copy full SHA for b890479

• Add an hparam to sample a random top-k rather than always the nearest neighbor.

  Show description for 5b2068e
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  5b2068e

  Copy full SHA for 5b2068e

• Working on super_lm. Add options for mixture of experts. Change max_length by default to 0 (defaults to max_length for the problem). Change optimizer_adam_beta2 to 0.999, which works better for …

  Show description for b95e3ab
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  b95e3ab

  Copy full SHA for b95e3ab

• Language modeling problem on cleaned-up version of wikipedia.

  Show description for 391413e
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  391413e

  Copy full SHA for 391413e

• hello_t2t with 1.5.0rc0

  Show description for cb3035f
  

  Ryan Sepassi

  committed
  cb3035f

  Copy full SHA for cb3035f

• Adding a short README with instructions on how to prepare and run the

  Show description for a64cbcd
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  a64cbcd

  Copy full SHA for a64cbcd

• Allow problems with a fixed length to override max_length. Restore previous functionality of using max_length=batch_size when hparams.max_length==0.

  Show description for 40d9ace
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  40d9ace

  Copy full SHA for 40d9ace

• The ps_devices are still necessary even when the no_data_parallelism hparam is set - this is for explicit model parallelism, which uses the ps_devices.

  Show description for 984b200
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  984b200

  Copy full SHA for 984b200

• No image summaries in while loops or in reused contexts - factor out this logic.

  Show description for f3a1f7d
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  f3a1f7d

  Copy full SHA for f3a1f7d

• Take gradient_norm out of default summaries. It over-spams the

  Show description for 004dec1
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  004dec1

  Copy full SHA for 004dec1

• Another bug where ema_count and ema_means was not being passed to bottleneck.

  Show description for 51f3d93
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  51f3d93

  Copy full SHA for 51f3d93

• Factor out fast_decode function from Transformer to allow re-use.

  Show description for f79f441
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  f79f441

  Copy full SHA for f79f441

• Fix ema_count() bug.

  Show description for 63ef9ee
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  63ef9ee

  Copy full SHA for 63ef9ee

• Create ChoppedTextProblem class for language modeling problems based on chopping up a corpus into fixed-sized segments. Refactor wikipedia language modeling problem to use this.

  Show description for 8fc0be2
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  8fc0be2

  Copy full SHA for 8fc0be2