Commits

Breadcrumbs

History for

onmaster

User selector

All users

All time

Commit History

Commits on Dec 4, 2017

• add translate_enfr_wmt32k_packed problem.

  Show description for b8aa4ea
  

  nshazeer

  authored and

  lukaszkaiser

  committed
  b8aa4ea

  Copy full SHA for b8aa4ea

Commits on Dec 1, 2017

• T2T depends on TF 1.4+, daisy_chain_getter bug fix, some Eager-mode improvements/fixes

  Show description for aa2c0b7
  

  Ryan Sepassi

  committed
  aa2c0b7

  Copy full SHA for aa2c0b7

• Remove TranslateEndeWmtBpe32kPacked. We have TranslateEndeWmt32kPacked.

  Show description for 01030eb
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  01030eb

  Copy full SHA for 01030eb

• Packed datasets - combine examples to constant length for efficient TPU training. Modify transformer to keep the packed-together examples from attending to one another.

  Show description for c9144df
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  c9144df

  Copy full SHA for c9144df

Commits on Nov 29, 2017

• fix bug with bpe32k - prepend <pad> and <EOS> to vocab.

  Show description for 7909c69
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  7909c69

  Copy full SHA for 7909c69

• Move Estimator input_fn and model_fn construction into Problem and T2TModel, respectively, which allows subclassing

  Show description for 398e85b
  

  Ryan Sepassi

  committed
  398e85b

  Copy full SHA for 398e85b

• merge PRs.

  Show description for 6d9b5e1
  

  Lukasz Kaiser

  authored and

  Ryan Sepassi

  committed
  6d9b5e1

  Copy full SHA for 6d9b5e1

• Versions of problems with combined examples. Good for TPU training.

  Show description for 936db05
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  936db05

  Copy full SHA for 936db05

• Bug fix, make class lab a VarLen feature to be compatible with img2img datasets that don't have labels.

  Show description for dccc9ac
  

  Niki Parmar

  authored and

  Ryan Sepassi

  committed
  dccc9ac

  Copy full SHA for dccc9ac

• Added an option to omit quotes from the subtoken voc format.

  Show description for 2e6d512
  

  T2T Team

  authored and

  Ryan Sepassi

  committed
  2e6d512

  Copy full SHA for 2e6d512

• Add ImageImagenet224; resnet50 now fits batch size of 128

  Show description for 717afe9
  

  Ryan Sepassi

  committed
  717afe9

  Copy full SHA for 717afe9

• Add Modality.targets_weights_fn

  Show description for 50f5515
  

  Ryan Sepassi

  committed
  50f5515

  Copy full SHA for 50f5515

• More tiny sketch fixes

  Show description for 6cf47f9
  

  Ryan Sepassi

  committed
  6cf47f9

  Copy full SHA for 6cf47f9

Commits on Nov 23, 2017

• Merge pull request #419 from wingsbr/librispeech

  Show description for 92983ea
  lukaszkaiser

  authored
  92983ea

  Copy full SHA for 92983ea

Commits on Nov 21, 2017

• Variety of fixes based on PR comments.

  wingsbr

  committed
  23129f2

  Copy full SHA for 23129f2

• Switched to full librispeech datasets.

  wingsbr

  committed
  98c7b41

  Copy full SHA for 98c7b41

Commits on Nov 20, 2017

• Added librispeech to data_generators/all_problems.py

  wingsbr

  committed
  844df4d

  Copy full SHA for 844df4d

• Expanded to include librispeech Problem and Modality.

  wingsbr

  committed
  5365113

  Copy full SHA for 5365113

Commits on Nov 14, 2017

• Raw text data generation for train, dev and test; script to compute rouge using the official pyrouge dist; a bash script to tokenize targets and predictions before computing rouge

  urvashik

  committed
  3f52fb9

  Copy full SHA for 3f52fb9

• Generating raw data files, completed pipeline for rouge

  urvashik

  committed
  a82231e

  Copy full SHA for a82231e

• Added a librispeech data generator.

  wingsbr

  committed
  0c026d2

  Copy full SHA for 0c026d2

Commits on Nov 11, 2017

• Rm ImageModality and ClassLabelModality 2d; rename SmallImage to Image

  Show description for eb5652f
  

  Ryan Sepassi

  committed
  eb5652f

  Copy full SHA for eb5652f

• Enable separate eval and t2t_usr_dir for TPU

  Show description for 955dad5
  

  Ryan Sepassi

  committed
  955dad5

  Copy full SHA for 955dad5

• Add CTC loss SymbolModality, correct get_bleu script, LSTM without inputs, small en-fr generator, play with adversarial transformer.

  Show description for e1bde97
  

  Lukasz Kaiser

  authored and

  Ryan Sepassi

  committed
  e1bde97

  Copy full SHA for e1bde97

• Fixing an issue in masked 2d masked local attention where the corner dim was being shifted to the front. Now, each block has a right shift. Adding an imagenet 64^2 problem with AREA resizing.

  Show description for 11a8d33
  

  Ashish Vaswani

  authored and

  Ryan Sepassi

  committed
  11a8d33

  Copy full SHA for 11a8d33

• Update RangedHParams to accept bools

  Show description for d007d47
  

  Ryan Sepassi

  committed
  d007d47

  Copy full SHA for d007d47

• Enable Xception and ImageNet on TPU

  Show description for ca22374
  

  Ryan Sepassi

  committed
  ca22374

  Copy full SHA for ca22374

Commits on Nov 3, 2017

• Work on generators: improve EnCs, add large EnFr and OCR test; LSTM corrections.

  Show description for c022afd
  

  Lukasz Kaiser

  authored and

  Ryan Sepassi

  committed
  c022afd

  Copy full SHA for c022afd

• Allow LSTMs to run on problems without inputs.

  Show description for 9a2a6f3
  

  Lukasz Kaiser

  authored and

  Ryan Sepassi

  committed
  9a2a6f3

  Copy full SHA for 9a2a6f3

• Ignore sync flag when in single-worker mode.

  Show description for f2e8e35
  

  nshazeer

  authored and

  Ryan Sepassi

  committed
  f2e8e35

  Copy full SHA for f2e8e35

Commits on Nov 2, 2017

• Merge pull request #392 from vince62s/translate

  Show description for 172a1b1
  lukaszkaiser

  authored
  172a1b1

  Copy full SHA for 172a1b1

Commits on Oct 30, 2017

• typo fix

  vince62s

  committed
  733de7b

  Copy full SHA for 733de7b

• Fix the EnZh task

  vince62s

  committed
  6b1267e

  Copy full SHA for 6b1267e

Commits on Oct 29, 2017

• fix mispells

  Mistobaan

  committed
  91d4e1c

  Copy full SHA for 91d4e1c

Commits on Oct 27, 2017

• Add back commented-out enfr datasets

  Show description for 9d86cf7
  

  Ryan Sepassi

  committed
  9d86cf7

  Copy full SHA for 9d86cf7