Commits · 58e2c449cd81895748eac7b39309e9ff77557dc9 · OpenDAS / Fairseq

15 Jun, 2018 40 commits
- default dropout to correct value for big transformer · 58e2c449
  alexeib authored May 17, 2018
  
  58e2c449
- also report sentence/s timing when generating · bf47b956
  Alexei Baevski authored May 17, 2018
  
  bf47b956
- allow specifying max_tokens for generation · 67af40c9
  Alexei Baevski authored May 15, 2018
  
  67af40c9
- Support integer learning rates · a5e49364
  Myle Ott authored May 14, 2018
  
  a5e49364
- fix alignment when using uneven batches and left pad · 9f1b37dd
  Alexei Baevski authored May 11, 2018
  
  9f1b37dd
- implement batching in interactive mode · 663fd806
  Alexei Baevski authored May 11, 2018
  
  663fd806
- Sampling doesn't work with interactive · 4ce453b1
  Sergey Edunov authored May 10, 2018
  
  4ce453b1
- Fix --remove-bpe to strip trailing BPE symbols · a04c4cf4
  Myle Ott authored May 08, 2018
  
  a04c4cf4
- Remove padding from --score-reference · c53b2ee0
  Myle Ott authored May 08, 2018
  
  c53b2ee0
- fix flag copy paste (decoder-normalize-before) · 12a47a64
  alexeib authored May 08, 2018
  
  12a47a64
- add support for averaging last n checkpoints · f6a5a54e
  Alexei Baevski authored May 04, 2018
  
  f6a5a54e
- make attn dropout 0.1 default for big en-de transformer · 23211c45
  Alexei Baevski authored May 03, 2018
  
  23211c45
- fix to adding tokens to dictionary while thresholding · d85b61d6
  Angela Fan authored May 01, 2018
  
  d85b61d6
- Fix --prefix-size · 7f538f54
  Myle Ott authored Apr 30, 2018
  
  7f538f54
- make sure tensor used to index is cuda if on gpu · 2a681d99
  Alexei Baevski authored Apr 25, 2018
  
  2a681d99
- Remove src-padding from generation output · 88df72c0
  Myle Ott authored Apr 25, 2018
  
  88df72c0
- Fix tests · 8afb7761
  Myle Ott authored Apr 24, 2018
  
  8afb7761
- Support --warmup-updates with fixed LR schedule · 7c7634f6
  Myle Ott authored Apr 22, 2018
  
  7c7634f6
- Save and restore wall time in checkpoints · 0daba38e
  Myle Ott authored Apr 21, 2018
  
  0daba38e
- Simplify train.py (merge with singleprocess_train.py) · dc40ac58
  Myle Ott authored Apr 21, 2018
  
  dc40ac58
- Fix embedding initialization for padding · c6d4386c
  Alexei Baevski authored Apr 23, 2018
  
  c6d4386c
- Use eval() to parse args.lr · 1ec5f0a0
  Myle Ott authored Apr 17, 2018
  
  1ec5f0a0
- Fix preprocess.py · fa7c575a
  Myle Ott authored Apr 12, 2018
  
  fa7c575a
- Small optimization for LSTM · f607d9e8
  Myle Ott authored Apr 13, 2018
  
  f607d9e8
- Fix Flake8 · 8fcdb9b7
  Myle Ott authored Apr 12, 2018
  
  8fcdb9b7
- remove completed sentences from batch · 2a84f46b
  Alexei Baevski authored Apr 12, 2018
```
remove completed sentences from batch and allow batching uneven lengths (with fixes to make padded sequences work correctly in all models)
```
  2a84f46b
- No more magical --fp16 · bcdc27dc
  Myle Ott authored Apr 12, 2018
  
  bcdc27dc
- Pad dictionary to be a multiple of 8 in preprocessing · 745d5fbd
  Myle Ott authored Apr 12, 2018
  
  745d5fbd
- Revert "Make dictionary size a multiple of 8" · 4cd2bb70
  Myle Ott authored Apr 12, 2018
```
This reverts commit b2e119c209363e6ff6d2878a69c7d1a507a2e9be.
```
  4cd2bb70
- Make dictionary size a multiple of 8 · 26f87c7d
  Myle Ott authored Apr 11, 2018
  
  26f87c7d
- Add FP16 support · 7ee1d284
  Myle Ott authored Apr 10, 2018
  
  7ee1d284
- Fix batching during generation · 73a87327
  Myle Ott authored Apr 07, 2018
  
  73a87327
- Allow schedule for update-freq · 47b3b81c
  Myle Ott authored Apr 07, 2018
  
  47b3b81c
- Improve dataloader speed and deprecate concept of batch_offset (use... · 4fa8760e
  Myle Ott authored Apr 07, 2018
```
Improve dataloader speed and deprecate concept of batch_offset (use --sample-without-replacement instead)
```
  4fa8760e
- better batching · c52f6ea4
  Sergey Edunov authored Apr 07, 2018
  
  c52f6ea4
- Use FP32 for multi-head attention softmax · d6be0c7e
  Myle Ott authored Apr 07, 2018
  
  d6be0c7e
- Simulated big batches · 2d27ae08
  Sergey Edunov authored Apr 07, 2018
  
  2d27ae08
- More improvements to weight init and FP16 support · 60c4081b
  Myle Ott authored Apr 06, 2018
  
  60c4081b
- Use PyTorch LayerNorm and improve weight init · 36e360d9
  Myle Ott authored Apr 05, 2018
  
  36e360d9
- smarter way to avoid applying encoder key mask · fc830685
  alexeib authored Apr 05, 2018
  
  fc830685