Commits · 753935efe83513298cb7d5e2bf1995eda7b48817 · OpenDAS / Fairseq

03 Sep, 2018 28 commits
- Merge internal changes · 753935ef
  Myle Ott authored Aug 27, 2018
  
  753935ef
- word stats in eval_lm · c7c567a7
  Alexei Baevski authored Aug 26, 2018
  
  c7c567a7
- Old checkpoints can't be loaded because of a new meter · c9b800d2
  Sergey Edunov authored Aug 24, 2018
  
  c9b800d2
- Add training wall time meter · 9c102784
  Myle Ott authored Aug 24, 2018
  
  9c102784
- disable final layer norm for transformer decoder as it makes things worse · f84e1ed4
  Alexei Baevski authored Aug 23, 2018
  
  f84e1ed4
- Fix adaptive softmax cutoff comment · 81ba4c4c
  Louis Martin authored Aug 23, 2018
  
  81ba4c4c
- Remove --normalization-constant from fconv · 47f095fd
  Myle Ott authored Aug 21, 2018
  
  47f095fd
- Diverse Beam Search · 8c0ca1a0
  Myle Ott authored Aug 10, 2018
  
  8c0ca1a0
- add warmup support back to cosine lr sched (important for mt) · ba9f32cc
  alexeib authored Aug 16, 2018
  
  ba9f32cc
- Warn when using FP16 on pre-Volta GPUs · 8d6665f2
  Myle Ott authored Aug 14, 2018
  
  8d6665f2
- Move read_binarized.py to scripts/ · e7422192
  Myle Ott authored Aug 14, 2018
  
  e7422192
- script to read binarized data · cd6590b6
  Alexei Baevski authored Aug 14, 2018
  
  cd6590b6
- Increase max buffer size in all_gather_list · 5b533aa6
  Myle Ott authored Aug 13, 2018
  
  5b533aa6
- fix tests · f1d81db8
  alexeib authored Aug 09, 2018
  
  f1d81db8
- Reset gnorm after each epoch · 97a6b139
  Sergey Edunov authored Aug 09, 2018
  
  97a6b139
- Factor out search logic in SequenceGenerator · ef43da72
  Myle Ott authored Aug 09, 2018
  
  ef43da72
- cosine + triangular lr scheduler · 75e12a27
  Alexei Baevski authored Aug 08, 2018
  
  75e12a27
- parameters to separate input/inner/out dims · 1d38624f
  alexeib authored Aug 04, 2018
  
  1d38624f
- load args from model for eval_lm · e4f51e18
  alexeib authored Aug 03, 2018
  
  e4f51e18
- make batching faster for monolingual dataset · 45082e48
  alexeib authored Aug 02, 2018
  
  45082e48
- fix tests · 0b5166db
  alexeib authored Jul 31, 2018
  
  0b5166db
- add flag that allows keeping optimizer config · 2dc074d8
  alexeib authored Jul 28, 2018
```
adds -reset-optimizer, --reset-lr-scheduler, and --optimizer-overrides flags
```
  2dc074d8
- make adaptive softmax dropout an optional arg · 6e3685ad
  Alexei Baevski authored Jul 28, 2018
  
  6e3685ad
- Always smaller soft · 19c25f47
  Alexei Baevski authored Jul 28, 2018
```
no need to have half-size option as behavior can be reproduced with existing flags
```
  19c25f47
- remove unneeded defaults · d8998173
  alexeib authored Jul 28, 2018
  
  d8998173
- character token embeddings for word level predictions · 885e7ec9
  Alexei Baevski authored Jul 28, 2018
  
  885e7ec9
- option for a smaller adaptive softmax · 616afddd
  Alexei Baevski authored Jul 28, 2018
  
  616afddd
- fix adaptive softmax indexing · f69206c8
  Alexei Baevski authored Jul 26, 2018
  
  f69206c8
16 Aug, 2018 3 commits
- Fix bidirectional LSTM concatenation (#249) · af38ed48
  Myle Ott authored Aug 16, 2018
  
  af38ed48
- Fix comment · 53c7d271
  Myle Ott authored Aug 16, 2018
  
  53c7d271
- add end-of-stack normalizations in case normalize_before has been set (#244) · fedc55ec
  ngimel authored Aug 16, 2018
  
  fedc55ec
01 Aug, 2018 2 commits
- Add ensemble for different architectures (#235) · f7f2dd01
  Maggie Li authored Aug 01, 2018
  
  f7f2dd01
- Fix bug when training with FP32 and --update-freq (#236) · 202e0bbe
  Myle Ott authored Jul 31, 2018
  
  202e0bbe
31 Jul, 2018 1 commit
- Correct the help name of the prefixes arguments (#234) · 9143dfab
  alvations authored Jul 31, 2018
  
  9143dfab
27 Jul, 2018 2 commits
- Correct path in the pre-processing example (#230) · cabcc254
  alvations authored Jul 27, 2018
  
  cabcc254
- Add load_optim option to load checkpoint but not optimizer state (#229) · 79bbe1d8
  theweiho authored Jul 26, 2018
  
  79bbe1d8
25 Jul, 2018 4 commits

fbshipit-source-id: 3f76eab2b42792fc8ed087fa0e2f4968bf980ad7 · 5d99e139
myleott authored Jul 25, 2018

5d99e139

Myle Ott authored Jul 25, 2018

Changelog:
- `f472d141`: Support tied embeddings in LSTM encoder/decoder
- `89e19d42`: Don't print alignment by default (use `--print-alignment` to re-enable it)
- `d2e2a1d4`: Add Transformer-based language model
- `c2794070`: Add new Transformer configuration for IWSLT
- `2fbfda0d`: Misc changes for pytorch-translate
- Miscellaneous bug fixes

fe4e185a

Merge internal changes · 2fbfda0d
Myle Ott authored Jul 25, 2018

2fbfda0d
Fix comment · 93fec886
Myle Ott authored Jul 25, 2018

93fec886