Commits · ee36a6f3e310ff905f6d74b8d9eac4c3a220b188 · OpenDAS / Fairseq · GitLab

22 Jan, 2018 30 commits
- Fixed Weight Decay Regularization in Adam · ee36a6f3
  Michael Auli authored Jan 19, 2018
```
See https://arxiv.org/abs/1711.05101
```
  ee36a6f3
- Fix tests · 66d9fcf5
  Myle Ott authored Jan 22, 2018
  
  66d9fcf5
- Output correct perplexity when training with --sentence-avg · f9362e87
  Myle Ott authored Jan 19, 2018
  
  f9362e87
- Fix max_positions calculation in train.py · 81ace092
  Myle Ott authored Jan 19, 2018
  
  81ace092
- Better warning message for inputs that are too long · 33469436
  Myle Ott authored Jan 18, 2018
  
  33469436
- ATen Fix · 66314a60
  Michael Auli authored Jan 17, 2018
  
  66314a60
- Momentum correction · 173c577b
  Michael Auli authored Jan 16, 2018
  
  173c577b
- Report log likelihood for label smoothing · dd31fa92
  Sergey Edunov authored Jan 16, 2018
  
  dd31fa92
- Share input/output embed · c5378602
  Sergey Edunov authored Jan 16, 2018
  
  c5378602
- Better support for torch.no_grad (since volatile is deprecated) · 907ca927
  Myle Ott authored Jan 12, 2018
  
  907ca927
- Fix training · 0b84ab19
  Myle Ott authored Jan 09, 2018
  
  0b84ab19
- Save dictionary in model base classes · 5eddda8b
  Myle Ott authored Jan 09, 2018
  
  5eddda8b
- Fix gradient clipping when --clip-norm=0 · 08a74a32
  Myle Ott authored Jan 09, 2018
  
  08a74a32
- Fix LearnedPositionalEmbedding · fd28c880
  Myle Ott authored Jan 08, 2018
  
  fd28c880
- Move normalization of model output (e.g., via LSM) into model definition · 4db6579a
  Myle Ott authored Jan 07, 2018
  
  4db6579a
- Move positional embeddings into LearnedPositionalEmbedding module · c21a6e29
  Myle Ott authored Jan 06, 2018
  
  c21a6e29
- Fix warning about deprecated `volatile` kwarg for Variables · 185a0df5
  Myle Ott authored Jan 06, 2018
  
  185a0df5
- Add option to SequenceGenerator to retain dropout · dccf7909
  Myle Ott authored Jan 01, 2018
  
  dccf7909
- Add --max-sentences-valid to train.py · c542884d
  Myle Ott authored Jan 01, 2018
  
  c542884d
- Streamline data formatting utils · eb005cdb
  Myle Ott authored Jan 01, 2018
  
  eb005cdb
- Add reduce kwarg to criterions · 6f6cb4ab
  Myle Ott authored Jan 01, 2018
  
  6f6cb4ab
- Raise FileNotFoundError if dictionary files don't exist · dcbf5e75
  Myle Ott authored Dec 26, 2017
  
  dcbf5e75
- Output number of model parameters in train.py · fa508492
  Myle Ott authored Dec 26, 2017
  
  fa508492
- Add explicit dimension to softmax calls · 18a6d85c
  Myle Ott authored Dec 26, 2017
  
  18a6d85c
- Support deprecation of volatile Variables in latest PyTorch · 7da4e062
  Myle Ott authored Dec 22, 2017
  
  7da4e062
- Minor fix for strip_pad functions · 5637d54e
  Myle Ott authored Dec 22, 2017
  
  5637d54e
- Better error message for --decoder-attention · a4f86a89
  Myle Ott authored Dec 22, 2017
  
  a4f86a89
- Fix BeamableMM · a09fe803
  Myle Ott authored Dec 22, 2017
  
  a09fe803
- Add support for sharded generation · 9f7c3ec6
  Myle Ott authored Dec 21, 2017
  
  9f7c3ec6
- Fix generation bug with large beam sizes (>50) · cc7705d3
  Myle Ott authored Dec 07, 2017
  
  cc7705d3
05 Jan, 2018 1 commit
- Directly decay weight instead of L2 penalty (#157) · 9430544a
  Yann N. Dauphin authored Jan 05, 2018
```
See https://arxiv.org/pdf/1711.05101.pdf
```
  9430544a
06 Dec, 2017 9 commits
- Merge pull request #77 from facebookresearch/oss-merge-internal · 94dae690
  Myle Ott authored Dec 06, 2017
  
  94dae690
- Fix conv padding for even kernel widths · 0a836276
  Myle Ott authored Dec 06, 2017
  
  0a836276
- Rebuild optimizer when loading checkpoints · 10bf4074
  Myle Ott authored Dec 04, 2017
  
  10bf4074
- Fix weight norm dimension in decoder (fixes #73) · 9f3ccaa6
  Myle Ott authored Dec 04, 2017
  
  9f3ccaa6
- Save number of GPUs in args (and checkpoints) · 99493a85
  Myle Ott authored Dec 02, 2017
  
  99493a85
- Prefer command-line configuration over checkpoint for optimizer state · bd46c5ec
  Myle Ott authored Nov 21, 2017
  
  bd46c5ec
- Allow --lr to specify a fixed learning rate schedule · 19fafae6
  Myle Ott authored Nov 21, 2017
  
  19fafae6
- Improve memory handling (recover from OOM and periodically empty caching allocator) · a233fceb
  Myle Ott authored Nov 17, 2017
  
  a233fceb
- Improve error when resuming training with a different model architecture · be274623
  Myle Ott authored Nov 15, 2017
  
  be274623