Commits · 0ef2856c3fcbbf732f8f7213d19ec3616139b36b · OpenDAS / Fairseq · GitLab

25 Jul, 2018 11 commits
- Don't compute unnecessary attention averages during training · 0ef2856c
  Myle Ott authored Jul 12, 2018
  
  0ef2856c
- Output positional scores in interactive.py · c37fc8fd
  Myle Ott authored Jul 12, 2018
  
  c37fc8fd
- Iterate on need_attn and fix tests · bb5f15d1
  Myle Ott authored Jul 12, 2018
  
  bb5f15d1
- Fix bug when --share-all-embeddings but no --encoder-embed-path · 498a186d
  Stephen Roller authored Jul 11, 2018
  
  498a186d
- default need_attn to False · 1018c333
  alexeib authored Jul 10, 2018
  
  1018c333
- disable printing alignment by default (for perf) and add a flag to enable it · 89e19d42
  Alexei Baevski authored Jul 10, 2018
  
  89e19d42
- Support tied embeddings in LSTM encoder/decoder · f472d141
  Stephen Roller authored Jul 03, 2018
  
  f472d141
- fix token block rotation · a7d0bd0e
  alexeib authored Jul 02, 2018
  
  a7d0bd0e
- make model access saner · 0e9e7f7b
  alexeib authored Jul 02, 2018
  
  0e9e7f7b
- fix raw text for language modeling · d494485f
  Alexei Baevski authored Jun 27, 2018
  
  d494485f
- fixed output_proj's input_dim in attention (#226) · 7358296b
  higgsfield authored Jul 25, 2018
  
  7358296b
20 Jul, 2018 1 commit
- stories data preprocessing needs padding factor 1 to match pretrained model, updating readme · 28adb200
  Angela Fan authored Jul 19, 2018
  
  28adb200
19 Jul, 2018 1 commit
- Pass sampling-temperature trough to the generator in interactive.py · eaa576b0
  Sergey Edunov authored Jul 19, 2018
  
  eaa576b0
11 Jul, 2018 1 commit
- Fix up model defaults (#211) · ac5fddfc
  Mehdi Drissi authored Jul 11, 2018
  
  ac5fddfc
10 Jul, 2018 1 commit
- assert that vocab size >= adaptive softmax cutoff (#214) · f26b6aff
  Alexei Baevski authored Jul 10, 2018
  
  f26b6aff
08 Jul, 2018 2 commits
- adding model arg override at generation time for interactive.py · 1d79ed9b
  Angela Fan authored Jul 08, 2018
  
  1d79ed9b
- add model override argument from load_ensemble_for_inference at generation... · 30ef667d
  Angela Fan authored Jul 07, 2018
```
add model override argument from load_ensemble_for_inference at generation time, updating readme for stories
```
  30ef667d
02 Jul, 2018 2 commits
- fix decoder_normalize_before typo (#205) · ff3db3cd
  ngimel authored Jul 02, 2018
  
  ff3db3cd
- adding pretrained stories model · 62578bbe
  Angela Fan authored Jul 02, 2018
  
  62578bbe
28 Jun, 2018 1 commit
- Fix preprocessing for WMT14 En-De to replicate Scaling NMT paper (#203) · a75c3092
  Myle Ott authored Jun 28, 2018
  
  a75c3092
26 Jun, 2018 1 commit
- Remove unnecessary assert (fixes #199) (#200) · 1f08602f
  Myle Ott authored Jun 26, 2018
  
  1f08602f
25 Jun, 2018 2 commits
- Remove more Variable() calls (#198) · 6edf81dd
  Myle Ott authored Jun 25, 2018
  
  6edf81dd
- Fix attention order in unit tests (fixes #195) (#197) · 74efc214
  Myle Ott authored Jun 25, 2018
  
  74efc214
24 Jun, 2018 3 commits
- Misc changes for pytorch-translate · bd4db8fb
  Myle Ott authored Jun 24, 2018
  
  bd4db8fb
- Fix for Dictionary.finalize · c6fe9fc5
  Myle Ott authored Jun 24, 2018
  
  c6fe9fc5
- Fix typo · 7bcb487a
  Myle Ott authored Jun 22, 2018
  
  7bcb487a
21 Jun, 2018 14 commits
- Merge pull request #193 from hmc-cs-mdrissi/minor_fixes · 3feaccf3
  Alexei Baevski authored Jun 21, 2018
```
default samples_per_token in eval_lm
```
  3feaccf3
- Merge branch 'master' of https://github.com/pytorch/fairseq into minor_fixes · 891f2307
  Mehdi Drissi authored Jun 21, 2018
  
  891f2307
- Add steps to reproduce WMT En-De results from Scaling NMT paper · b458977a
  Myle Ott authored Jun 21, 2018
  
  b458977a
- Move reorder_encoder_out to FairseqEncoder and fix non-incremental decoding · 6ec5022e
  Myle Ott authored Jun 21, 2018
  
  6ec5022e
- Fix interpretation of --max-epoch · e9967cd3
  Myle Ott authored Jun 21, 2018
  
  e9967cd3
- add sinusoidal pos initialization · 678d0bbc
  alexeib authored Jun 20, 2018
  
  678d0bbc
- fix sinusoidal embedding init size · 03fffb98
  Alexei Baevski authored Jun 20, 2018
  
  03fffb98
- Support FP16 during inference · 930c9580
  Myle Ott authored Jun 19, 2018
  
  930c9580
- Support pretrained embeddings for Transformer. · 9a88b71d
  Stephen Roller authored Jun 19, 2018
```
Also show a nicer error message.
```
  9a88b71d
- sort descending when evaluating lm because it is faster (17k wps vs 11k) and will fail early if oom · 5bf07724
  alexeib authored Jun 19, 2018
  
  5bf07724
- respect max tokens and ignore invalid inputs when evaluating lm · a091d239
  alexeib authored Jun 19, 2018
  
  a091d239
- Fix a bug with using GloVe 840B tokens for initialization. · cc85d411
  Stephen Roller authored Jun 19, 2018
  
  cc85d411
- Store full checkpoints instead of symlinking · 9dcee4c7
  Myle Ott authored Jun 18, 2018
  
  9dcee4c7
- Better failure message when loss explodes during FP16 training · d9a13180
  Myle Ott authored Jun 18, 2018
  
  d9a13180