Skip to content

GitLab

Switch branch/tag
  1. 01 Jun, 2019 1 commit
  3. 31 May, 2019 1 commit
  5. 30 May, 2019 7 commits
  7. 29 May, 2019 7 commits
  9. 28 May, 2019 1 commit
  11. 24 May, 2019 2 commits
  13. 23 May, 2019 3 commits
    • Jason Fried's avatar
      collections.abc python 3.8 · 6b3a516f
      Jason Fried authored 
      Summary:
In python 3.7 collections.abc warns when importing abc classes from `collections`
In 3.8 this will not work at all.

This changes all code using abc's from collections to attempt to import from `collections.abc`

I am not fixing existing lint's don't ask, if `arc lint` auto-fixed I accepted, except for spelling in code.

Reviewed By: lisroach

Differential Revision: D15461049

fbshipit-source-id: ac2bf2ec8cffacd8ba5572882b0832bbf99a1646
      6b3a516f
    • Myle Ott's avatar
      Fix gating for find_unused_parameters · 128f4bea
      Myle Ott authored 
      Summary: Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/600

Differential Revision: D15469322

Pulled By: myleott

fbshipit-source-id: fdefa8efbb10e48b2a04a6bc10404fd2f3f21ecf
      128f4bea
    • Kritika Singh's avatar
      Allow unused params in distributed training · 72a5487c
      Kritika Singh authored 
      Summary:
Context from https://fb.workplace.com/groups/1405155842844877/permalink/2785095451517569/:

I am adding a model to pyspeech (formerly fairspeq) with the following `forward`:
```
def forward(self, src_tokens, src_lengths, prev_output_tokens, name):
    encoder_out = self.encoder(src_tokens, src_lengths)
    if name == Dataset.d1:
        decoder_out = self.decoder1(prev_output_tokens, encoder_out)
    elif name == Dataset.d2:
        decoder_out = self.decoder2(encoder_out)
    return decoder_out
```
When I run distributed training on this model, I get the following error:

```
RuntimeError: Expected to have finished reduction in the prior iteration before starting a
new one. This error indicates that your module has parameters that were not used in
producing loss. You can enable unused parameter detection by (1) passing the keyword
argument `find_unused_parameters=True` to `torch.nn.parallel.DistributedDataParallel`; (2)
making sure all `forward` function outputs participate in calculating loss. If you already have
done the above two steps, then the distributed data parallel module wasn't able to locate the
output tensors in the return value of your module's `forward` function. Please include the loss
function and the structure of the return value of `forward` of your module when reporting this
issue (e.g. list, dict, iterable). (prepare_for_backward at
caffe2/torch/csrc/distributed/c10d/reducer.cpp:410)
```

The recommended fix is to pass find_unused_parameters=True to DistributedDataParallel's initialization

Reviewed By: myleott

Differential Revision: D15439726

fbshipit-source-id: 7fd80d4a3f49ac90182dec723b49b14e6689406a
      72a5487c
  15. 22 May, 2019 2 commits
    • Matt Le's avatar
      Fix semisupervised translation · c11aaf14
      Matt Le authored 
      Summary: Fixes semisupervised translation task to deal with change in order of data loading and model creation (D15428242).  When we build the model, we create the backtranslation function, which we can then pass in to the constructor of BacktranslationDataset

Reviewed By: myleott

Differential Revision: D15455420

fbshipit-source-id: 95101ca92f8af33702be3416147edd98da135a20
      c11aaf14
    • zhiqiang's avatar
      Remove duplicate code (#754) · 886ef6bc
      zhiqiang authored 
      Summary:
Remove duplicate definition of PositionalEmbedding in `lightconv.py`
Pull Request resolved: https://github.com/pytorch/fairseq/pull/754

Differential Revision: D15451443

Pulled By: myleott

fbshipit-source-id: a3d82ab2c1335d66be3c5d67a07893162d138c7a
      886ef6bc
  17. 21 May, 2019 3 commits
  19. 20 May, 2019 4 commits
  21. 19 May, 2019 1 commit
  23. 17 May, 2019 2 commits
  25. 16 May, 2019 5 commits
  27. 15 May, 2019 1 commit
    • Ruty Rinott's avatar
      Fix biTransformer export (#583) · 2a3adcdc
      Ruty Rinott authored 
      Summary:
Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/583

D14610694 fixed issues in layerNorm exporting by making it conditional.  D15260838 changed the implementation of TransformerDecoderLayer to the one under transformer, thus losing the fix. Bringing it back here.

Reviewed By: myleott, geof90, liaimi

Differential Revision: D15357119

fbshipit-source-id: e29e053ca5beca0008d7a8dad9880a483a14c7b9
      2a3adcdc