Skip to content

GitLab

Switch branch/tag
  1. 23 Jul, 2019 3 commits
  3. 22 Jul, 2019 8 commits
  5. 21 Jul, 2019 4 commits
    • Myle Ott's avatar
      Update GPT-2 BPE · 62b5498b
      Myle Ott authored 
      Summary: Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/749

Differential Revision: D16410984

Pulled By: myleott

fbshipit-source-id: 7698df46b8a179afccb287990f9705358690454a
      62b5498b
    • Myle Ott's avatar
      Default to mmap and infer dataset implementations automatically · 5f78106a
      Myle Ott authored 
      Summary: Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/751

Differential Revision: D16410989

Pulled By: myleott

fbshipit-source-id: ddbbee49756f9ff6c4487977a3f5d2259b7abafe
      5f78106a
    • Liang Wang's avatar
      Fix topp sampling issues (#882) · 1f96d284
      Liang Wang authored 
      Summary:
Two issues here:

1. `last_included` should be the last included index `cumsum_mask[:, :, -1:]` instead of `cumsum_mask[:, :, :1]`  (which is either 0 or 1);

2. If `--no-repeat-ngram-size` is set, the sum of `probs` may less than 1, we need to re-normalize to make it a valid probability distribution

The following code can reproduce this issues:

```
import torch
import numpy as np

def _sample_topp(probs):

    # =====  Code from  fairseq/search.py _sample_topp ======

    # sort the last dimension (vocab dimension) in descending order
    sorted_probs, sorted_indices = probs.sort(descending=True)

    # compute a mask to indicate the words to be included in the top-P set.
    cumsum_probs = sorted_probs.cumsum(dim=2)
    mask = cumsum_probs.lt(sampling_topp)

    # note that mask was computed by 'lt'. One more word needs to be included
    # so that the cumulative probability mass can exceed p.
    cumsum_mask = mask.cumsum(dim=2)
    last_included = cumsum_mask[:, :, :1]
    mask = mask.scatter_(2, last_included, 1)

    # truncate unnecessary dims.
    max_dim = last_included.max()
    truncated_mask = mask[:, :, :max_dim + 1]
    truncated_probs = sorted_probs[:, :, :max_dim + 1]
    truncated_indices = sorted_indices[:, :, :max_dim + 1]

    # trim the words that are not in top-P by setting their probabilities
    # to 0, so that they would not be sampled later.
    trim_mask = 1 - truncated_mask
    trimed_probs = truncated_probs.masked_fill_(trim_mask, 0)
    return trimed_probs, truncated_indices

    # ========================================================

if __name__ == '__main__':
    np.random.seed(1234)
    torch.manual_seed(1234)

    sampling_topp = 0.9
    probs = torch.softmax(torch.randn(1, 1, 10), dim=-1)
    # probs = tensor([0.0545, 0.0779, 0.0189, 0.0647, 0.0282, 0.0862, 0.0656, 0.1041, 0.0399, 0.4600])
    print('probs =', probs[0][0])

    trimed_probs, truncated_indices = _sample_topp(probs)

    cum_probs = trimed_probs.cumsum(dim=-1)[0][0]
    # cumsum = tensor([0.4600, 0.5641])
    print('cumsum =', cum_probs)
    # Will throw AssertionError
    assert float(cum_probs[-1]) >= sampling_topp

```
Pull Request resolved: https://github.com/pytorch/fairseq/pull/882

Differential Revision: D16409269

Pulled By: xingz9

fbshipit-source-id: 94b1122eed50c656057b64e22af6f4a6ea7a68af
      1f96d284
    • Myle Ott's avatar
      Rename data.transforms -> data.encoders · f812e529
      Myle Ott authored 
      Summary: Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/747

Differential Revision: D16403464

Pulled By: myleott

fbshipit-source-id: ee3b4184f129a02be833c7bdc00685978b4de883
      f812e529
  7. 19 Jul, 2019 7 commits
  9. 17 Jul, 2019 5 commits
  11. 14 Jul, 2019 1 commit
  13. 11 Jul, 2019 2 commits
  15. 10 Jul, 2019 1 commit
  17. 09 Jul, 2019 1 commit
  19. 08 Jul, 2019 1 commit
    • Guanheng Zhang's avatar
      Integrate torch.nn and fairseq MultiheadAttention (#772) · 6d2e0831
      Guanheng Zhang authored 
      Summary:
Integrate torch.nn and fairseq MultiheadAttention modules. In the future, both libraries will be benefited from performance optimization together.

Under the following circumstances, the calculation of the MultiheadAttention will still remain in fairseq, including:
1. onnx trace
2. incremental state
3. static kv

We plan to gradually mitigate those capabilities to PyTorch's core library.

Faieseq users can user the attribute self.enable_torch_version to force the calculations in either torch or fairseq. We use the following script to ensure both versions yield the same results.

------------------------------------------------------------------------------------
```
import torch
from fairseq.modules import MultiheadAttention
import time

embed_dim = 64
kv_embed_dim = 1208
num_heads = 16
src_len = 20
tgt_len = 30
bsz = 10

model = MultiheadAttention(embed_dim, num_heads, kdim=kv_embed_dim, vdim=kv_embed_dim,
                           bias=True, add_bias_kv=True, add_zero_attn=True)

query = torch.rand((src_len, bsz, embed_dim))
key = torch.rand((src_len, bsz, kv_embed_dim))
value = torch.rand((src_len, bsz, kv_embed_dim))

attn_mask = torch.randint(0, 2, (src_len, src_len)).float()
attn_mask.masked_fill_(attn_mask == 0, float('-inf'))
attn_mask.masked_fill_(attn_mask > 0, float('0.0'))

seq_mask = torch.randint(0, 2, (1, src_len))
key_padding_mask = seq_mask
for i in range(bsz-1):
    key_padding_mask = torch.cat([key_padding_mask, seq_mask], axis=0)
key_padding_mask = key_padding_mask == 1

# Apply torch.nn version
model.enable_torch_version = True
torch_output, torch_weight = model(query, key, value, key_padding_mask=key_padding_mask, attn_mask=attn_mask)

# Apply fairseq version
model.enable_torch_version = False
fairseq_output, fairseq_weight = model(query, key, value, key_padding_mask=key_padding_mask, attn_mask=attn_mask)

print("torch and fairseq generate same results: outputs are same ? ",
      torch.allclose(torch_output, fairseq_output, atol=5e-6, rtol=1e-6),
      ", weights are same ? ",
      torch.allclose(torch_weight, fairseq_weight, atol=5e-6, rtol=1e-6)
)
```
------------------------------------------------------------------------------------
Expected results:
torch and fairseq generate same results: outputs are same ?  True , weights are same ?  True

------------------------------------------------------------------------------------
Similar performance is expected for both two versions. Using the following setup and have the initial performance benchmark results:

#########################
embed_dim = 32
kv_embed_dim = 32
num_heads = 4
src_len = 3
tgt_len = 2
bsz = 4
num_samples = 50000

#########################
torch-version MultiheadAttention cpu time: 0.46589  ms per iteration.
fairseq-version MultiheadAttention cpu time: 0.47861  ms per iteration.
torch-version MultiheadAttention gpu time: 0.82330  ms per iteration.
fairseq-version MultiheadAttention gpu time: 0.79410  ms per iteration.
Pull Request resolved: https://github.com/pytorch/fairseq/pull/772

Reviewed By: myleott

Differential Revision: D16108450

Pulled By: zhangguanheng66

fbshipit-source-id: cd2eb5a6eeeab6c274999b7928c2af14fc211565
      6d2e0831
  21. 06 Jul, 2019 2 commits
  23. 04 Jul, 2019 1 commit
    • Spencer Poff's avatar
      support streaming iterator · 5c241c8c
      Spencer Poff authored 
      Summary:
For tasks that involve streaming data directly from an API, we need a simpler epoch iterator.

Also included in this change is support for initializing a dictionary with an arbitrary list of special symbols.

Reviewed By: myleott

Differential Revision: D16110603

fbshipit-source-id: be6d9f680292dec1512614871f9269c95ac84861
      5c241c8c
  25. 02 Jul, 2019 1 commit
    • Xutai Ma's avatar
      add --max-tokens-valid option for validation · bccfddbb
      Xutai Ma authored 
      Summary: Add the max-token-valid option. Sometime a separate max batch tokens for validation may be helpful, for example when there is a long sequence in validation set thats larger than max_tokens (it's rare in MT but could happen in ASR or AST).

Reviewed By: myleott

Differential Revision: D16076951

fbshipit-source-id: ae7f4218594580b9450a8196d7afa1e7e2018aee
      bccfddbb
  27. 01 Jul, 2019 3 commits