Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[mask] set max_chunk_size according to subsample rate #2520

Merged
merged 1 commit into from
May 8, 2024
Merged

[mask] set max_chunk_size according to subsample rate #2520

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 11 additions & 6 deletions wenet/transformer/encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -157,12 +157,17 @@ def forward(
xs = self.global_cmvn(xs)
xs, pos_emb, masks = self.embed(xs, masks)
mask_pad = masks # (B, 1, T/subsample_rate)
chunk_masks = add_optional_chunk_mask(xs, masks,
self.use_dynamic_chunk,
self.use_dynamic_left_chunk,
decoding_chunk_size,
self.static_chunk_size,
num_decoding_left_chunks)
chunk_masks = add_optional_chunk_mask(
xs,
masks,
self.use_dynamic_chunk,
self.use_dynamic_left_chunk,
decoding_chunk_size,
self.static_chunk_size,
num_decoding_left_chunks,
# Since we allow up to 1s(100 frames) delay, the maximum
# chunk_size is 100 / 4 = 25.
max_chunk_size=int(100.0 / self.embed.subsampling_rate))
if self.use_sdpa:
chunk_masks = mask_to_bias(chunk_masks, xs.dtype)
if self.gradient_checkpointing and self.training:
Expand Down
11 changes: 6 additions & 5 deletions wenet/utils/mask.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,7 +130,8 @@ def add_optional_chunk_mask(xs: torch.Tensor,
decoding_chunk_size: int,
static_chunk_size: int,
num_decoding_left_chunks: int,
enable_full_context: bool = True):
enable_full_context: bool = True,
max_chunk_size: int = 25):
""" Apply optional mask for encoder.

Args:
Expand All @@ -151,8 +152,8 @@ def add_optional_chunk_mask(xs: torch.Tensor,
>=0: use num_decoding_left_chunks
<0: use all left chunks
enable_full_context (bool):
True: chunk size is either [1, 25] or full context(max_len)
False: chunk size ~ U[1, 25]
True: chunk size is either [1, max_chunk_size] or full context(max_len)
False: chunk size ~ U[1, max_chunk_size]

Returns:
torch.Tensor: chunk mask of the input xs.
Expand All @@ -167,15 +168,15 @@ def add_optional_chunk_mask(xs: torch.Tensor,
chunk_size = decoding_chunk_size
num_left_chunks = num_decoding_left_chunks
else:
# chunk size is either [1, 25] or full context(max_len).
# chunk size is either [1, max_chunk_size] or full context(max_len).
# Since we use 4 times subsampling and allow up to 1s(100 frames)
# delay, the maximum frame is 100 / 4 = 25.
chunk_size = torch.randint(1, max_len, (1, )).item()
num_left_chunks = -1
if chunk_size > max_len // 2 and enable_full_context:
chunk_size = max_len
else:
chunk_size = chunk_size % 25 + 1
chunk_size = chunk_size % max_chunk_size + 1
if use_dynamic_left_chunk:
max_left_chunks = (max_len - 1) // chunk_size
num_left_chunks = torch.randint(0, max_left_chunks,
Expand Down
Loading