Skip to content

Fix attentive statistics pooling dtype mismatch with FP16 and BF16 #3081

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

Open
Con-Benksl wants to merge 1 commit into
base: develop
Choose a base branch
from
+40 −1
Open

Fix attentive statistics pooling dtype mismatch with FP16 and BF16 #3081

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
2 changes: 1 addition & 1 deletion speechbrain/lobes/models/ECAPA_TDNN.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -285,7 +285,7 @@ def _compute_statistics(x, m, dim=2, eps=self.eps):
mean, std = _compute_statistics(x, mask / total)
mean = mean.unsqueeze(2).repeat(1, 1, L)
std = std.unsqueeze(2).repeat(1, 1, L)
attn = torch.cat([x, mean, std], dim=1)
attn = torch.cat([x, mean, std], dim=1).to(x.dtype)
else:
attn = x

Expand Down
39 changes: 39 additions & 0 deletions tests/unittests/test_ECAPA_TDNN.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,39 @@
import pytest
import torch


@pytest.mark.parametrize(
"dtype", [torch.float16, torch.bfloat16, torch.float32]
)
@pytest.mark.parametrize("global_context", [True, False])
@pytest.mark.parametrize("use_lengths", [True, False])
def test_attentive_statistics_pooling_dtype(
device, dtype, global_context, use_lengths
):
from speechbrain.lobes.models.ECAPA_TDNN import AttentiveStatisticsPooling

pool = AttentiveStatisticsPooling(
channels=2, attention_channels=2, global_context=global_context
).to(device=device, dtype=dtype)
# Uniform attention gives the mean and standard deviation of valid frames.
with torch.no_grad():
pool.conv.conv.weight.zero_()
pool.conv.conv.bias.zero_()
x = torch.tensor(
[[[1, 3, 1, 3], [2, 4, 2, 4]], [[1, 3, 5, 7], [2, 4, 6, 8]]],
device=device,
dtype=dtype,
requires_grad=True,
)
lengths = torch.tensor([1.0, 0.5], device=device) if use_lengths else None

output = pool(x, lengths)

expected = [[2, 3, 1, 1], [2, 3, 1, 1]]
if not use_lengths:
expected[1] = [4, 5, 2.2360679775, 2.2360679775]
expected = torch.tensor(expected, device=device, dtype=dtype).unsqueeze(2)
torch.testing.assert_close(output, expected)
output.sum().backward()
assert torch.isfinite(x.grad).all()
assert all(torch.isfinite(param.grad).all() for param in pool.parameters())