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

Use 64bit ints to guard against issues with very large FSQ codebook size #171

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Use 64bit ints to guard against issues with very large FSQ codebook size #171

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
12 changes: 7 additions & 5 deletions stable_audio_tools/models/fsq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,10 +35,10 @@ def __init__(
super().__init__()
self.levels = levels

_levels = torch.tensor(levels, dtype=int32)
_levels = torch.tensor(levels, dtype=torch.int64)
self.register_buffer("_levels", _levels, persistent = False)

_basis = torch.cumprod(torch.tensor([1] + levels[:-1]), dim=0, dtype=int32)
_basis = torch.cumprod(torch.tensor([1] + levels[:-1]), dim=0, dtype=torch.int64)
self.register_buffer("_basis", _basis, persistent = False)

codebook_dim = len(levels)
Expand Down Expand Up @@ -90,7 +90,9 @@ def _indices_to_codes(self, indices):

def _codes_to_indices(self, zhat):
zhat = self._scale_and_shift(zhat)
return (zhat * self._basis).sum(dim=-1).to(int32)
zhat = zhat.round().to(torch.int64)
out = (zhat * self._basis).sum(dim=-1)
return out

def _indices_to_level_indices(self, indices):
indices = rearrange(indices, '... -> ... 1')
Expand All @@ -100,7 +102,7 @@ def _indices_to_level_indices(self, indices):
def indices_to_codes(self, indices):
# Expects input of batch x sequence x num_codebooks
assert indices.shape[-1] == self.num_codebooks, f'expected last dimension of {self.num_codebooks} but found last dimension of {indices.shape[-1]}'
codes = self._indices_to_codes(indices)
codes = self._indices_to_codes(indices.to(torch.int64))
codes = rearrange(codes, '... c d -> ... (c d)')
return codes

Expand All @@ -116,7 +118,7 @@ def forward(self, z, skip_tanh: bool = False):
# make sure allowed dtype before quantizing

if z.dtype not in self.allowed_dtypes:
z = z.float()
z = z.to(torch.float64)

codes = self.quantize(z, skip_tanh=skip_tanh)
indices = self._codes_to_indices(codes)
Expand Down