Skip to content

[Feature] Improve VISReg loss numerical stability and defaults #431

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

[Feature] Improve VISReg loss numerical stability and defaults #431

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
10 changes: 5 additions & 5 deletions stable_pretraining/methods/visreg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,13 +85,13 @@ def __init__(
self._cached_B = -1
self._cached_target = None

def _get_target(self, B: int, device, dtype) -> torch.Tensor:
def _get_target(self, B: int, device) -> torch.Tensor:
"""Theoretical standard-normal quantiles for ``B`` sorted samples."""
if self._cached_B != B:
q = torch.linspace(1, B, B, device=device, dtype=torch.float32) / (B + 1)
self._cached_target = torch.erfinv(2 * q - 1).mul_(math.sqrt(2))
self._cached_B = B
return self._cached_target.to(device=device, dtype=dtype)
return self._cached_target.to(device=device)

def forward(self, z: torch.Tensor) -> torch.Tensor:
""":param z: Embeddings [V, B, D] (views, batch, dim).
Expand All @@ -104,13 +104,13 @@ def forward(self, z: torch.Tensor) -> torch.Tensor:
center_loss = mu.pow(2).mean()

z_centered = z - mu
std = z_centered.norm(dim=1).div(math.sqrt(B)) + 1e-6
std = z_centered.norm(dim=1).div(math.sqrt(B)).clamp(min=1e-6)
scale_loss = (std - 1.0).pow(2).mean()

z_norm = z_centered / std.detach().unsqueeze(1)
W = F.normalize(torch.randn(D, self.K, device=z.device, dtype=z.dtype), dim=0)
p_sorted = (z_norm @ W).sort(dim=1).values
target = self._get_target(B, z.device, z.dtype).view(1, B, 1)
target = self._get_target(B, z.device).view(1, B, 1)
shape_loss = (p_sorted - target).pow(2).mean()

return (
Expand Down Expand Up @@ -207,7 +207,7 @@ def __init__(
lambda_scale: float = 1.0,
lambda_shape: float = 1.0,
lambda_center: float = 1.0,
lamb: float = 0.02,
lamb: float = 0.9,
pretrained: bool = False,
drop_path_rate: float = 0.1,
):
Expand Down
Loading