You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Question about PositionalEmbedding used for DiscreteDiffusionSDE #54
In the forward function of the class PositionalEmbedding,there is x = x.ger(freqs.to(x.dtype))
However, if x.dtype is int64, freqs.to(int64) will lose some values.
For example, in line 528 of diffusion/diffusionsde.py , the class DiscreteDiffusionSDE's sample method, t.dtype is torch.long. t = torch.full((n_samples,), sample_step_schedule[i], dtype=torch.long, device=self.device)
Maybe my understanding is not correct, would you please to show where to handle this case? Thank you!
Activity