Your code yields absolute errors <=1e-7 due to the limited FP32 precision, so use `torch.allclose` to compare the output. Also the general statement is not correct, as e.g. batchnorm layers will return results which depend on the batch size.
shouldn't f(batch)[:k] == f(batch[:k])?
what am i missing? cc @PyTorch @ptrblck_dehttps://github.com/pytorch/pytorch/issues/31919 …
-
-
-
Good point on the batchnorm layer, didn't think about that. I understand that FP32 precision can lead to absolute errors <= 1e-7, however when you compare `f(batch)[:batch] == f(batch[:batch])`, they're all true. i.e FP32 is no more a problem...
- Još 3 druga odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.