Assert prevents mixed_precision?

Question

Assert prevents mixed_precision?

mattmisk opened this issue a year ago · comments

Line 2661 of imagen_pytorch:

assert images.dtype == torch.float, f'images tensor needs to be floats but {images.dtype} dtype found instead'

This assert triggers when using mixed_precision = True. Am I doing something wrong, or is this a bug? Everything works fine if I comment the line out.

Phil Wang · Answer 1 · Thu Jul 20 2023 23:05:48 GMT+0800 (China Standard Time)

@mattmisk oh nice catch! put in a quick fix

Phil Wang · Answer 2 · Thu Jul 20 2023 23:44:43 GMT+0800 (China Standard Time)

@mattmisk generated images look ok?

mattmisk · Answer 3 · Fri Jul 21 2023 00:07:35 GMT+0800 (China Standard Time)

@mattmisk generated images look ok?

Yes, but my application is unique and still in development (dataset acquisition, processing, etc.). It's on my list to do a head-to-head comparison with and without mixed precision.