Closed Maw-Fox closed 1 year ago
There may be an issue that wasn't found during preliminary tests, I am going to close this current pull request until I can manage to make sure this issue is unrelated to these commits.
What's the issue?
What's the issue?
Not sure yet, still trying to figure out what is happening on their end, it could be a variety of factors but it's likely not this code (just making sure that it isn't before I opt to pull req something that could potentially break things with little-to-no error reporting regarding it).
Seems to be freezing up and not progressing for them, but allegedly it works with 1:1 512x512 images. Though this makes me feel like this isn't an issue with this code in particular. They have struggled to get the newer version of WD's trainer to function on Windows, so it's really hard to say whether or not the issue lay with WD, their env, Windows, etc. My money is on env, but can't be sure.
Seems to be working for someone else though after some more testing so chalking it up to the other person's environment.
Features:
--resize
argument (boolean) to allow people to use not previously-prepared datasets more easily and perform the associated scaling and cropping as necessary to match the best-fitting bucket. Defaults toFalse
.Fixes:
collate_fn
that hasn't been normalized in this way will cause atorch.Size
discrepancy. E.g:torch.Size([3, 512, 512])
,torch.Size([4, 512, 512])