Manually set step_per_epoch over patch based training? #14481
Unanswered
Paddy-Xu
asked this question in
code help: CV
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Hi, I am using patch based training with torchio. It seems that the step_per_epoch is not calculated correctly.
Suppose I only have one image of 512 * 512 * 512, and using patch of 64 * 64 * 64. With a batch size of 2, I would assume that a proper step_per_epoch would be ((512/64) * (512/64) * (512/64)) / 2 = 256 in order to cover the whole volume (stochastically though). However, it seems that each epoch ends with only one batch of patch generation.
I can see there is a min_steps argument in trainer, but that seems to be the total steps, but not steps per epoch. Is there a way to manually overwrite steps per epoch?
Beta Was this translation helpful? Give feedback.
All reactions