Manually set step_per_epoch over patch based training?

[Paddy-Xu]

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?