Mentions légales du service

Skip to content

Fix #313 bug on reporting progress in torch TrainingPlan

CREMONESI Francesco requested to merge bugfix/313-torch-training-progress-reporting into develop

See #313 (closed) for the bug description.

Main changes:

  • Introduced a test that reproduces the problem
  • I reordered the batch_maxnum exit condition to be after the logging, otherwise there was no logging output during the last batch (especially frustrating in the common scenario where batch_maxnum=1
  • batch_size is no longer inferred from len(data), but it is obtained directly from the data loader.
  • to report progress on the number of samples processed until now, only doing batch_size*batch_ may lead to problems in the last batch; instead, I wrapped it in a min function min(batch_size*batch_, total_num_samples)
  • Fixed researcer-side reporting as well (did not implement the corresponding test for this)

Merge request reports