zeonzir
commented
5 years ago Completed fix by adding a running batch counter to account to true batch size being processed as opposed the hard calculation of pseudo_batch_looper * batch_size. Will be reviewed soon and pushed.
The logging loss sum appears to be divided by the expected number of samples (i.e., the full minibatch size), instead of the actual samples processed. This results in an abrupt reduction in magnitude of the logged loss when an epoch ends.