search

kdmalc / personalization-privacy-risk

Privacy analysis for ML and classical filtering personalization parameters
0 stars 0 forks source link

Fix batch size issues: some bs's break the pipeline for some reason #52

Open kdmalc opened 9 months ago

kdmalc commented 9 months ago

This might be related to sequence length and such too... maybe not explicitly setting the sequence length would fix this? Not sure. I don't have the values that break the code so would have to experiment.

This is only for the deep networks, AFAIK.

kdmalc commented 8 months ago

If the batch_size times the sequence_length results in not enough columns ("samples"), then you either get imbalanced batches, or if drop_last is True, then you drop the last batch. Note that if batch_size times sequence length is greater than the total number of samples, then you only have one (unfilled) batch, which would then get dropped due to drop_last, resulting in empty trainloaders.