[Questions] Only calling Dataloader once per epoch? Training classifier from scratch at each cycle?

[WurmD]

[Questions] Only calling Dataloader once per epoch? Training classifier from scratch at each cycle?

[pranaygupta36]

I believe we could have used a better naming convention for the variables. The variable 'epoch' here refers to each iteration. Basically a random batch of 64 samples is loaded each time next(iter(train_loader)) is called. Therefore the VAEs sees all the samples while training.

To answer the second question, no we didn't resume training the classifier from where it was left off in the last cycle, because after each cycle the generated embeddings pertaining to the unseen classes are different, and the classifier is trained such that it optimally classifies the embeddings generated at the end of the current cycle.