Clarification about keras.utils.Sequence

1. Anything you want, considering that one epoch will get len batches from the Sequence.
2. There is no secret, use it as any other generator, with the difference that you may do steps_per_epoch=len(generator) or steps_per_epoch=None.
   • max_queue_size: any value, this will load batches that will be waiting in memory until their turn to get into the model
   • workers: any value, this will be the number of parallel "threads" (forgive me if the name is not precise) that will be loading batches
   • use_multiprocessing: I don't know this one. It was not necessary for me and the only time I tried it was buggy enough to freeze my machine
   • shuffle: From the documentation: Boolean. Whether to shuffle the order of the batches at the beginning of each epoch. Only used with instances of Sequence (keras.utils.Sequence). Has no effect when steps_per_epoch is not None.
3. I think this is it. If you want to thread the model itself, then you might want to read about multi GPU training, I guess.

Advantages of Sequence over a regular generator:

With sequence, it's possible to keep track of which batches were already taken, which batches are sent to which thread for loading, and there will never be a conflict because it's based on indices.

With generator, parallel processing will lose track of what batches were already taken or not because threads don't talk to each other and there is no other option than yielding batch by batch sequentially.

Advantages of generators and sequences over a loop

In a loop, you will "wait for batch load", "wait for model training", "wait for batch load", "wait for model training".

With fit_generator, batches will be loaded "while" the model is training, you have both things happening simultaneously.

For very simple generators, there won't be a big impact. For complex generators, augmentators, big image loaders, etc., the generation time is very significant and may severely impact your speed.