Does Subclassing Keras Sequence Require implementing __next__ method?

Question

The Keras Sequence class documentation says

Every Sequence must implement the __getitem__ and the __len__ methods. If you 
want to modify your dataset between epochs you may implement on_epoch_end. The 
method __getitem__ should return a complete batch.

However the source code for fit_generator & other similar methods in training_generator.py calls next method with the generator passed as the argument. From my understanding this means, the class which subclasses from Sequence should be an iterator which requires __next__ method to be implemented.

In order to use a class subclassed from Sequence with methods like fit_generator, predict_generator, etc. is it required to implement __next__ method?

Answer 1

No, there is no need to implement __next__, if you carefully check fit_generator in training_generator.py, you will see that another API is used if the generator is a subclass of Sequence, __next__ is not used for sequences.

This was implemented that way because a Sequence can be read by multiple workers, that's why it uses a index-based API and not a stateful API like __next__.