Keras uses parallelism extensively in its codebase. Most tasks are embarrassingly parallel: data loading, data augmentation, and other preprocessing techniques are all easily parallelized.

To feed a Model, keras takes a batch of inputs from an infinite generator and feed it to the Model.

# Simple example
for batch in infinite_gen:
    model.train_on_batch(batch)
    ....

The generator can be anything, but right now we support two types: a Python generator and a keras.utils.Sequence generator.

A generator is an object allowing you to iterate over the return value of a function. In Keras, we require that those generators are infinite. Example:

while True:
    for batch in my_batches:
       yield batch

To manipulate those generators, we use a keras.utils.GeneratorEnqueuer.

A GeneratorEnqueuer creates a pool of processes and sends the generator object to each process. These objects are NOT shareable so all processes get a copy of the original, this causes the results to be duplicated.

Then, the enqueuer asks each process for a new input. The result is store inside of a Queue. Keras can then pop the queue to get a new batch.

A keras.utils.Sequence is a safer way of doing it. To manipulate those, we use a keras.utils.OrderedEnqueuer. This enqueuer is almost the same as the GeneratorEnqueuer, but it asks for a specific input to each process. This is feasible because keras.utils.Sequence is indexable. This way, we avoid duplicates.

• https://docs.python.org/3.4/library/multiprocessing.html?highlight=process
• https://docs.python.org/3/library/concurrent.futures.html
• https://dref360.github.io/managers/
• https://dref360.github.io/executors/