Keras training process does not respect the given-as-argument training image number

[delhomer]

When training a model, if one passes a specific value to nb-training-images, one expects that the Keras model iterates over this specific amount of images, as many times as the number of training epochs.

However, it seems that flow_from_directory method (from Keras API) consider all the images belonging to the provided training image folder. Here one has to modify the generator building to taking this behavior into account.