keras model fit generator several times slower than model fit

Question

In Keras 1.2.2, there are referencing merge and it has multiprocessing included, but model.fit_generator() is still about 4-5x slower than model.fit() due to disk reading speed limitations. How can this be sped up, say through additional multiprocessing?

Can anyone help me with this?

Answer 1

You may want to check out the workers and max_queue_size parameters of fit_generator() in the documentation. Essentially, more workers create more threads for loading the data into the queue that feeds data to your network. There is a chance that filling the queue might cause memory problems, though, so you might want to decrease max_queue_size to avoid this.