Does model.save() save the model of the last epoch or the best epoch?

Question

This single liner is used to save the keras deep learning neural network model.

model.save('my_model.h5')

Does model.save() save the model of the last epoch or the best epoch? Sometimes, the last epoch does not provide improvement to performance.

Answer 1

It saves the model in its exact current state. If this statement is after the Model#fit method completion, then it represents the last epoch.

For best epoch (assuming best == smallest loss or greater accuracy), you can use the ModelCheckpoint for this:

epochs = 100
# other parameters...

model.fit(x, y,
          epochs=epochs,
          validation_data=valid,
          verbose=2,
          callbacks=[
              TerminateOnNaN(),
              TensorBoard('./logs'),
              ModelCheckpoint('best.h5',
                              save_best_only=True),
              ...
          ])
# the model is holding the weights optimized for 100 epochs.
model.load_weights('best.h5')  # load weights that generated the min val loss.