What does self.iterations refer to in Keras' Optimizer class

Question

I've been trying to understand Keras' Optimizer class, and realize that there's a variable I don't quite understand - self.iterations. Does this refer to:

1. The number of individual samples for which updates have been performed?
2. The number of individual batches for which updates have been performed? (This is what I believe)
3. The number of total epochs (i.e. complete passes thru a training set) for which updates have been performed?

Answer 1

It's 2.

The entire keras function is iterated once per batch.

One way to test is to get a small array of data and train for one epoch:

#get 3 batches of size 32 from the data
small_X = X_train[:3*32]
small_Y = Y_train[:3*32]

#print the initial value of iterations
print(keras.backend.eval(model.optimizer.iterations))

#train for 1 epoch with batch size 32
model.fit(small_X, small_Y, epochs=1, batch_size=32, verbose=0)

#see the new value of iterations
print(keras.backend.eval(model.optimizer.iterations))