Five Epochs, Three to One

A GAN is trained on a dataset of 48,000 photographs, following the standard alternating recipe: the two networks never move at the same time, so each update is computed against an opponent that is frozen for that step.

• One discriminator update consumes a batch of $64$ real photographs together with $64$ freshly generated samples.
• One generator update consumes a batch of $64$ noise vectors and no real photographs at all.
• The team uses an update ratio of 3 discriminator updates to 1 generator update: after every third discriminator update, the generator is updated once.

An epoch ends when the discriminator has used every one of the 48,000 real photographs exactly once. The run lasts 5 epochs, and every batch divides evenly throughout.

How many times are the generator's weights updated over the whole run?

Give the exact whole number.