C1W1: how many steps will it run to train the GAN?

On the last step to train the GAN, how many steps it’ll run? Mine has gone to 69,000 steps and it is still running.

From the parameters, n_epochs = 200, batch_size = 128. Should it run 200 * 128 = 25,600 steps?

@seanjiang, the number of steps is the number of batches that are run. The number of batches per epoch depends on how big the dataset is, so unless you know that, it’s hard to say how many steps will be run. To get a better sense of how much longer you have to wait, some of the labs also print the epoch number in the line that prints the step. You could add that yourself here or in any other lab if you notice it’s missing.

For this lab, it looks like there are about 93000 steps. The note in the assignment says, “On a GPU, this should take about 15 seconds per 500 steps, on average, while on CPU it will take roughly 1.5 minutes”. Actual times can definitely vary from this, but that gives you a rough idea. If you’re using the Coursera environment, you should be on a GPU.

93000 steps at 15 seconds per 500 steps works out to just over 45 minutes - sort of long, which I suspect is why they made this part optional. But, you should definitely run it for a while so you can see the digits start to emerge as the model trains.

1 Like

Thanks much for the clarification, Wendy!

1 Like