CODEWITHSUNDEEP
pythonkeras
Jash Shah
Jash Shah

Reputation: 2164

Keras Sequence, fit_generator and steps_per_epoch

I've noticed that with fit_generator, the steps_per_epoch parameter is usually assigned total_samples//batch_size, where one can create a generator/use ImageDataGenerator and pass it as an argument to fit_generator.

However I am using the Sequence class (keras.utils.Sequence()) to create my generator and passing steps_per_epoch an integer less than total_samples//batch_size.

What I would like to know is would the generation of data start in the generator start from the beginning once each epoch is completed?

For example, I have 3200 samples in my training set and I use a batch size of 32. So ideally for one complete epoch I should set steps_per_epoch to 100. However what would happen if I set my steps_per_epoch to 50? Once the first epoch is completed would data point number 1601 (32*50) be generated or would it start from the beginning (data point number 1) ?

Upvotes: 3

Views: 783

Answers (1)

Dr. Snoopy
Dr. Snoopy

Reputation: 56357

When using Sequence, you do not need to pass steps_per_epoch, as this information can be inferred from the __len__ method of your Sequence.

If you pass steps_per_epoch while using Sequence, this will override any use of the __len__ method and it will effectively only use steps_per_epoch samples from your sequence (from 0 to steps_per_epoch - 1), and it will reset back to zero at the end of the epoch. You can check this behavior in the keras source code.

Upvotes: 4

Related Questions