Calculate the gradient of pytorch model

1

I have a model trained for four epochs and I saved this model after each epoch. Is there is a way to calculate the gradient as the difference between two models I mean between the saved model after the fourth epoch and the model saved after the third epoch.

2

I’m unsure what exactly your use case is, but in case you want to calculate any difference between parameters you could directly iterate these as seen here:

modelA = models.resnet18()
modelB = models.resnet18()

diffs = {}
for (name, paramA), (paramB) in zip(modelA.named_parameters(), modelB.parameters()):
    diffs[name] = (paramA - paramB).abs()
3

@ptrblck Thanks for your reply, what I want to do it to calculate the gradient as the difference between the two model weights i.e.
gradient = (W_n ) - (W_n-1), where n is the number of epochs.

4

My code shows how to calculate the difference between the parameters of two different models. In your case both models would have been trained for some epochs while I just use randomly initialized models as an example.