Hi everyone,
I want to modify the value of my learning rate at each step instead of doing it at the end of each epoch.
In particular at each step of my training I compute a generic measure x and I want to modify the learning rate for the next step as a function of x.
Just to give a more explicit idea of the pipeline, given the constant C what I want to do is something like:
for epoch in range (0, n_epochs):
for data, label in train_loader:
#training
#computation of x
#modify of the learning rate as a function of x (for example lr = C*(1+x))
Till now I’ve only used lr_schedule = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda = lr_rule)
to modify the learning rate value.
Is there a more flexible method in pytorch to modify it at each step and with a generic function of a generic computed measure?