I use this web(https://github.com/dusty-nv/pytorch-classification/blob/5107de352cade326e7aacd44bf8d625b7055fd4e/train.py)to use train.py. I want to add the early stopping method to train.py. How to do it?
You can either use the validation loss or accuracy to determine the early stopping. Its better to do it on the epoch validation loss. In the validate function, you can keep track of the running loss and compute the epoch loss outside of the loop. In addition, save the best validation loss (initialized with a very large value) and introduce a patience variable (initialized with 0). Both can be global variables or passed as an argument to the validate function:
def validate(val_loader, model, criterion, num_classes, args):
running_loss = 0.0
for i, (images, target) in enumerate(val_loader):
running_loss += loss.item() * images.size(0)
epoch_val_loss = running_loss / len(val_loader)
if epoch_val_loss < best_val_loss:
best_val_loss = epoch_val_loss
patience = 0 # reset patience if the loss improved
else:
patience += 1 # increase patience counter
if patience == patience_threshold: # loss didn't decrease for 'patience_threshold' epochs
return early_stopping # can be a flag to indicate early stopping and break out of the epoch loop
You can do it similarly using the accuracy metric. This is a very crude way to do early stopping. There can be more formal ways and modifications. Feel free to change it according to your needs.
I am very sorry to say. Would it be convenient for you to show or write the early_stopping code with train.py to me?
How to use dropout method on train.py