How to compute loss during evaluation?

❓ Questions and Help

How would you recommend to compute loss during evaluation? for example I set my model to evaluation --> model.eval()

get prediction

with torch.no_grad(): predictions = model(images)

calculate loss between predictions and targets

SOMETHING LIKE THAT criterion = nn.crossentropyloss() loss = criterion(predictions, targets)

But above code doesn’t work

[Igal20]

My goal is to see convergence of loss of train dataset and validation dataset. Something like that. Obviously I can not set model to train mode because than my model will learn validation set as well.

Is there other way to do it?

[fmassa]

You can normalize it to be in the same range. Also, the validation accuracy is expected to go up, while the training loss should go down.