Error

[jgdwyer]

No description provided.

[jgdwyer]

I noticed that when I weight my training examples non-uniformly when training a regressor type MLP, my validation error becomes an order of magnitude better than my training error. Digging into the code it looks like the loss function purposely doesn't account for the weights when calculating the validation error. Given that my stop condition for training my neural network is that the validation error is steady for some number of iterations, I would have thought the validation error should be calculating with weights. Is there a specific reason it's not?

if mode == 'train': loss += processor(Xb, yb, wb if wb is not None else 1.0) else: loss += processor(Xb, yb) count += 1

See lines 322-325 of backend/lasagne/mlp.py