In my code, I print out the loss for training part, and loss and accuracy (correct/total) for the validation and test parts.
1: I find that loss in validation and test are usually 2 magnitude larger than training loss.
2: validation loss looks not useful at all, all I need right now is to use validation accuracy history and training loss history to decide whether the mode is suitable and whether the traning is overfitting or not.
Do you guys care about the validation loss value? Thanks.