Update loss file

[borgesa]

With reference to issue #21: I have done the update I described in the issue.

Now the loss functionality can use all PyTorch loss functions (by specifying the name in config.json). A template of a custom loss function class is included.

Remaining:

• Functionality that checks if config['loss'] matches one of the custom classes in 'model/loss.py'
  • When we implement this, we need to decide if custom loss functions ('model/loss.py' ones) or standard PyTorch losses has precedence (in case someone implements a loss function with the same name as PyTorch standard losses, either by accident or on purpose)

[SunQpark]

for loss.py

1. line 7 has minor typo: *kwargs need to be replaced with **kwargs
2. line 26: super(_Loss, self) need to be changed to super(CustomLossClass, self)
3. line 32: for the loss function and metrics I prefer using (output, target) over (input, target) since input is a reserved word in python. That way would also match my implementations of metric functions too.

for other files It seems that you want to pass arguments to the pytorch loss class, than train.py and config.json should be changed too.

"loss": "NLLLoss",
"loss_args": {
    reduction='elementwise_mean'
},

for config, and then

loss = get_loss_function(config['loss'], **config['loss_args'])

for train.py (line23)

[borgesa]

@SunQpark,

Thank you for your comments. I agree with them all, and have updated the pull request.

With regards to number 3, I used the convention that PyTorch use in the "loss" module (although I agree with you: I do not like using reserved names). Since "output, target" is already used in the repository, I followed your proposal.

[SunQpark]

Thank you for quick response! It seems that there are some discussions about using the input as variable name. I was surprised that this comes first in the google search, although the keyword was 'pytorch convention' which does not include 'input' Anyway, I will merge this PR now, thank you.