Ported Contextual loss and Gradient loss from Victorca25's fork.
Now that contextual loss has been shown to be useful I figured I'd create a PR for this. Gradient loss seems useful as well since it's basically what SPSR does without having to use SPSR.
Ported Contextual loss and Gradient loss from Victorca25's fork.
Now that contextual loss has been shown to be useful I figured I'd create a PR for this. Gradient loss seems useful as well since it's basically what SPSR does without having to use SPSR.