Adding MAE loss and more customizability to discriminator

[ramanakumars]

• Adding MAE loss for img2img translation
• Moving filters into generator and discriminator in the configuration so that we can also pass the boolean norm into the discriminator
• Moved InstanceNorm before activation since this was causing issues

• Move additional customizability into the generator and discriminator configs (e.g., activation, use_dropout, final_activation, n_layers)
• Test norm change on other datasets

[AgentM-GEG]

For future: Have we historically used Tanh()? would it benefit to have an activation_layer = nn.Tanh() as input args?

[AgentM-GEG]

Looks good; In the ethos of our other repos, for future it may be beneficial to have Adam vs. NAdam vs. Adagrad etc as input configurable variables. Of note again are the hyper params for Adam/NAdam .. sometimes betas = (0.9, 0.99) work better and sometimes betas (0.9, 0.999).