Hand-implement ResNet

[jiwoncpark]

Hand-implementing ResNet instead of calling torchvision's will enable us to:

• modify of the layers (dropout, concrete dropout)
• remove fluff (like pretrained options)
• save GPU memory since we can bypass concatenating 3 channels
• save training time since we can bypass normalization of the 3 channels

[jiwoncpark]

Done and closing (training time halved! ~35s/it --> ~20s/it).