This is to become a great project someday...
TODO / IDEAS:
DATA:
-balancen
-random crop
-noise (gaussian, salt & pepper)
-flipping
-rotating
-lighting
-scaling (padding || distortion && no padding)
NETWORK:
-resnet blocks
-inception blocks
-batchnorm
Training:
-cyclical lr
-batch growth
-learning rate warmup (bei scratch train)
-cosine lr decay (?)