Basic residual network for CIFAR-10 classification from Deep Residual Learning for Image Recognition. Uses TensorFlow and Keras.
Wow, it's true! Residual connections really do make a magical difference for deep nets.
plain20
: 20-layer network without residual connections. Trains fine.plain56
: 56-layer network without residual connections. Does terribly.resnet56
: 56-layers network with residual connections. Performance matches and exceeds 20-layer network.
$ pipenv sync
$ pipenv run python -m resnet.train with net=resnet20