Highway networks implemented in PyTorch.

Just the MNIST example from PyTorch hacked to work with Highway layers.

• Make the Highway nn.Module reuseable and configurable.
• Why does softmax work better than sigmoid? This shouldn't be the case...
• Make training graphs on the MNIST dataset.
• Add convolutional highway networks.
• Add recurrent highway networks.
• Experiment with convolutional highway networks for character embeddings.

• ELU doesn't work better than RELU for the layer activation.
• Softmax seems to work better than sigmoid for the gate function?!