Tools to assist modeling and training with PyTorch.
Currently, one useful training algorithm is offered: training RNNs with backpropagation-through-time without requiring training sequences to be of the same length, dynamically swapping out sequences in the batch as they get exhausted.
See doc string for
train.py for more details. Also, my blog post, Better Backprop Through Time Using Pytorch for some discussion.