Candle

Simple PyTorch helpers. Example features:

• Simple modules like Lambda, Flatten, View, Concat, Split, SkipConnection.
• NoInputSpec, which wraps another module to dynamically figure out its input size the first time it's called.
• CannedNet and as an example generalisation CannedResNet, which provide a straightforward and extensible way to assemble sequential neural networks, but with some extra features, like not having to figure out input sizes, and easy specification of ResNet (or ResNet-in-ResNet etc.) style architectures.
• create_supervised_trainer for a simple trainer with all the boilerplate written for you. (Depends on the excellent ignite framework.)
• Window and Recur for a simple way to create complicated recurrent architectures.
• Utilities like tensor_product and batch_flatten.

Plus a few more things!