Antipasti is a lightweight toolkit for building and training deep networks with Theano.

• Tensorflow to leverage multiple GPUs: we have bought enough GPUs in the mean time to leverage efficient multi-GPU training, which is still somewhat experimental in Theano. We wish to retain the non-functional aspects and the syntax of Antipasti, but with a NetworkX backend for more powerful graph manipulation capabilities.
• More pythonic code (PEP8, streamlined control flow, proper packaging, etc.)

• Although Keras remains a beautifully written library, we have found that supporting two libraries simultaneously does neither justice. Our path of choice is to support Keras layers in antipasti-tf without committing to multi-framework support.