Mumu is a serialization toolkit for neural networks built with keras. It is being built concurrently with an efficient consumer in Java. The goal is to make a portable abstraction of arbitrarily structured nets so that they can be deployed with low-latency on commodity hardware. Eventually this might expand to include a limited subset of torchnn.
Pull requests are welcome.
Requires protobuf >= 2.6
make
nosetests
python setup.py install