ionn 0.0.12

io for neural networks

# IONN io-operations for artificial neural networks

IONN provides functionality to store, load and freeze neural networks and convert networks between different neural network frameworks. The current version provides

1. Storing, loading, freezing of tensorflow models in google protobuf files (submodule tfpb)
2. Dumping keras models as google protobuf files and loading them into a pure tensorflow environment (submodule k2tf)

## tfpb - freeze and store graphs

Tensorflow provides a graph freezing tool that works ok, but is hardly documented and not particularly modular. The tfpb module provides a simplified interface to storing frozen graphs. There are two main entrypoints, load_protobuf and save_protobuf. Furthermore, you can directly call tfpb to freeze stored graphs like this

tf-freeze <input_graph_file_name> <output_file_name> <checkpoint_file_name>

## k2tf - From keras to tensorflow

Keras is nice if we want to quickly draft out a neural network architecture. Unfortunately, it differs considerably in how it stores models and can therefore not well co-exist with tensorflow infrastructure. k2tf supports storing keras models in tensorflow protobuf files that can later be loaded without keras. There are currently two drawbacks though:

1. Models have to be frozen, which isn’t exactly desirable because most of tensorflow’s strength is in tweaking the models during the learning phase.
2. Models have to be reloaded in a separate process to avoid confusion about the tensorflow graph.