Braid is a flexible and modular neural network library. It is written in python and uses Tensorflow backend. It is designed for fast development and at the same time to be able to support arbitrary network designs.
LeNet's Network Architecture
https://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf
Flexibility
There should be multiple ways to interact with the library. Design the network in Tensorflow, braid or simply provide it in a protobuf file.
Modular
The structure of the library should not be rigid. User should be able to modify parts of library according to preference.
Openness
Keep the Tensorflow backend open rather than behind layers of abstraction. It helps in modifying existing structures instead of building from ground up for unavailable network attributes and nodes.