TensorX is a minimalistic high-level library to build and train neural network models in TensorFlow. It was developed to design simple neural network models with minimum verbose.
Both this library and its documentation are a work in progress, any input is welcome. You can follow the project on Github and read its documentation on readthedocs. I’m focusing on features I myself use in my research, so I’ll add components as I need them. If more people get interested in the project, I’ll create some contribution guidelines.
- Consistent API: tensorx is designed to have a simple intuitive API focused on modular neural networks with multiple layers.
- Pragmatic Code: verbose-free code is more readable, reproducible, and easier to debug and experiment with. Make it easy to use for common use cases. Do not overwhelm users with features they will not use.
- Transparency: the main goal is not to replace the use of TensorFlow or hide it behind abstractions, but to complement it with easy-to-use modular API to create and manipulate tensors.
- Focus: this is not a library to create every single “Deep Learning” model one might read about. Its about taking advantage of TensorFlow flexibility while compensating for some of its shortcomings.