Athena is a deep learning framework in C++17. It is designed for maxima performance and the ablity to run on any systems and accelerators without being vender locked-in.

Athena provides

• The ablity to easily implement a computation backend
• Tensor operators colsely resambles numpy's
• High level neural netowrk API
• Native NNPACK support
• Integration with xtensor

Athena handles tensor like how a game engine handles shaders and meshes. The Tensor class does not creare nor operate a tensor itself. Instead, it calls a Backend that does all the job. This design allows Athena to load backends at runtime and even utilize multiple backened at once.
Backends in Athena can do more than performing tensor operations. Algoritms can be inserted into a backend then quary at runtime to perforem specific tasks if the uderlying library has a way to do such operation faster then using tensor operations.

Currently there are 3 backends avliable. XtensorBackend using the xtensor, NNPackBackend using NNPACK, ArrayFireBackend using ArrayFire. XtensorBackend is aimed at providing full fledged tensor operators and useable algorithms. And the NNPackBackend provides fast algorithms for other backends if running on sutable machine.

The ArrayFireBackend is capable of using the CPU the GPU to compute. But it have sevier limits. Ex: tensors created from ArrayFireBackend can only have at mose 4 dimentions. And the ArrayFireBackend suffers from performance and stablity issue within ArrayFire it self. This is under addressing now. Use it at your own risk.

More backends built on faster libraries like Eigen and, MIOpen are planed.

example1 in the examples folder demostrates how to implement a MLP to solve the XOR problem in Athena.
mnist contains an example using a MLP to classify MNIST digits.

Since deep learning and it's community is rapidly growing, we'd love to get contribution from you to accelerate Athena's development. Feel free to open a pull request or an issue to report a bug or request features.