Simple low level bindings for CNTK library from Microsoft.
Currently exploring ways how to interact with C++ API nicely from Rust. Expect a lot of breaking changes.
Build scripts are not ready yet, might not work outside of 64bit linux.
- Ability to train all reasonable architectures
- No memory leaks
- Reasonable ease of use
- Idiomatic Rust everywhere
- Super good ergonomics
- Avoiding all unnecessary copies
You need to have CNTK-2.3 installed and paths to includes and library files in relevant enviroment variables (cntk activate scripts does this well). You also need g++-4.8 installed (because CNTK uses it to compile things).
See examples folder.
Only works with single precision (f32 in Rust, float in C++) types. Only works with dense representations of vectors/matrices/tensors. Only works with ASCII strings for variable names and filenames.
- Passing data in and out of computation.
- Backpropagation.
- Training fully connected feedforward, convolutional and recurrent network.
- Saving and loading the model.
- Code for most operations. - Almost all, except couple of helpers.
- Demo of seq2seq model training.
- Turning c++ exception into rust panics (most of the time).
- Interop with NDArray library.
- GPU.
- Better docs.
- Finish all operations.
- Better build scripts.
- Figure out whether we want NDArrayView or go directly from Rust data to Value and back.
- Builder pattern where appropriate (Variable?).