klebs6/caffe2-rs

caffe2-rs

This library is based on the flexible Caffe2 C++ project.

It provides a high performance and modular rust environment for deep learning.

As of July 1, 2023, the codebase is under construction: it is ready for interface experimentation and development work, but not yet for production.

If you are interested in this project and would like to contribute, your contribution is both welcome and appreciated!

The current bottleneck is in the translation of c++ statements into rust. It is possible to do this manually. However, it is better to do it automatically. Some work is being done in this domain, however there are several challenges to be overcome.

For more information, see chomper: https://github.com/klebs6/chomper

Note from original translator (klebs6)

Greetings! I have a few preliminary work-items to attend for a few months. Once these are done, I plan to productionize the chomper.

This will include integrations with the rust-ecosystem compiler tools.

During the productionization, we will be able to enable name resolution and type inference during transpilation (not to mention several performance wins).

This work will provide a massive speedup for the translation. Subsequently, caffe2-rs development will be free to progress.

If there are developers out there who would like to contribute to the c++ to rust transpiler development track, I would love to hear from you. Please feel free to reach out. I am sure the rust ecosystem will be grateful. C++ to Rust transpilation is a major bottleneck for many projects.

Acknowledgments

This project owes its DNA to the original authors of the Caffe2 C++ library, without whom, this work would not exist.

The author of this project would like to thank the original authors of the Caffe2 C++ library for their work, as well as the Rust community for their support.

caffe2-rs's Issues

caffe2-aten crate does not build

known issue! working on it!

As of 4/24/2023, i have a medium sized stack of work to do before i can bring full attention back to this, but i am grateful for any support/feedback in the meantime!

The plan for now is to modularize caffe2-aten and caffe2-c10 along the same lines as the rest of the workspace. That way, dependent crates can pluck just what they need from aten/c10, instead of having to link the whole thing.

collapse caffe2 core crates into a single crate

It looks like the core of this workspace, namely these:

will be more useful once collapsed into a single crate.

We can use cargo feature flags to keep a "pay for what you use" model in the core.

This is an open issue -- will accept a PR implementing this change.

merge families of operators

For example, caffe2op-tan, caffe2op-sin, caffe2op-abs, etc may be better suited in a single caffe2op-math crate.

I don't think we ought to do this everywhere, but for certain groups of operators it makes sense.

Currently, the crate layout is roughly maximally granular. This can be good in certain cases. For certain groups of small operators, though, I don't think it needs to be this way.

Another example is merging caffe2op-do, caffe2op-while, caffe2op-if, etc into caffe2op-controlflow

Open to discussion.

Too many crates?

Searching for "neural" or "tensor" on crates.io almost exclusively yields results from this repository, as you have about 250 crates published on crates.io.

This is a bit much, if you'd ask me. It's flooding the search results and making using crates.io's service harder.

The fact that you had to automate generating the READMEs of all these crates using ChatGPT should have been a hint that this wasn't a good project architecture.

I hope you will listen to my concerns and address this issue.

Whats up with the readme?

Why is the readme just a transcript of ChatGPT? I am confused...