tracel-ai/burn

Autodiff benchmark comparison

Open

#1,559 opened on Mar 29, 2024

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Rust (964 forks)auto 404
help wantedperformance

Repository metrics

Stars
 (15,550 stars)
PR merge metrics
 (PR metrics pending)

Description

Hey, thanks again for the call. As discussed, it would be nice to compare the AD performance on Microsoft's ADBench Benchmark set. Here is a PyTorch implementation, since you mentioned that this might be best to get started from burn: https://github.com/microsoft/ADBench/blob/master/src/python/modules/PyTorch/gmm_objective.py Here is a Rust implementation: https://github.com/EnzymeAD/Enzyme/blob/64fd1d341bd171f3f54beacc92f81bd81f59a53d/enzyme/benchmarks/ReverseMode/gmm/src/lib.rs

If someone has the time to set up a burn version of any of gmm/ba/lstm/hand, then I would love to run some benchmarks against it. The main issue is probably that burn is more high-level so this benchmarks might be harder to model, while rust-enzyme isn't stable enough yet to test classical Neural Network implementations. So it might also might sense to postpone this for a while.

Contributor guide