benchmark with scipy or numpy implementation?

[Xiaojieqiu]

Hey @ritchieng thanks for implementing this method so quickly! I am curious about the performance of this method comparing to other standard methods, implemented in scipy or numpy. Is it possible for us to see some benchmarking regarding this implementation?

[ritchieng]

Hi @Xiaojieqiu, this should be easy. I'll provide a comparison of NumPy vs PyTorch by this week.

For large matrices, I would say a highly optimized PyTorch version (JIT and CUDA) would be very fast because you can run many function calls in parallel on the GPU and once you compile the function, you don't need the Python interpreter on subsequent calls. This will be implemented #2

[Xiaojieqiu]

that will be wonderful, I am looking forward to seeing those benchmarks

[hypnopump]

any progress??