rust-ndarray VS fundsp

Is the ndarray crate a good fit for this scenario? https://github.com/rust-ndarray/ndarray

ndarray and numpy crates provide good way to operate on numpy ndarrays from python

I don't think any of the major ML projects have GPU acceleration because ndarray doesn't support it.

Not sure about ndarray: https://github.com/rust-ndarray/ndarray/issues/840

I used basic_dsp a while back and found it lacking. I was hoping to find something that uses the ndarray datatype but i'm not seeing this yet. If you're primarily trying to learn though it might not really matter which library you contribute to. As for myself, I just picked the one that was most used and actively worked on at the time. However I keep an eye out on other libraries; if I see something take off, I might switch over. Either way you'll learn and can point to it as work accomplished.

Most libraries use either an older C/C++ dynamically or statically. I was looking into maybe rewriting some image manipulation functions in Rust using ndarray (which would then allow for both hardware, gpu and multithreading acceleration), but it would take a very long to get fully into the subject.

The best way would to use a crate for this, such as ndarray, which provides some good ways to do this.

First, let me explain a crate, where old dependencies do matter. There is one very handy crate called nshare. It's used for converting images into ndarrays. The libraries between nshare converts are regularly updated, for example to support new formats or to make the usage more convenient. If one writes a new software there could be conflicts in versions of ndarray (for example), so we make sure, that nshare is always up-to-date and works with the newest versions in order to build the best code on it. Another example could be everything cryptographic. If there would be a security issue with ring we really want every dependent to be updated.

well you can represent categorical values in `ndarray` for sure (even structured arrays [here](https://github.com/rust-ndarray/ndarray/issues/32)), but the memory has to be contiguous for BLAS/LAPACK and therefore it is impossible to mix continuous and categorical values. I was thinking that we could emulate categorical values with a descriptor field for the type of each feature and then just use floats to represent them.

ndarray Readme links to matrix https://github.com/rust-ndarray/ndarray

- https://github.com/SamiPerttu/fundsp

Check this video, you can even live coding audio effect in DAWs with Glicol:

Here's a code fragment that can be plugged into the beep example to replace the let mut c = ... part:

That's a nice one focusing on audio: https://github.com/SamiPerttu/fundsp

https://github.com/SamiPerttu/fundsp seems promising and I know they're trying to make the code DSL like to the point where it could be parsed from a file but I don't think that actually works yet.