Daniel Liu

This was intentional. The problem is that with an add function that directly modifies the Tensor, the backprop and forward prop code will be more complicated as we need to...

I think the current convolution layer supports 3 color channels, so not much work needs to be done in terms of updating the library. There was a bug fix that...

Yes! Adding tests is very important, though I never got around to doing it. If you want to, you can add some. Also, one day I'll figure out how to...

I know JUnit is cool and all, but I really do not want to add another dependency...

Yup, this is something we will work on. I'm wondering if there is a better way to generate C bindings that is not just making multiple copies of a function...

I'm not sure if `*` will every be directly supported internally. It will always have to be mapped to some character that fits within the scoring matrix, so SIMD lookups...

Yup, #4 is pretty much about the same thing. The fix would be to remove it, along with the hamming distance functions that rely on reading bytes as aligned integers....

IIRC the reason is because it saves a couple instructions when computing tracebacks (we can make certain assumptions on the scores). I'll look into allowing `open >= extend`.

This probably won't be implemented because I'm fairly certain adaptive-banding-like approaches don't work well when `open == extend` (this was measured in the original adaptive banding paper, and I think...

It would require a bit of work to support Neon with the library as it is currently. My implementation directly uses my own wrapper for SIMD instructions, so a separate...