Kyle Daruwalla

I think the correct thing to do would be to define `matmul(x, y)` similar to `mul` in this repo. You can define different methods of `matmul` for different types of...

> Did you actually download the model and test these lines, or are these just an example? Yes, these are just an example. Do you know what size input your...

Okay, I downloaded your model and took a look. It seems like your input should be "400 x N" where N is the batch size? In that case, the dummy...

> Flux crowd seem to prefer that Flux import of ONNX models return a Flux.Chain, even if this means failing to import models which can’t be expressed with Chain. Potentially...

Having recently worked with `Ghost.Tape`, I can say it would be a really nice structure for this. We might want to add some more graph-like helper functions for easier traversal...

Looks like a good starting point to me. Perhaps the tape context can store a backend selection (defaulting to `NNlib`). Then `load_node!` can dispatch on both the stored `OP` and...

> So far adding/migrating operations is easy as long as there's a counterpart in NNlib. Unfortunately it's not always the case This is why I think handling multiple libraries now...

> Would be in Flux or a separate package, not ONNX.jl Yup, exactly. It will also help iterate on this package without waiting for Flux to move stuff to NNlib...

> I suppose backends and translating a Tape to some other representation are two kind of redundant features and we should try to decide which one is the recommended way....

Nice, thanks for the response. That design looks great for examples in docstrings where we do some kind of check to indicate to the reader that "yes, this function returns...