Marco

Opened https://github.com/ncase/nutshell/issues/13 to address the other issues.

There are two things to improve: 1) add support for `padding_idx` (and in `Embedding`) 2) reduce allocations (some of the methods don't operate entirely inplace on a workspace matrix)

I made a few updates following the suggestions presented above. In particular, I have changed the `inputs`/`offsets` method to be more Julian. I originally wanted it to be identical to...

Thanks for the comments. I'll review them tomorrow. > The main outstanding question I have is the utility of input/offset style input. What's the value of specifying inputs in this...

> > I included it specifically for feature parity with Pytorch. I agree that it is cumbersome compared to the vector of vectors input, but I think it has utility...

Is there anything else necessary for this PR? Maybe some improvements with mapreduce can be done, but are not possible right now.

Thanks for the comments, I'll check it soon and hopefully we can move forward.

Yes I lost the thread also. I'm happy to wrap it up this weekend, if the proposed changes are what we want.

I have noticed that wrapping multiple JSON objects in `[ ]` with a comma separator causes this to parse correctly: ```julia julia> s = "[{\"name\":\"Marco\"}, {\"name\":\"Julia\"}]" "[{\"name\":\"Marco\"}, {\"name\":\"Julia\"}]" julia> JSON.parse(s)...

Sorry, one more thing. It works if you repeatedly read from a stream: ```julia julia> s = "{\"name\":\"Marco\"} {\"name\":\"Julia\"}" "{\"name\":\"Marco\"} {\"name\":\"Julia\"}" julia> stream = IOBuffer(s) IOBuffer(data=UInt8[...], readable=true, writable=false, seekable=true, append=false,...