Ben Johnson

Interesting -- are you thinking just for inference or for both inference and training?

Relevant code and paper, if you haven't seen it: https://github.com/rdspring1/LSH_DeepLearning I don't think they did it on GPUs or big models, but maybe interesting

Any updates on this? I'm trying to find an example of a library that uses approximate kNN methods to speed up the output layer.

You can use the Chebyshev option instead of the full eigendecomposition, but it's still not going to be great. I've done a fairly major refactor of this code that scales...

There are scripts for converting data here: https://github.com/bkj/pytorch-graphsage/blob/master/utils/run-convert.sh And invocations of train here: https://github.com/bkj/pytorch-graphsage/blob/master/run.sh You may find this repo helpful: https://github.com/bkj/graphsage-age-prediction That's a boiled down implementation of graphsage w/ instructions...

Yeah, likely you need an older version of pytorch. I can put a requirements.txt here later. If I don't do it in the next few days ping me here again...

That sounds like a pytorch version problem, I think this expectes 0.3.1, but I'll have to double check

I'm trying to run some larger-scale hyperparameter tuning experiments --`warm_start` would be particularly useful. A `verbose` flag would also be very nice to keep logs clean.

Ah thanks. I figured that Levenshtein automata would work for this, but I couldn't find a python implementation. The one that you linked computes Levenshtein distance but according to the...

OK thanks. Do you have any pointers to a well-tuned model? ~ Ben