Accelerate via CUDA / gpu

[Alex Lane]

Hey all,

I'm a bit lost into my first setup utilizing my GPU and training with gensim. Googling and stackoverflow reveal people looking for tensorflow-gpu, but not much on gensim. How can I utilize my NVIDIA gpu and gensim to improve training times?

[Ivan Menshikh]

Hi Alex,

Please check issue Word2vec to use GPU #449 and this PR [WIP] TensorFlow wrapper for using GPU #1033. 

[Alex Lane]

Thanks Ivan,

alright so gensim CPU is best we got so far then?

[Alex Lane]

I am trying to Doc2Vec

[Ivan Menshikh]

Yep, you can read this benchmark

[Alex Lane]

Would a benchmark test on an EVGA 1080ti ftw3 help? I'm building up a tensorflow Word2Vec - Doc2vec program as it hasn't been tested.

[Radim Řehůřek]

Absolutely! Thanks for the offer Alex.

A proper, thorough benchmark on some proper HW would be awesome :)

Ivan, Lev -- can you please assist Alex there?

I'm thinking -- let's also throw our AWS and IBM SoftLayer machines at it, and compare CPU/GPU across these managed services as well (cost/benefit/perf on w2v/d2v). Should be an interesting data point for others.

Best,

Radim

[Alex Lane]

Thanks so much for all of your and your team's work Radim. I have a simple question. I need to compare certain docvecs with eachother. the tags look like tags=[u'HeaderKey + '_' +str(idx)]. i need to compare docvecs with the same headerkey to eachother. If I can infer_vector() and get a similarity, that would also work, but it appears its still a TODO?

[Ivan Menshikh]

You can check this tutorial to see how you can work with tags.

 

I can infer_vector() and get a similarity, that would also work

You are right, this is more intuitive and clean way

[Alex Lane]

I can infer_vector() and get a similarity, that would also work

So I can do that? It looks like its still a TODO- model.similarity(model.infer_vector('sentence and words and stuff'),model.infer_vector('another sentence and words and stuff'))

sorry, the tutorial didn't really have what I was looking for.

[Gordon Mohr]

You can `infer_vector()` on new texts – but they should be preprocessed/tokenized in the same way that texts were for training the model. (Don't pass strings, as in your post, but lists-of-tokens.) 

Also, note that better inference may be achieved with far more `steps` (an optional parameter to `infer_vector()`, default 5) or a lower starting `alpha` (default 0.1, but the same default as used in training, 0.025, may be more appropriate). 

- Gordon

[Alex Lane]

Thanks Gordon! I'll tokenize and pass that, but my issue is, these have been trained in the model. How do I grab a set of docvecs with similar tags?