search

lyst / lightfm

A Python implementation of LightFM, a hybrid recommendation algorithm.
Apache License 2.0
4.77k stars 691 forks source link

Serving Big LightFM Model #342

Closed ahayamb closed 6 years ago

ahayamb commented 6 years ago

I've watched your talk and it sounds interesting to serve the model via postgres. If it's permissible to know, what are your use cases? How fast the prediction can be done? I've tried apache madlib to do array operation in postgres and want to see whether I did something wrong Please close this if you found it irrelevant, thank you

maciejkula commented 6 years ago

I would not recommend doing that. In hindsight, a service with some models in memory and responding to requests is a much simpler solution.

RothNRK commented 6 years ago

In my case latency is an issue so I make predictions for each user/lot, store the top_n recommendations for each user, and serve those results from an endpoint. The endpoint is essentially a dictionary lookup.

ahayamb commented 6 years ago

Thanks for the answers. I'm sorry for not stating my point clearer before. I've used lightfm to do prediction and it is fast enough. I have approximately 3gig model and serve it using uwsgi since some of python web frameworks doesn't support multi threaded request. If I spawn 4 uwsgi processes, it will consume approx. 12gig of RAM. I've tried to take out model's parameter and put it to postgres, but prediction time is increasing. Is there any strategies to overcome memory issue but still getting an acceptable prediction time? I'm sorry this is so demanding 😂

RothNRK commented 6 years ago

Yeah I missed that. I wouldn't know how to address that but would be interested to see what you come up with.

maciejkula commented 6 years ago

@ahayamb you can get the get_user/item_representations methods to dump user and item embeddings and biases as numpy arrays. Once you have those the score for (user i, item j) is given by

score = np.dot(user_embeddings[i], item_embeddings[j]) + item_biases[j]

Since these are plain numpy arrays, you can keep them on disk and memmap them into the memory of each of your individual processes.

ahayamb commented 6 years ago

Since these are plain numpy arrays, you can keep them on disk and memmap them into the memory of each of your individual processes.

Great!!! Sounds interesting, I will try. Thanks for the suggestion 🎊