search

galsang / BIMPM-pytorch

Re-implementation of BIMPM (Bilateral Multi-Perspective Matching for Natural Language Sentences, Zhiguo Wang et al.) on Pytorch.
106 stars 31 forks source link

Improving performance #8

Closed Arjunsankarlal closed 5 years ago

Arjunsankarlal commented 5 years ago

I have trained a model and I am able to get way better results than the previous method that I was using. But now I face performance issues. For some queries, the model takes more than 6 seconds to do prediction. Any heads up to tuning the prediction time would be helpful! :)

Albert-Ma commented 5 years ago

What's the param you used for QQP and SNLI

Arjunsankarlal commented 5 years ago

The default params itself works well, I reduced the batch size to 32, since 64 was causing memory issues while training. Well, instead of using Glove embeddings, try using FastText Crawl embeddings, they give 0.5% increase in accuracy.