kentonl / e2e-coref

End-to-end Neural Coreference Resolution
Apache License 2.0
524 stars 173 forks source link

Training specs #14

Closed amttar closed 6 years ago

amttar commented 6 years ago

You mentioned that training took you 48 hours to complete 400K steps to convergence, but I can't find the specs of your training machine. Was that on GPU or CPU? How big? I am training your model on Arabic CoNLL and it took me about 11 hours to complete 7600 steps only. A simple math shows that to reach 400K steps following this rate, I would need 578 hours! I am running on a machine that has 8 vcpus, and I wanted to check with you before questioning my training.

amttar commented 6 years ago

I tried training on GPUs and I managed to get it to 400K in about 48 hours; so yeah, I assume it was trained on GPU.