search

grig-guz / rst-coref

RST discourse parsing with coreference information.
MIT License
3 stars 3 forks source link

training, epochs and iteration #5

Open arantzadev opened 1 month ago

arantzadev commented 1 month ago

Hi I had a question concerning the training, is it normal that the there are about 6000 iterations for each epochs ?

I am only training on the rst-dt train set with an NVIDIA RTX A6000 and it takes about an hour for a single epoch.

Also why is the the current number of epochs set at 20 ?

Thank you

grig-guz commented 1 month ago

Yes, it is normal. There are ~300 trees but since we have a shift-reduce parser, each shift-reduce action is treated as an individual datapoint. So there are way more datapoints than trees. We tried 10 and 15 epochs but the validation scores didn't saturate. They didn't improve beyond 20.