You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have an engine I created with ./mmt create using an old dataset I had, and I would like to retrain it again using a new dataset for more fine-tuning.
I've tried ./mmt train --from-model but it messes up the BLEU Score.
I also tried creating a new engine from scratch with ./mmt create to train it on both datasets together, but I got an exit code -11 while training, i'm assuming the jvm ran out of memory at some point (it was using a lot while I was monitoring memory usage).
I've seen from other issues people suggesting ./mmt memory import -p parallel_src parallel_tgt -e model_name, but i'm not sure how much this impacts the performance or translation speed.
Any tips or suggestions are appreciated, thank you.
The text was updated successfully, but these errors were encountered:
I have an engine I created with
./mmt create
using an old dataset I had, and I would like to retrain it again using a new dataset for more fine-tuning.I've tried
./mmt train --from-model
but it messes up the BLEU Score.I also tried creating a new engine from scratch with
./mmt create
to train it on both datasets together, but I got anexit code -11
while training, i'm assuming the jvm ran out of memory at some point (it was using a lot while I was monitoring memory usage).I've seen from other issues people suggesting
./mmt memory import -p parallel_src parallel_tgt -e model_name
, but i'm not sure how much this impacts the performance or translation speed.Any tips or suggestions are appreciated, thank you.
The text was updated successfully, but these errors were encountered: