Aarrghh xD
#14
by
fconti
- opened
I'm curious, what approach did you take? Most of the stuff I tried didn't outperform the baseline and my best performing model is a slight modification of the baseline only.
I fine-tuned the embedding model in the base-line without the classifier. I employed a triplet loss approach with online hard batch mining. This allowed me to utilize all the sequences available in the training set!
I am going to share my code soon!