Spaces:

competitions

/

movie-genre-prediction

like 75

Paused

Hi All,
First of all i like to thanks @huggingface-co and Data driven science for hosting an amazing competition where we can really learn from different people across the community and grow . So thanks a lot for that , and also a big thanks to @abhishek .

The data was super clean and I have concat the movie_name + synopsis , as the only synopsis based model was not helping in generating the Best LB score :D and used stratified 5 fold for the prediction i used the idxmax function to match with the corresponding genre. My solution is super naaive :)

Coming on to my solution :
Things that worked:

1. Deberta V3 large + AWP + Mean Pooling+ Cross Entropy Loss : 512 max length
2. Deberta Large + AWP + Mean Pooling+ Cross Entropy Loss : 512 max length
3. Deberta Xlarge + AWP + Mean Pooling+Cross Entropy loss : 512 max length

and Average of these models result in public LB : 0.44x and private LB: .43x

What don't worked:

1. GPT2 medium model , for variety i trained GPT2 medium model but it worsen the score.
2. Even tried with pooled output for the forward pass , didn't help much.
3. Prepoccesing of the text even worsen the score

I used A6000 for all the training and I'll share the clean code soon.

Thanks once again and Happy Learning

Regards,
Aman Kapoor