1st place solution in Yandex Cup 2023 - ML RecSys (audio classification).

Make final submission (inference of each model + blending):

1. pip install -r requirements.txt
2. Download and unpack embeddings
3. Download and unpack models. Structure is exactly the same as in models, but files pytorch_model.bin are in checkpoint-* directories.
4. Modify models_dir, embeddings_dir, blended_output_path (optional, ensemble output), ensemble_path (optional, ensemble config path: choose between two files in ensembles directory) in bin/predict_ensemble.sh
5. Run inference:

   cd bin
   bash predict_ensemble.sh

I have loaded two best submissions:

• ensembles/ensemble_v30.tsv - public score: 0.3095, private score: 0.3128.
• ensembles/ensemble_v31.tsv - public score: 0.3087, private score: 0.3120. Was selected as final.

Re-train whole ensemble (train each model on 3 splits of 10-folds kf):

cd bin
bash re_train_ensemble.sh

Code for models selection: src/beam_search.py