yandex-cup-2022-1st-place-solution

1st place solution of ML Audio Content track of Yandex Cup 2022

UPD: links to the LB/competition will be posted soon D:

Hardware

GPU: 1x4090
CPU: Ryzen 9 5950x
RAM: 128Gb

Software

Python 3.8.10
CUDA: 11.8
torch build: torch==1.14.0.dev20221021+cu117
NVIDIA Driver Version: 520.56.06

Directory structure

├── data
├── ensemble
├── ensemble.py
├── ensemble.sh
├── exps
├── prepare_env.sh
├── requirements.txt
├── train_arcface.py
├── train.sh
└── utils

Data

Download and unzip data to ./data/ dir inside project root

Install reqs

bash prepare_env.sh

Train model

It takes approximately 15 mins on the mentioned machine to finish one fold train, which is alone enough to take first place with score of 0.505-0.512 on public LB.

bash train.sh

Inference

bash ensemble.sh

Approximate tensorboard logs

Note: validation is done on one fold out of 10.