Repository contains the code for 1st place solution of BirdClef2022 competition. The goal of the competition was to identify bird calls from given soundspaces.
reproduce
run
sh 5folds.sh
train
sh train_single.sh 0 ./outs cls_nf0_v3 0 /kaggle/input nf0_v3
respectively, 0 number of gpus, ./outs validation output directory, cls_nf0_v3 config , 0 fold number, /kaggle/input/ is root data directory that contains birdclef-2021 and birdclef-2022 data.
inputs
Download noise_30sec, ff1010bird_nocall, train_soundscapes folders from here.
manually chunked data
maupar
- Download audio clips from here.
- Move audio files under birdclef-2022/train_audio/maupar
- Use folds.csv
references
Parts of solution taken from