infiniteemo / blef

Repository contains the code for 1st place solution of BirdClef2022 competition. The goal of the competition was to identify bird calls from given soundspaces.

• Solution post
• Inference notebook

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

• https://github.com/selimsef/xview3_solution
• https://github.com/ChristofHenkel/kaggle-birdclef2021-2nd-place
• https://github.com/ChristofHenkel/kaggle-landmark-2021-1st-place