dazzle-me / birdclef-2022-3rd-place-solution

CNN part of 3rd place solution of BirdCLEF 2022 challenge

birdclef-2022-3rd-place-solution

This repository contains the CNN part of the third place solution of BirdCLEF 2022 challenge.

The overview of our solution can be found here on Kaggle forum.

Hardware

Deep Learning Server

CPU : Intel(R) Core(TM) i9-10940X CPU @ 3.30GHz

GPU : 1x3090

RAM : 64Gb

Software

Python 3.8.10

CUDA: 11.6

cudnn version: 8.0.5.39

NVIDIA Driver Version: 510.73.05

Data

Aside from competition data, the solution used following external data:

1. freefield1010 no-call dataset
2. aircrowd_2020 no-call dataset
3. no-call parts of 2021 training data

All of these datasets were kindly provided by the last year 2nd place participants here

Download all data above and place it into ./input/ dir.

Setup env

cd docker
docker build . -t bclf
cd ..

Launch env

Inside "launch_docker.sh"

docker run --gpus '"device=2"'  --name=em2 \
				 -u $(id -u):$(id -g) \
				 --shm-size 64G \
				 --log-driver=none \
				 --rm -v /scratch/eduard.martynov/datasets/:/workspace/datasets \
				 -v /scratch/eduard.martynov/bclf_4/:/workspace/bclf \
				 -it bclf

you should change

• --gpus: device id you want to use, for example 0

change the highlighted part to where competition dataset is stored

• -v /scratch/eduard.martynov/datasets/:/workspace/datasets

change the highlighted part to this directory

• -v /scratch/eduard.martynov/bclf_4/:/workspace/bclf

bash launch_docker.sh

How to train models

This scripts goes into saved sources from 'exp-dir', trains the models and copies them into 'save-dir'

bash launch_docker.sh
cd ./bclf/
python3 walk_and_train.py --save-dir /workspace/bclf/reproduced_experiments \
                          --exp-dir ./experiments

How to do inference

arguments:

• inference-dir: test dir
• experiment-dir: dir with all experiments stored
• exp: experiment name
• fold: fold used as validation during training
• thresholds: per-bird threshold values
• num-workers: number of workers used in DataLoader
• batch_size: batch size to speed-up inference
• cfg: always 'base_config.yaml' except for 'new-metrics-exp-23', for this experiment pass 'tf_effnet_v2_s_in21k.yaml'

bash launch_docker.sh
cd ./bclf/
python3 inference_on_dir.py --experiment-dir ./reproduced_experiments \
                            --exp start_fold4_exp17_full \
                            --inference-dir /workspace/datasets/bird-clef/test_soundscapes/ \
                            --thresholds 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 \
                            --num-workers 16 \
                            --batch-size 32 \
                            --fold 4 \
                            --cfg base_config.yaml                            

References

• https://github.com/ChristofHenkel/kaggle-birdclef2021-2nd-place