The repository to show the way I trained https://www.kaggle.com/hidehisaarai1213/birdcall-resnet50-init-weights
- Python >= 3.6
- ffmpeg
- sox
- GPU (P100 / V100) is the best
pip install -r requirements.txt
- First you need to put dataset in
/input/birdsong-recognition
. The data structure should be the same as that of Kaggle dataset. - Run
make prepare
at the top of this repository. This command will perform resampling the dataset to 32kHz. Note that this will take a bit long (1 Hours on NVIDIA DGX Station). - Run
make train
at the top of this repository.