arpoyda/BirdCLEF_2024

BirdCLEF 2024 1st place solution

PythonApache-2.0

Below you can find an outline of how to reproduce our solution for the BirdCLEF 2024 competition.

Hardware/Software:

Kaggle kernel with Tesla P100.

Python packages are detailed separately in requirements.txt

Data Setup

Competition data can be downloaded from here. Put it in the ./data/ directory.

Model Build

There are two options to reproduce the solution.

With pretrained models.

a) Download models checkpoints from here and put in models_weights/.

b) Run python run_inference.py from the top directory.

Submission file will be in tmp/ directory.
Retrain from sratch.

Run the following scripts from the top directory.

a) Download models checkpoints (fold0.ckpt, fold1.ckpt, fold3.ckpt) from here and put in models_weights/. This models are trained only on train_audio and used for pseudo labeling of unlabeled_soundscapes.

a) Create mels and pseudo labeles for data/train_audio:
```
`python run_train_preprocessing.py`
```
b) Create mels and pseudo labeles for data/unlabeled_soundscapes:
```
`python run_unlabeled_preprocessing.py` 
```
c) Train 6 models (for each config from configs/ in a similar way):
```
`python run_training.py --cfg configs/effnet_seg20_80low.json`
```
d) Make prediction for data/test_soundscapes:
```
`python run_inference.py`
```
Submission file will be in tmp/ directory.

Inference on Kaggle

Inference is published in a Kaggle kernel here. Weights from our trained models are provided in a kaggle dataset linked to the inference kernel here.

Also see our write-up for more details.