Birds-Recognition

Download mp3 files, data processing and training functions for birds classification, using pretrained CNN models.

Download audiofiles and process data

Download the metadata of the birds in your BirdsList (Choose your bird species here)

python create_birds_dataset.py

Download the audio files (mp3)

python download_birds_audio.py

Process the audio (convert mp3 to wav), resample audios and create training labels csv

python mp3towav.py

(! The data_processing.py file is for data processing just before training !)

Download pretrained models

Download your model from https://zenodo.org/record/3576403#.XqsOuyPVJPY and put it in the pretrained_models directory.

Launch training (Transfer Learning)

Example: ResNet22, 75% of the dataset

python main.py --model_path="pretrained_models/ResNet22_mAP=0.430.pth" --trained_models_path="models" --graphs_path="graphs" --saving=1 --model_type="Transfer_ResNet22" --lr=1e-4 --batch_size=32 --epochs=25 --frac_data=0.75

Example: Cnn6, 100% of the dataset

python main.py --model_path="pretrained_models/Cnn6_mAP=0.343.pth" --trained_models_path="models" --graphs_path="graphs" --saving=1 --model_type="Transfer_Cnn6" --lr=1e-3 --batch_size=16 --epochs=25

--model_path: Path to your pretrained model

--saving: 0 = do not save; 1 = save training graph and trained model to .pth

--model_type: The class name of your Pytorch model in transflearn_models.py

--frac_data: Fraction of the dataset for training (0 <= frac_data =< 1, default = 1)

Inference on the test set and show stats with a Confusion Matrix

Example: Plot CM for trained Transfer_ResNet22

python plot_cm.py --trained_model_path="models/Transfer_ResNet22_trained_testAcc=0.6330.pth" --graphs_path="graphs" --saving=1 --model_type="Transfer_ResNet22"

Credits

Author: Richard Huang, under the guidance of Nicolas Farrugia

Dataset retrieved from https://www.xeno-canto.org

Download function inspired by AgaMiko

Code for Audioset Tagging CNN from Qiu Qiang Kong