birds_dnn

Feature extraction (frontend) and machine learning for classification of birds songs using Kaggle dataset.

Install

After using git clone, copy the folder with ogg files to a folder called train_audio, and locate train_audio in the same folder as src, etc.

For instance, if you cloned the repository to C:\github\birds_dnn\ and now have

C:\github\birds_dnn\src
C:\github\birds_dnn\txt

then, create the folder with your ogg files:

C:\github\birds_dnn\train_audio

From src folder, you can use:

 python .\listen_and_see.py --should_plot --features=stft --normalization_method=minmax

to listen to the files and also plot images with spectrogram or similar features.

One can find more information in the file processing_steps.docx located in folder documentation.

One may have more than a bird singing in one file. But we don't know how the label file describes it
In the Kaggle challenge, one can use metadata (example: lat / long)
About the submission file: for each row, we print the 182 values of the output softmax layer