Entry to the Bird Audio Detection challenge

Based on Densely Connected Convolutional Networks (DenseNets) Theano/Lasagne

About the BAD challenge

To get information about the challenge, please visit its Website. The purpose was to propose a solution to the binary classification task of detecting singing birds in 10-s duration audio files.

Requirements

• Theano (0.9.0.dev3)
• Lasagne (0.2.dev1)
• h5py (2.6.0)
• scikit-learn (0.17)
• fuel (0.2.0)
• MIR toolbox for feature extraction with Matlab

Usage

1- Feature extraction

Features: 56 log-Mel F-BANK coefficients, 58 bands, hop size: 50 ms, frame size: 100 ms, fmin: 50 Hz, fmax: 22050 Hz

• MIR_extract_logSpectrumBands.m: extracts F-BANK coefficients from WAV files
• create_hdf5_ff1010bird_public.py: creates an HDF5 file with Train, Valid and Test subsets

2- To train a densenet for 30 epochs:

python train.py densenet 30

3- To test a model:

python test.py densenet <model path> <HDF5 datafile path> <CSV id file path>

Example (the HDF5 data file is not provided in this repo):

python test.py densenet models/bad16_densenet_bn_static-fbank-0.019326000-sub4.npz hdf5/bad2016test_melLogSpec56.hdf5 hdf5/Test_files.csv

Model Architecture

The code builds the following model. It is based on this recipe.

input layer: (None, 1, 200, 56)
first conv layer: (None, 32, 200, 56)
dense block 0: (None, 107, 200, 56)
transition 0: (None, 107, 100, 28)
dense block 1: (None, 182, 100, 28)
transition 1: (None, 182, 50, 14)
dense block 2: (None, 257, 50, 14)
post Global pool layer: (None, 257)
output layer: (None, 2)

total number of layers: 74
number of parameters in model: 328004

Each dense block corresponds to 5x[BatchNorm - ReLu - Conv3x3]
Each transition block corresponds to 1x[Conv1x1 - Max-Pool2x2]

Saliency maps

To generate saliency maps with guided backprop (based on this recipe):

python saliency_maps.py densenet <modelpath>

Example:

For any questions, please email me: thomas.pellegrini@irit.fr

If you use this code, please consider citing my paper:

T. Pellegrini. Densely Connected CNNs for Bird Audio Detection. In Proc. European Signal and Image Processing Conference (EUSIPCO 2017), EURASIP, pp. 1734-1738, September 2017, Kos