Introduction

This repository includes the source code of my submitted Acoustic Scene Classification (ASC) system to Task 1A of the DCASE challenge 2019. The code is based on Python 3.5 and uses PyTorch 1.0.0.

The proposed system is based on an AlexNet-like model with stratified log-Mel features. "stratify" means that a given log-MEL image is decomposed as the combination of a number of component images, which correspond to sound patterns of different nature. Then each component image is modeled independently by a portion of convolution kernels in the CNN model.

If you would like to know more details about my ASC system, you can read my technical report here.

How to Use

There are two steps to run the system. First thing is to do audio feature extraction, i.e., extract log-Mel feature for each audio signal and decompose it into 3 component images. Then, we train and test the CNN model based on the component images.

Pre-requisite:

The code is based on Python 3.5 and uses PyTorch 1.0.0. The libraries' versions for running the code are listed below. However, the code should be able to run with libraries of newer versions.

• numpy.version=='1.14.0'
• soundfile.version=='0.9.0'
• yaml.version=='3.12'
• cv2.version=='3.4.2'
• scipy.version=='1.0.0'
• imageio.version=='2.5.0'
• pickle.version=='$Revision: 72223 $'
• sklearn.version=='0.18.2'
• matplotlib.version=='2.0.2'

I ran the code on a computer with 128GB RAM. In my code, I simply load the entire dataset into the memory. If your RAM size is small (e.g. 16GB), maybe you will need to optimize the feature loading step, i.e., loading data batch by batch.

Audio Feature Extraction

To do feature extraction, use the script "extr-asc.py" in the folder "feat_extract_asc". Before running the script, set the paths of the raw dataset and the output folders. 'raw_data_folder' is the path to audios in development dataset. 'output_feature_folder' is where extracted features are stored. 'spectrograms_folder' includes the feature images for visualization purpose only.

config = { ...
	'raw_data_folder': '.../TAU-urban-acoustic-scenes-2019-development/audio',
	'output_feature_folder': '.../features/development/logmel-128-S',
	'spectrograms_folder': '.../features/development/logmel-128-S-imgs',
	...
	}

Then run the script by

python extr-asc.py

Training and Testing

To train and test the CNN model with the officially provided setup, use the script "main.py" in "cnn_for_asc" folder. Before running the script, modify the following variables:

gpu = 0 # specify which gpu is used for training and testing.
development_data_path = '.../features/development/logmel-128-S' # path to feature folder.

Then run the script by

python main.py

After the training and testing are completed, a result folder named "results-logmel128S-AlexNetS-Mixup-20eps" is generated. Check for model accuracy, confusion matrix and learning curve there. A single model that I trained with the scripts has an accuracy of 76.6% on development dataset, and the confusion matrix is as follows: