This repo contains an unofficial MATLAB implementation of DCASE2021 Task 1A baseline code, which is part of the DCASE 2021 challenge.
Copyright 2021 The MathWorks, Inc.
MathWorks Products (http://www.mathworks.com)
Requires MATLAB release R2021a or newer. To train the baseline system, the following toolboxes are required:
To accelerate training, the following toolbox is recommended:
To quantize a network, the following package is required:
To deploy a quantized network to CUDA code, the following toolbox is required:
See Quantization Workflow Prerequisites for a list of required products depending on your target.
This unofficial baseline has the following known differences with the official baseline. There may be additional differences.
- The mini-batch size was increased from 16 to 256.
- A piecewise learn rate schedule was added with a drop period of 100 epochs. The max number of epochs was reduced from 200 to 120.
- This example uses and evaluates the final state of the network, after all epochs are complete. The official baseline uses the best peforming model over all of the epochs.
- This example only trains the network and evalutes the system once. The official baseline trains and evaluates 10 times to provide additional statistical analysis.
- This example uses int8 quantization instead of the half-precision quantization in the official baseline.
To run this baseline, add Unofficial_DCASE2021_Task1A_Baseline_Using_MATLAB.mlx and classifyAcousticScene.m to your current folder in MATLAB and then run Unofficial_DCASE2021_Task1A_Baseline_Using_MATLAB.mlx. The example loads and examines the data, defines and trains a model, quantizes the model, and evaluates the quantized model.
You can view a PDF of the executed example in the file Unofficial_DCASE2021_Task1A_Baseline_Using_MATLAB.pdf.
The license is available in the License.txt file in this repository.
[1] Irene Martin-Morato, Toni Heittola, Annamaria Mesaros, and Tuomas Virtanen. Low-Complexity Acoustic Scene Classification for Multi-Device Audio: Analysis of DCASE 2021 Challenge Systems. 2021. URL: https://arxiv.org/abs/2105.13734.