DSMN_dcase2023

KO | EN

gitlite — search

#ai #llm #golang #android #machine-learning #rust #docker #windows #awesome #cli #kubernetes #vue #mcp #ios

(★ 25)

[DCASE 2023] Official Implementation for "Low-Complexity Acoustic Scene Classification Using Deep Space Separable Distillation And Mutil-Task Learning"

DSMN_dcase2023 Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# DSMN_dcase2023
![name](https://img.shields.io/badge/dsmn-v0.1.1-brightgreen)
![license](https://img.shields.io/badge/license-MIT-blue)
![ipython](https://img.shields.io/badge/iPython-v8.4.0-orange)
![build](https://img.shields.io/badge/build-passing-yellowgreen)
## Some Related Works
**2023-06**: Low-Complexity Acoustic Scene Classification Using Deep Space Separable Distillation And Mutil-Task Learning [(Paper Link)](https://dcase.community/documents/challenge2023/technical_reports/DCASE2023_Wang_61_t1.pdf)  
**2024-05**: Deep Space Separable Distillation for Lightweight Acoustic Scene Classification  [(Paper Link)](https://arxiv.org/abs/2405.03567)
## Introduction 
### Abstract
This technical report describes our system for Task 1 in Detec-tion and Classification of Acoustic Scenes and Events (DCASE) 2023. We propose a deep space separable distillation block as the basic unit of the model, using its strong block processing ability to continuously cut the high-frequency and low-frequency parts of the log-Mel spectrogram. The accuracy is improved by multi-scale embedding and multi-task learning methods. To prevent overfitting, we adopt data augmentation methods such as mixing, speculation and spectral modulation. Quantization aware training is adopted to quantize the model to meet the requirements of edge devices with low complexity constraints. The proposed system achieves a 53.3% accuracy on the development dataset with only a parameter count of 45.16 kB and  the MACs of 8.64M .
### Result
|| Parms | MACs | Acc(float32) | Acc(int8) |
|----|----|----|----|----|
|model1| 45.1k | 8.64M | 53.4 | 53.3 |
|model2| 56.1k | 16.74M | 59.2 | 56.4 |
|model3| 56.5k | 25.44M | 53.4 | 50.8 |
|model4| 121.8k | 20.92M | 54.7 | 52.4 |
 
## Requirements
The required library files are placed in requirements.txt  
our environment: RTX3080 + cuda11.3 + torch1.11
## Usage
1、Clone project
```
git clone https://github.com/Wangkkklll/DSMN_dcase2023.git
```
2、Install the required library files
```
pip install -r requirements
```
3、Dataset  
Please place dataset under the path /DSMN_dcase2023/data  
If your data is in another path, please create soft link
```
ln -s /your_path/ /root/DSMN_dcase2023/data
```
4、Nessi clone  
This library is used to calculate the number of statistical model parameters.  
```
git clone https://github.com/AlbertoAncilotto/NeSsi.git
```
5、How to train  
Modify the imported model to obtain a different model:from models.model1 import Cnn (**line313**)  
If you are running the code for the first time, change **True** in **lines 220 and 252** of the **train.py** file to **False** to preprocess the data. After the data is processed, you can keep it as True when running the code later.
```
python train.py
```
## Contact us
kangliwang@stu.pku.edu.cn

## Bibtex
If you think this work will help you, please refer to the following documents
```
@techreport{Wang2023,
    Author = "Wang, Kangli and Wu, Yiling and Li, Yanxiong",
    title = "Low-Complexity Acoustic Scene Classification Using Deep Space Separable Distillation Module and Multi-Label Learning",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "This technical report describes our system for Task 1 in Detection and Classification of Acoustic Scenes and Events (DCASE) 2023. We propose a deep space separable distillation block as the basic unit of the model, using its strong block processing ability to continuously cut the high-frequency and low-frequency parts of the log-Mel spectrogram. The accuracy is improved by multi-scale embedding and multi-task learning methods. To prevent overfitting, we adopt data augmentation methods such as mixing, speculation and spectral modulation. Quantization aware training is adopted to quantize the model to meet the requirements of edge devices with low complexity constraints. The proposed system achieves a 53.3\% accuracy on the development dataset with only a parameter count of 45.16 kB and the MACs of 8.64 M. "
}
```

## License
MIT

DSMN_dcase2023

DSMN_dcase2023 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

DSMN_dcase2023

DSMN_dcase2023 Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README