The E-PANNs [1, 2] {~92MB, 24M parameters} are obtained from the original PANNs (CNN14) {~312MB, 81M parameters)[3].
A sample video of demo can be viewed at: (https://youtu.be/HiZw0pGXGQk)
[1] Arshdeep Singh, Haohe Liu and Mark D PLumbley, "E-PANNS: Sound Recognition using Efficient Pre-Trained Audio Neural Networks", accepted in Internoise 2023.
[2] Singh, Arshdeep, and Mark D. Plumbley. "Efficient CNNs via Passive Filter Pruning." arXiv preprint arXiv:2304.02319 (2023).
[3] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, Mark D. Plumbley. "PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition." arXiv preprint arXiv:1912.10211 (2019).
and the current demonstration shown in above Figure is a modification on the top of the AI4S project (https://ai4s.surrey.ac.uk/) previous demo https://github.com/yinkalario/General-Purpose-Sound-Recognition-Demo built by Yin Cao and Andres Fernandez.
The prediction/confidence in the above Figure is obtained by applying the audio tagging system (E-PANNs) on consecutive short audio segments.
We recommend using Anaconda to install the dependencies as follows:
conda create -n panns python=3.7
conda activate panns
pip install -r requirements.txt
conda install -c anaconda pyaudio
A comprehensive list of working dependencies can be found in the full_dependencies.txt file.
Download the model into your preferred <model_location> from https://doi.org/10.5281/zenodo.7939403
<model_location> == <repo_root>/models
Assuming our command line is on <repo_root>, the panns environment is active and the model has been downloaded into <repo_root>/models, the following command should run the GUI with default parameters (tested on Ubuntu20.04):
python demo_tag.py
Note that the terminal will print model summary upon start. The syntax to alter them is the same as with MODEL_PATH, e.g. to change the number of classes displayed to 10, add TOP_K=10.
Convergence plot of E-PANNs when 25%, 50% and 75% filters are removed from the C7 to C13 layers of PANNS (CNN14).
[1] Arshdeep Singh, Haohe Liu and Mark D PLumbley, "E-PANNS: Sound Recognition using Efficient Pre-Trained Audio Neural Networks", accepted in Internoise 2023.
[2] Singh, Arshdeep, and Mark D. Plumbley. "Efficient CNNs via Passive Filter Pruning." arXiv preprint arXiv:2304.02319 (2023).
- https://research.google.com/audioset/dataset
- https://github.com/qiuqiangkong/audioset_tagging_cnn
- https://github.com/qiuqiangkong/panns_inference
- https://github.com/yinkalario/Sound-Event-Detection-AudioSet
This work was supported by Engineering and Physical Sciences Research Council (EPSRC) Grant EP/T019751/1 “AI for Sound (AI4S)”. Project link: https://ai4s.surrey.ac.uk/