This is a repo of our following paper,
Singh, Arshdeep, and Mark D. Plumbley. "Efficient Similarity-Based Passive Filter Pruning for Compressing CNNS." ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023.
This is a repository containing scripts to prune CNNs using similarity-based pruning algorithm in a fast manner.
-
Efficient_Similarity_Pruning.py: To compute indexes of important filters for a given convolution layer in an efficient way by approximating the distance matrix or using complete distance matrix. The input to the script is the pre-trained weights.
-
Plot_generation.py : To generate Figures in the result and analysis section.
-
Fine_tuning_DCASE21_Net.py: Given important filter indexes for all layers, Obtaining a pruned network and peforming fine-tuning.
- important_filter_indexes for VGGish_Net and DCASE21_Net: The important set of filters obtained using "Efficient_Similarity_Pruning.py", l1-norm based method and GM pruning method..
-
DCASE2018 dataset and VGGish_Net baseline (pre-trained weights, model with 64.69% accuracy): https://drive.google.com/drive/folders/1b-eOYzNm2-IjTLf6jeaGr9twpd79LKFm?usp=sharing
-
DCASE2021 dataset and DCASE21_Net baseline (pre-trained weights, model with 48.58% accuracy): https://drive.google.com/drive/folders/1b-eOYzNm2-IjTLf6jeaGr9twpd79LKFm?usp=sharing
if above links does not work, please see below link