Pruning_MTCNN_MobileFaceNet_Pytorch

This repo demonstrates how to prune MTCNN and MobileFaceNet using Pytorch implementation. The whole model size can be reduced by x2.5 with x1.7 speedup and negligible accuracy loss. The results are shown as the below table. The base model and pruned model are provided in this repo.

Pruning Criteria

The pruning criteria is based on the paper Pruning Convolutional Neural Networks for Resource Efficient Inference. Please go through the "Pruning MTCNN Tutorial" and "Pruning MobileFaceNet Tutorial" for detailed information.

How to Use

Prepare the training dataset

put the dataset into folder "data_set"

1. MobileFaceNet: faces_emore, LFW, CFP-FP and AgeDB-30 from Model Zoo
2. MTCNN:

• WIDER FACE, CNN_FacePoint,
• prepare 12x12, 24x24, 48x48 images respectively,
• put the annotation files into "data_preprocessing/anno_store"
• Refer to MTCNN_Tutorial for the data preparation

Pruning

• MTCNN
• python MTCNN/Pnet_Prune.py
• python MTCNN/Onet_Prune.py
• python MTCNN/Rnet_Prune.py
• MobileFaceNet
• python MFN/MFN_Prune.py

Reference

• pytorch-pruning

Please give me a star if it is helpful for your research