Interpretable Customized Neural Network Pruning by Identifying Critical Data Routing Path

This repository contains：

• The Final Report: Interpretable Customized Neural Network Pruning by Identifying Critical Data Routing Path.pdf
• The Report Latex Source Files: LatexSource
• Our Experiment Codes: SourceCode

We implement our pruning algorithm based on the TensorFlow 1.4.1 with CUDA 8.0. We use CIFAR-100 dataset and VGG-16 network for all the experiments. Codes are available at Github Link

To run our code, you have to download:

• CIFAR-100 Dataset： Assuming the code is put in directory ".", please download the dataset and save it in the directory "./cifar-100-python".

• Pretrained VGG-16 Model：Assuming the code is put in directory ".", please download the following three files and save it in the directory "./vggNet".

  • https://drive.google.com/open?id=1fDZDf7UpsVCn4CGGvI-Jssm3iFQgpyJw
  • https://drive.google.com/open?id=1ZJm8-6HIDOLBWXBt92MQjUKqWRbq_xpz
  • https://drive.google.com/open?id=1nXcmco9zrJIkOTQTTa4V3_VbTVajExWI

  Credit to BoyuanFeng, Github site: https://github.com/BoyuanFeng

And you need to install following python pachages:

• pickle
• json
• keras
• numpy
• tensorflow
• sklearn

We suggest you to install Anaconda for convenience

Therre are several steps to run the code:

1. run.py: this file generate class encodes. You should change classes or set loops to run all classes in this file
2. trim_and_test.py: this file trim the model and test the accuracy with pruned model which has not been fine tuned yet. Change target_class_id for models of different classes
3. run_finetune.py: this file fine tune the pruned model and test the accuracy with fine tuned models. You should change target_class_id for models of different classes

Notice: we use GPU for training, so you should designate certain GPU for training in these files.