twcmchang / CP-CNN

Channel-Prioritized Convolutional Neural Networks for Sparsity and Multi-fidelity

Channel-Prioritized Convolutional Neural Network with Sparsity and Multi-fidelity

This repository contains the code to reproduce the core results from the paper Channel-Prioritized Convolutional Neural Networks with Sparsity and Multi-fidelity in the review process.

Dependencies

This work uses Python 3.6.0. Before running the code, you have to install

• tensorflow==1.4.0
• numpy==1.13.0
• pandas==0.20.3
• Pillow==4.3.0
• progress==1.3

The above dependencies can be installed using pip by running

pip install -r requirement.txt

Usage

To have a quick start on the experiment of CIFAR-10 by running

bash quick_start.sh <GPU_ID>

Training stage for channel prioritization and network sparsity

python3 train.py  --init_from <pre-trained_net_params.npy> --save_dir <save_directory> --tesla 0 --keep_prob 1  --lambda_s 0.001 --lambda_m 0.001 --decay 0.00005 --prof_type linear

Fine-tuning stage (set tesla=1) for loss aggregation

python3 train.py --init_from <pruned_net_params.npy> --save_dir <save_directory> --tesla 1 --keep_prob 1  --prof_type linear

Testing multi-fidelity inference

python3 test.py --init_from <finetuned_net_params.npy> --output <output_file> --keep_prob 1

To run the experiment on CIFAR-100, please add another arguement --dataset CIFAR-100 in above commands.

If you have questions, please write an email to [e-mail].