ffeng1017/power-law

Recommend install in virtual environment

$ conda create -n yourenvname python=2.7 anaconda

Activate the enviorment

conda activate yourenvname

Install the required the packages inside the virtual environment

sh installation.sh

For experiments in section 3.1

bash expt1/fc/run.sh

For CNN on MNIST experiment in section 3.2

bash expt1/cnn/runMNIST.sh

For AlexNet on ImageNet experiment in section 3.2

bash expt1/cnn_large/runMNIST.sh

For VGG-16 on ImageNet experiment in section 3.2

bash expt1/cnn/runCIFAR.sh

For experiments in section 4.2

bash expt2/fc/run.sh

For experiments in section 5.1

bash expt3/fc/run_pathnet.sh

For mlp on mnist in section 5.2

bash expt3/fc/run_pnn.sh

For cnn on cifar in section 5.2

bash expt3/cnn/run_cifar_pnn.sh

• MNIST is inside this folder
• CIFAR dataset need to be prepared using pylearn2 following https://github.com/MatthieuCourbariaux/BinaryConnect

• First download CIFAR dataset and put the data in expt3/cnn/cifar10(100)
• Run the scripts to build up pickle input format (need minor changes on the path):

cd pylearn2/pylearn2/datasets
python cifar10.py
python cifar100.py
cd power-law/pylearn2/pylearn2/scripts/datasets
python make_cifar10_gcn_whitened.py
python make_cifar100_gcn_whitened.py

• Or you can directly download from this Google drive link.

• Due to github file size limitation, models in VGG and AlexNet on ImageNet in section 3.2 cannot upload in the repo. You can download on this Google drive link. Please place the three files under:

expt1/cnn_large

• For GPU support during training, please refer to this tutorial.
• Please make sure the configuration of environmental variables has been done under root path.