GAN-based Projector for Faster Recovery with Convergence Guarantees in Linear Inverse Problems.

This repository provides code for this paper: GAN-based Projector for Faster Recovery with Convergence Guarantees in Linear Inverse Problems.

Requirements

Linux system, a GPU

1. Python 3.4
2. Pytorch 1.2.0
3. numpy
4. matplotlib
5. imageio
6. scikit-image
7. scikit-learn
8. requests
9. torchvision

Preliminaries

1. Create a conda virtual environment and install packages

   conda create --name NPGD python=3.4
   conda activate NPGD
   conda install pytorch=1.2.0 imageio=2.6.1 matplotlib=3.1.1 numpy=1.17.2 scikit-image=0.15.0 scikit-learn=0.21.3 requests=2.22.0 torchvision=0.4.0 

2. Clone this repo

   git clone https://github.com/yuqili3/NPGD_linear_inverse_prob.git
   cd NPGD_linear_inverse_prob

3. Download the dataset as well as the test images, we need to crop and resize the faces from the original images.

   unzip celebAtest.zip -d data/celebA/testing_images
   unzip mnisttest.zip -d data/mnist/
   python download_data.py celebA
   rm -rf ./data/celebA/celebA_data_raw
   rm ./data/celebA/celebA_data.zip

   this may take several minutes to download the large file

4. Train the network projector models.

   1. for mnist dataset, train a DCGAN first and then a concatenated network projector:

      python MNIST_DCGAN.py --k=100 --epochs=40 --lr=1e-4
      python MNIST_invDCGAN_multi_task_loss.py --k=100 --epochs=100 --lr=1e-4 --lamda=0.1 --sigma=1

   2. for celebA dataset:

      python CELEBA_DCGAN.py --k=100 --epochs=40 --lr=1.5e-4
      python CELEBA_invDCGAN_multi_task_loss.py --k=100 --epochs=100 --lr=1e-4 --lamda=0.1 --sigma=1

Demos

1. Compressed sensing using random Gaussian matrix

   • mnist

     python MNIST_projection_using_inv_gen.py --k=100 --task=cs --m=100 --designed=0  --lr=1 --ite=30 --info=1

   • celebA

     python CELEBA_projection_using_inv_gen.py --k=100 --task=cs --m=1000 --designed=0  --lr=0.5 --ite=30 --info=1

2. Image inpainting

   • mnist

     python MNIST_projection_using_inv_gen.py --k=100 --task=ip --mask_size=8 --lr=2 --ite=30 --info=1

   • celebA

     python CELEBA_projection_using_inv_gen.py --k=100 --task=ip --mask_size=32 --lr=1 --ite=30 --info=1

3. Image super resolution

   • mnist

     python MNIST_projection_using_inv_gen.py --k=100 --task=sr --sr_ratio=2 --lr=6 --ite=30 --info=1

   • celebA

     python CELEBA_projection_using_inv_gen.py --k=100 --task=sr --sr_ratio=2 --lr=6 --ite=30 --info=1

4. Use a designed matrix in compressed sensing

   • mnist

     python design_measurement_matrix.py --dataset=MNIST --k=100
     python MNIST_projection_using_inv_gen.py --k=100 --task=cs --m=200 --designed=1 --lr=1 --ite=30 --info=1

   • celebA

     python design_measurement_matrix.py --dataset=CELEBA --k=100
     python CELEBA_projection_using_inv_gen.py --k=100 --task=cs --m=1000 --designed=1 --lr=1 --ite=30 --info=1