Compressed Sensing with Deep Image Prior

May 22, 2019 ยท View on GitHub

This repository provides code to reproduce results from the paper: Compressed Sensing with Deep Image Prior and Learned Regularization.

Here are a few example results:

MNIST at 75 measurementsX-ray at 2000 measurements
mnist_reconstrmnist_reconstr

Preliminaries

  1. Clone the repository

    $ git clone https://github.com/davevanveen/compsensing_dip.git
$ cd compsensing_dip

    Please run all commands from the root directory of the repository, i.e from compsensing_dip/

  2. Install requirements

    $ pip install -r requirements.txt

Plotting reconstructions with existing data

  1. Open jupyter notebook of plots

    $ jupyter notebook plot.ipynb

  2. Set variables in the second cell according to interest, e.g. DATASET, NUM_MEASUREMENTS_LIST, ALG_LIST. Existing supported data is described in the comments.

  3. Execute cells to view output.

Generating new reconstructions on the MNIST, xray, or retinopathy datasets

  1. Execute the baseline command

    $ python comp_sensing.py

    which will run experiments with the default parameters specified in configs.json

  2. To generate reconstruction data according to user-specified parameters, add command line arguments according to those available in parser.py. Example:

    $ python comp_sensing.py --DATASET xray --NUM_MEASUREMENTS 2000 4000 8000 --ALG csdip dct

Running CS-DIP on a new dataset

  1. Create a new directory /data/dataset_name/sub/ which contains your images
  2. In utils.py, create a new DCGAN architecture. This will be similar to the pre-defined architectures, e.g. DCGAN_XRAY, but must have output dimension equal to the size of your new images. Output dimension can be changed by adjusting kernel_size, stride, and padding as discussed in the torch.nn documentation.
  3. Update configs.json to set parameters for your dataset. Update utils.init_dcgan to import/initiate the corresponding DCGAN.
  4. Generate and plot reconstructions according to instructions above.

Note: We recommend experimenting with the DCGAN architecture and dataset parameters to obtain the best possible reconstructions.

Generating learned regularization parameters for a new dataset

The purpose of this section is to generate a new (\mu, \Sigma) based on layer-wise weights of the DCGAN. This functionality will be added soon.