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 measurements | X-ray at 2000 measurements |
---|---|
![]() |
![]() |
-
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/
-
Install requirements
$ pip install -r requirements.txt
-
Open jupyter notebook of plots
$ jupyter notebook plot.ipynb
-
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. -
Execute cells to view output.
-
Execute the baseline command
$ python comp_sensing.py
which will run experiments with the default parameters specified in
configs.json
-
To generate reconstruction data according to user-specified parameters, add command line arguments according to those available
parser.py
. Example:$ python comp_sensing.py --DATASET xray --NUM_MEASUREMENTS 2000 4000 8000 --ALG csdip dct
- Create a new directory
/data/dataset_name/sub/
which contains your images - 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. - Update
configs.json
to set parameters for your dataset. Updatecs_dip.py
to import/initiate the corresponding DCGAN. - 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.
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.