Repository files navigation

This repository provides code to reproduce results from the paper:Compressed Sensing using Generative Models.

Here are a few example results:

Reconstruction	Super-resolution	Inpainting

Here we show the evolution of the reconstructed image for different number of iterations:

NOTE: Please runall commands from the root directory of the repository, i.e fromcsgm/

1. Python 2.7
2. Tensorflow 1.0.1
3. Scipy
4. PyPNG
5. (Optional : for lasso-wavelet)PyWavelets
6. (Optional)CVXOPT

Pip installation can be done by$ pip install -r requirements.txt

1. Clone the repository and dependencies

   $ git clone https://github.com/AshishBora/csgm.git$cd csgm$ git submodule update --init --recursive

2. Download/extract the datasets:

   $ ./setup/download_data.sh

3. Download/extract pretrained models or train your own!

   • To download pretrained models:$ ./setup/download_models.sh
   • To train your own
     • VAE on MNIST:$ ./setup/train_mnist_vae.sh
     • DCGAN on celebA, seehttps://github.com/carpedm20/DCGAN-tensorflow

4. To use wavelet based estimators, you need to create the basis matrix:

   $ python ./src/wavelet_basis.py

The following are the supported experiments and example commands to run the demos:

[Note: For celebA experiments, we run very few iterations per experiment (30) to give a quick demo. To get results with better quality, increase the number of iterations to at least 500 and use at least 2 random restarts.]

1. Reconstruction from Gaussian measurements
   • $ ./quick_scripts/mnist_reconstr.sh
   • $ ./quick_scripts/celebA_reconstr.sh "./images/182659.jpg"
2. Super-resolution
   • $ ./quick_scripts/mnist_superres.sh
   • $ ./quick_scripts/celebA_superres.sh "./images/182659.jpg"
3. Reconstruction for images in the span of the generator
   • $ ./quick_scripts/mnist_genspan.sh
   • $ ./quick_scripts/celebA_genspan.sh
4. Quantifying representation error
   • $ ./quick_scripts/mnist_projection.sh
   • $ ./quick_scripts/celebA_projection.sh "./images/182659.jpg"
5. Inpainting
   • $ ./quick_scripts/mnist_inpaint.sh
   • $ ./quick_scripts/celebA_inpaint.sh "./images/182659.jpg"

1. Create a scripts directory$ mkdir scripts

2. Identfy a dataset you would like to get the quantitative results on. Locate the file./quant_scripts/{dataset}_reconstr.sh.

3. ChangeBASE_SCRIPT insrc/create_scripts.py to be the same as given at the top of./quant_scripts/{dataset}_reconstr.sh.

4. Optionally, comment out the parts of./quant_scripts/{dataset}_reconstr.sh that you don't want to run.

5. Run./quant_scripts/{dataset}_reconstr.sh. This will create a bunch of.sh files in the./scripts/ directory, each one of them for a different parameter setting.

6. Start running these scripts.

   • You can run$ ./utils/run_sequentially.sh to run them one by one.
   • Alternatively use$ ./utils/run_all_by_number.sh to create screens and start proccessing them in parallel. [REQUIRES: gnu screen][WARNING: This may overwhelm the computer]. You can use$ ./utils/stop_all_by_number.sh to stop the running processes, and clear up the screens started this way.

7. Create a results directory :$ mkdir results. To get the plots, seesrc/metrics.ipynb. To get matrix of images (as in the paper), run$ python src/view_estimated_{dataset}.py.

8. You can also manually access the results saved by the scripts. These can be found in appropriately named directories inestimated/. Directory name conventions are defined inget_checkpoint_dir() insrc/utils.py

For a complete list of images not used while training on celebA, seehere.