Sparse-Gen

This repository provides a reference implementation for learning Sparse-Gen models as described in the paper:

Modeling Sparse Deviations for Compressed Sensing using Generative Models
Manik Dhar, Aditya Grover, Stefano Ermon
International Conference on Machine Learning (ICML), 2018
Paper: https://arxiv.org/abs/1807.01442

Requirements

The codebase is implemented in Python 2.7. To install the necessary requirements, run the following commands:

pip install -r requirements.txt

Setup

The following command will download the CelebA, OMNIGLOT, and MNIST datasets:

bash ./setup/download_data.sh

The following command will unzip the trained model weights for the experiments:

unzip models.zip

The following command will create wavelet basis for the celebA experiments

python ./src/wavelet_basis.py

Options

Learning and inference of Sparse-Gen models is handled by the main.py script which provides the following command line arguments.

  --pretrained-model-dir PRETRAINED_MODEL_DIR
                        Directory containing pretrained model
  --dataset DATASET     Dataset to use
  --input-type INPUT_TYPE
                        Where to take input from
  --input-path-pattern INPUT_PATH_PATTERN
                        Pattern to match to get images
  --num-input-images NUM_INPUT_IMAGES
                        number of input images
  --batch-size BATCH_SIZE
                        How many examples are processed together
  --measurement-type MEASUREMENT_TYPE
                        measurement type
  --noise-std NOISE_STD
                        std dev of noise
  --num-measurements NUM_MEASUREMENTS
                        number of gaussian measurements
  --model-types MODEL_TYPES [MODEL_TYPES ...]
                        model(s) used for estimation
  --mloss1_weight MLOSS1_WEIGHT
                        L1 measurement loss weight
  --mloss2_weight MLOSS2_WEIGHT
                        L2 measurement loss weight
  --zprior_weight ZPRIOR_WEIGHT
                        weight on z prior
  --dloss1_weight DLOSS1_WEIGHT
                        -log(D(G(z))
  --dloss2_weight DLOSS2_WEIGHT
                        log(1-D(G(z))
  --sparse_gen_weight SPARSE_GEN_WEIGHT
                        weight for sparse deviations
  --optimizer-type OPTIMIZER_TYPE
                        Optimizer type
  --learning-rate LEARNING_RATE
                        learning rate
  --momentum MOMENTUM   momentum value
  --max-update-iter MAX_UPDATE_ITER
                        maximum updates to z
  --num-random-restarts NUM_RANDOM_RESTARTS
                        number of random restarts
  --decay-lr            whether to decay learning rate
  --lmbd LMBD           lambda : regularization parameter for LASSO
  --lasso-solver LASSO_SOLVER
                        Solver for LASSO
  --const_dummy CONST_DUMMY
                        dummy hack
  --save-images         whether to save estimated images
  --save-stats          whether to save estimated images
  --print-stats         whether to print statistics
  --checkpoint-iter CHECKPOINT_ITER
                        checkpoint every x batches
  --image-matrix IMAGE_MATRIX
                        0 = 00 = no image matrix, 1 = 01 = show image matrix 2
                        = 10 = save image matrix 3 = 11 = save and show image
                        matrix

Examples

You will need to download the datasets to run the experiments. To run the quantitative experiments as given in the paper, run the scripts in the quant_scripts directory:

bash ./quant_scripts/celebA_reconstruction.sh
bash ./quant_scripts/omniglot_reconstruction.sh
bash ./quant_scripts/mnist_reconstruction.sh

This will generate the scripts in multiple directories for the required experiments which can be run using the utils/run_sequentially.sh script. The exact commands are as follows:

bash ./utils/run_sequentially.sh scripts_mnist
bash ./utils/run_sequentially.sh scripts_mnist2omni
bash ./utils/run_sequentially.sh scritps_omni
bash ./utils/run_sequentially.sh scritps_omni2mnist
bash ./utils/run_sequentially.sh scritps_celebA

When all experiments have finished running the graphs can be generated using:

bash ./setup/make_graphs.py

Portions of the codebase in this repository uses code originally provided in the open-source Compressed Sensing with Generative Model (https://github.com/AshishBora/csgm) repositories.

Citing

If you find Sparse-Gen useful in your research, please consider citing the following paper:

@inproceedings{dhar2018modeling,
title={Modeling Sparse Deviations for Compressed Sensing using Generative Models},
author={Dhar, Manik and Grover, Aditya and Ermon, Stefano},
booktitle={International Conference on Machine Learning},
year={2018}}