deep-ternary

This repository contains the source code for the paper titled "Deep Learning Sparse Ternary Projections For Compressed Sensing of Images"

Prerequisites

- Tensorflow 
- Numpy 
- h5py 
- opencv (cv2) 

The code has been tested in Ubuntu 14.04 and MacOSX, with

- Tensorflow v.1.2.1
- Numpy v.1.12.1 
- h5py v.2.7.0
- opencv (cv2) v.2.4.11

Usage

Prepare training data

Currently, my dataset of patches is not too big (~5GB) so during training, a DataLoader instance will load all the training patches into memory to feed the network. To train with your own data, you may need to create a training set of image patches and store it as a hdf5 database. Alternatively, you can write your own DataLoader to load and feed your network from your own dataset.

A sample hdf5 database, with 2000 32x32 image patches, is provided in data folder. Note that this small database is just a sample to show how the data should be prepared, and it is too small to train a well-performing model.

In order to get good model, a much larger training database is necessary. To reproduce the results reported in the paper, you are recommended to download the dataset with 5 million patches. (Note: Due to the strict policy of my university's admin, I cannot share the permanently the download link to this dataset. If you want to download it, please create a new issue here or send me an email.)

Configurations

The configs.py file sets all the default configurations and hyper-parameters. There are some other hyperparameters inside train.py and test.py To use your own configurations, you can either edit these files, or put your configurations as flags

For example:

python train.py --n_epochs=50 --initial_lr=0.005

Training

After preparing your dataset and set all the necessary hyper-parameters, you are ready to train your model. Run train.py, together with our hyper-parameters, to start the training, for example:

python train.py --db_fname=./data/patches_32x32_2k.h5 --batch_size=50 --output_basedir=output

or to train the model with the big database downloaded from the link above:

python train.py --db_fname=./data/imagenet_val_32x32_5m.h5 --batch_size=5000 --output_basedir=output

The trained model will be saved into output/snapshots

Testing

After training the model, you can test it with your images using the test function.

python test.py --test_folder=test_images

All images (jpg, png, tif) inside the test folder will be used for evaluation.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Reference

If you find the source code useful, please cite us:

D. M. Nguyen, E. Tsiligianni and N. Deligiannis, "Deep learning sparse ternary projections for compressed sensing of images," 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), 2017, pp. 1125-1129.