wushanshan / L1AE

Code for our paper "Learning a compressed sensing measurement matrix via gradient unrolling".

Toy example

Under the folder toy_example, we provide a jupyter notebook jan22_toy_example.ipynb that works through the training and evaluation of our autoencoder (as well as other baseline algorithms) for the synthetic1 dataset. We highly recommend interested readers to take a look before diving deep into our code.

Overview

The source code contains four parts:

1. Core
   • model.py
   • utils.py
   • datasets.py
   • baselines.py
2. Code for each dataset
   • synthetic_main.py
   • synthetic_powerlaw_main.py
   • amazon_main.py
   • amazon_parallel_l1.py
   • rcv1_main.py
   • rcv1_parallel_l1.py
3. Scripts for reproducing our results
   • scripts/synthetic1.sh
   • scripts/synthetic2.sh
   • scripts/amazon.sh
   • scripts/rcv1.sh
   • scripts/synthetic_powerlaw.sh
4. Code and scripts for one of the baselines Simple AE + l1-min
   • synthetic_simpleAE.py
   • amazon_simpleAE.py
   • rcv1_simpleAE.py
   • scripts are under simpleAE_scripts/

Run

To reproduce our experimental results, first run chmod +x scripts/*.sh to make the scripts executable. After that, run the given scripts:

• $ ./scripts/synthetic1.sh
• $ ./scripts/synthetic2.sh
• $ ./scripts/amazon.sh
• $ ./scripts/rcv1.sh
• $ ./scripts/synthetic_powerlaw.sh

Note:

1. The results are stored in a python dictionary which is then saved under the folder ckpts/. They can be used to reproduce the figures shown in our paper.
2. Before running amazon.sh, download train.csv from this kaggle competition and specify its location via --data_dir.
3. The RCV1 dataset will be fetched automatically using the sklearn.datasets.fetch_rcv1 function.
4. To reproduce results of one of the baselines Simple AE + l1-min, run scripts under the folder simpleAE_scripts/.
5. For high-dimensional vectors, solving l1-min using Gurobi takes a long time on a single CPU. To speed up, we solve l1_min in parallel on a multi-core machine. In amazon_main.py and rcv1_main.py, performance evaluation is performed on a small set of the test samples (while training is still done using the complete training set). After training the autoencoder, we use a multi-core machine and solve l1_min in parallel on the complete test set using amazon_parallel_l1.py and rcv1_parallel_l1.py. Depending on your multi-core machine, solving l1_min in parallel on the complete test set may still take a long time, I would recommend running amazon_parallel_l1.py and rcv1_parallel_l1.py first with a small subset (by setting a small number for the parameters num_core and batch in the python file).

Environment

Here is our software environment.

1. Python 2.7.12
   • numpy 1.13.3
   • sklearn 0.19.1
   • scipy 1.0.0
   • joblib 0.10.0
2. Tensorflow r1.4
3. Gurobi 7.5.1