MikeXydas/DeepSqueeze

Implementation of the DeepSqueeze paper: https://cs.brown.edu/people/acrotty/pubs/3318464.3389734.pdf

DeepSqueeze

In this repo I attempt to reproduce the compression utility described in DeepSqueeze: Deep Semantic Compression for Tabular Data by Amir Ilkhechi, Andrew Crotty, Alex Galakatos, Yicong Mao, Grace Fan, Xiran Shi, Ugur Cetintemel from Brown University.

Demo (with sound 🔉)

deep_squeeze_demo.mp4

You can read the report.pdf for info about the original paper, my implementation, my additions and results.

There are 3 branches in this repo:

• master, used in my demo presentation
• experiment, mainly used to run experiments and producing results
• mixture_of_experts, since I was not able to achieve better results using the Mixture of Experts architecture, I decided to keep it into a separate branch reducing code complexity of the master and demo branches

Running

I suggest running DeepSqueeze in the master branch which is cleaned-up following the steps below:

1. Create a python environment. The DeepSqueeze package was developed in python3.8
2. Install the requirements in requirements.txt
3. Download one of the processed tables (no header, only numerical values).
   • Corel dataset
   • Intel Berkeley Research Lab Sensor Data
   • Monitor dataset, due to its size I have not uploaded the preprocessed version. You can download it here and preprocess it using notebooks/preprocessing.ipynb
4. Compress the table with the command:

python compress.py -i path/to/input.csv -o path/to/output/dir/ -e <error_threshold_percentage>

Note that the -e parameter takes a value between [0, 100] with suggested values being: 0.5, 1, 5, 10.

5.Decompress the table with the command:

python decompress.py -i path/to/compressed_tables.zip

This is the full pipeline of DeepSqueeze with some simplification presented in report.pdf.