lattice-ai / Compressed-DNNs-Forget

Minimal Reproducibility Study of (https://arxiv.org/abs/1911.05248). Experiments with Compression of Deep Neural Networks

What Do Compressed Deep Neural Networks Forget?

This Project aims to reproduce the claims of the paper titled "What Do Compressed Deep Neural Networks Forget?" by Sara Hooker, Aaron Courville, Gregory Clark, Yann Dauphin and Andrea Frome.

Weights and Biases Project Page

Weights and Biases Client were used to conduct all the experiments. You can explore the Project here.

Metrics

Train Top-1 Accuracy

Validation Top-1 Accuracy

Training Loss

Web Application

You can interact with the web app at this link. Some identified PIE's are also available on the website.

Methods of Pruning

Steps for Reproduction

Conda Approach

(After Cloning and switching to the web-app branch)

1. Create the conda environment conda env create -f environment.yml
2. Activate the Environment conda activate compression
3. Run the Application streamlit run app.py

Docker Approach

docker pull docker.pkg.github.com/sauravmaheshkar/compressed-dnns-forget/compression-app:latest
docker run -p 8501:8501 compression-app:latest                                                             

Contribute

If you want to contribute to the project kindly mail me at [e-mail].

Step 1

• Option 1 🍴 Fork it!
• Option 2 👯‍♂️ Clone this repo to your local machine using https://github.com/SauravMaheshkar/Compressed-DNNs-Forget.git

Step 2

• HACK AWAY! 🔨🔨🔨

Step 3

• 🔃 Create a new pull request using https://github.com/SauravMaheshkar/Compressed-DNNs-Forget/compare/

License

The data for this project was taken from kaggle datasets. You can find the Large-scale CelebFaces Attributes (CelebA) Dataset here.

• Copyright 2020 @Saurav Maheshkar
• MIT License

Credits

The inspiration for this readme file came from

• fvcproductions
• Navendu Pottekkat