Bag of Tricks for Faster & Stable Image Classification

This is a Python library for using the latest SOTA techniques to improve the training pipeline of DNNs, making them faster and robust.

This was done for the course CS231n (Spring 2022) at Stanford. Find our poster here: http://cs231n.stanford.edu/reports/2022/pdfs/122p.pdf, and our paper here: http://cs231n.stanford.edu/reports/2022/pdfs/122.pdf

Installation

Use the package manager conda to install torch==1.7 torchvision==0.8

Run the following script to install custom PyTorch layers for Approximate Tensor Operations

./setup.sh

Usage

Sample Notebook showing how to use the library is attached in ./src/demo.ipynb

Adding a New Model

Edit ./src/models.py, and add your model to the model_dict in get_model

def get_model(model_name, model_params, learning_rate, loader_train, num_channels, device):
    model_dict = {
        'VGG16' : models.vgg16, 
        'Resnet18' : models.resnet18, 
        'Resnet50' : models.resnet50, 
    }

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

If you use this repository or extend it, please cite us.

Authors

Aman Bansal

Shubham Anand Jain

Bharat Khandelwal