Overhead MNIST

Dataset info

The dataset used in this project is the Overhead-MNIST dataset and it can be found here
The dataset contains 10 classes

Project info

This repo contains an end-to-end deep learning project deployment for overhead image classification
The purpose of this project is to try multiple new things. The following new things that have been explored
- In this project, the recently popular Kolmogorov-Arnold Networks (KAN) is also explored
- Simple and Complex CNNs using the KAN layer is developed and tested
- Weights and Biases has been used for MLOps including its new feature of model registry
- For training, the Automatic Mixed Precision (AMP) has been explored by using float16 for some layer weights to speed up training
- For deployment, an API has been developed and deployed using FastAPI and docker
For the training, the dataset is split into 95% - 5% for train and validation sets respectively
The python packages are listed in requirements.txt
The docker container can be deployed using Dockerfile
For training and logging the model, use the modeling/train.py script
The FastAPI app deployment code is in app.py script
To test the deployed FastAPI app on a local machine, the test_post_request.py script can be used
Some sample test images are available in sample_test_images

Training using docker image

Use the following command for training using the docker image

docker run --rm -it --init   --gpus=all   --ipc=host   --user="$(id -u):$(id -g)"   --volume="$PWD:/app"   my_pytorch python3 modeling/train.py --dir_dataset /app/dir_dataset/

Docker deployment instructions on a local machine

For deployment requirements_deployment.txt needs to be used
To build the container, run the following command

docker build -t fastapi_overhead_mnist .

To the run the container, run the following command

docker run -p 7860:7860 -t fastapi_overhead_mnist

HuggingFace deployment

The FastAPI application has also been deployed to HuggingFace
To test the deployed FastAPI app on HuggingFace, use the test_post_request.py script in the HuggingFace repo since the endpoint is different