daniabib / overweight-dogs

:dog: PyTorch image classifier to classify images of overweight dogs. The model is a pretrained EfficientNet-B0 with a custom classify layer on top.

Overweight Dogs Image Classifier

A simple example of a PyTorch image classifier to classify images of overweight dogs. The model is a pretrained EfficientNet-B0 with a custom classify layer on top.

The dataset was built scraping Bing Image for terms like "overweight dogs", "obese dogs" and similars. I also wrote some scripts to extract frames of videos containing overweighted dogs to achieve a more robust dataset. The dataset is composed of 5400 images already randomly split into train, validation, and test folders. It can be downloaded here: https://drive.google.com/file/d/1uFDfpyCkMxzdz2Mez5zUBJJHBkh_OJjq/view?usp=sharing

Just for the sake of completeness (and fun!) I added an API call to the AWS Rekognition Label Detection service, so our model is called only when a dog is in the picture. That way our application can function as an object detector too.

A simple Streamlit frontend was built to simplify the use of the model by clients. The model is integrated with FastAPI, and can be accessed on the /prediction endpoint. It is already deployable via Docker and has GPU support if needed.

The application is also already deployable on AWS ECS.

How To Use

To clone and run this application, you'll need Git and Docker:

# Clone this repository
$ git clone https://github.com/daniabib/overweight-dogs

# Go into the repository
$ cd overweight-dogs

Then you need to build and run the Docker image:

# Build and run the image
$ docker-compose up -d --build

The Streamlit frontend will be avaiable at http://localhost:8501.

To test the model you can use the FastAPI documentation page. It will be avaiable on http://0.0.0.0:8080/docs.

You can also use a HTTP client like httpie, for example:

$ http --form http://0.0.0.0:8080/prediction [email protected]

It will return something like this:

HTTP/1.1 200 OK
content-length: 25
content-type: application/json
date: Fri, 18 Mar 2022 21:09:45 GMT
server: uvicorn

{
    "category": "overweight"
}

Things I'm still implementing

• A simple Streamlit frontend. ✅
• A docker-compose pipeline to simplify deployment on AWS ECS. ✅

References

For this code, I used a variety of sources, but mainly these:

• The book Building Data Science Applications with FastAPI, by François Voron.
• Machine learning model serving in Python using FastAPI and streamlit, by Davide Fiocco.
• Deploying PyTorch Model to Production with FastAPI in CUDA-supported Docker, by Ming Choi.

Disclaimer

This repository is just an exercise on deep learning, computer vision, and model deployment. It has no intention of having any scientific validation. Please, use it only as a code example.