Tutorial for using mmclassification

Outline

This is a tutorial for using mmclassification to:

• train an image classification model on your own custom dataset
• run inference using your trained model
• serve the model using a plain vanilla FastAPI wrapper (you should really look at TorchServe for production workloads, mmclassification has out of the box support for it.)

Prerequisites

To run the training and inference notebooks, you need a Google account to access Colab.

1. When in Google Colab, click on File > Open Notebook and select Github as the source.
2. For Training, put in https://github.com/billcai/mmclassify-tutorial/blob/master/food_model_classifier.ipynb as the link.
3. For Inference, put in https://github.com/billcai/mmclassify-tutorial/blob/master/food_model_inference.ipynb as the link. Load the notebooks and they should be good to go.

For serving the model, we have a Dockerfile, which you can build using

cd serve-api
docker build . -t mmclassify-api:latest

This requires you to have a GPU-enabled device that has compute ability 7.5 and above (basically a GTX 1060 and above roughly). Then, assuming that you have nvidia-docker installed, you can start serving the model with

docker run --name mmclassify-api -p 8000:8000 --gpus all mmclassify-api:latest

You would be able to see the OpenAPI docs and test your API at http://localhost:8000/docs.