Task

MNIST Classification API Service

Goal

Develop an API service for training and serving a deep learning model for MNIST classification, integrated with a model registry and experiment tracking.

Tasks

Design and Implement an API with three endpoints:

Train

• Description:

  Receives hyperparameter and trains a deep neural network for MNIST classification.

• Details:

  • Logs training metrics and artifacts to a local experiment tracking server.
  • Upon completion, returns information about the tracked experiment entry.

Register

• Description: Receives a reference to a training experiment, exports the corresponding model to ONNX format, and promotes it to a model registry.
• Details:
  • May include some model governance processes before registering.
  • Returns information about the registered model.

Predict

• Description: Receives an image file, predicts the digit using the latest registered model, and returns the prediction result.

Introduction

This API service is designed to streamline the training, registration, and prediction processes for a deep learning model focused on MNIST classification. The service features three primary endpoints, each serving a unique purpose in the model lifecycle.

Endpoints Overview

Endpoint	Method	Input	Output
Train	POST	hyperparameters (json)	model ID, and experiment tracking information.
Register	POST	model ID,artifact ID	register ID
Predict	POST	image	predicted digit and confidence score.

Installation

docker

1. To use the docker, you need to use the docker image. Build and run the image in the docker folder.

   $ docker build -t [IMAGE_NAME] .
   $ docker run [OPTIONS] IMAGE[:TAG|@DIGEST] [COMMAND] [ARG...]

2. Install the package using requirements.txt within the docker.

   # pip install -r requirements.txt
   

conda

1. Create a virtual environment

   $ conda create -n mnist python=3.10
   $ conda activate mnist

2. Install the package

   $ pip install -r requirements.txt

Quick Start

There are two ways to use this project.

1. Use docker-compose
2. Use curl
3. Use Swagger API
4. insomnia

docker-compose

1. localhost

   $ export IP_ADDRESS=localhost && \
     docker compose -f .devcontainer/docker-compose.yml up

2. setting host IP

   $ export IP_ADDRESS=[USER_IP] && \
     docker compose -f .devcontainer/docker-compose.yml up

Use curl

1. endpoints

   • /train/ Run the train endpoint to train a CNN model MNIST. To train the model, the user must enter hyper parameters, as shown below.

     • example input:

       $ curl -X POST "http://localhost:8000/train/" \
         -H "Content-Type: application/json" \
         -d '{"learning_rate": 0.001, "epochs": 10, "batch_size": 128}'

   • /register/ To register a trained model with mlflow register endpoint.

     • example input:

       $ curl -X POST "http://localhost:8000/register/" \
         -H "Content-Type: application/json" \
         -d '{ "run_id": "{run_id}", "artifact_path": "model", \
         "registered_model_name": "mnist_model", "registered_artifact_path": "onnx_model" }'

   • /predict/ Run a prediction endpoint to predict your trained model.

     • example input:

       $ curl -X POST "http://localhost:8000/predict/" \
         -H "Content-Type: application/json" \
         -H 'Content-Type: multipart/form-data' \
         -F 'file=@{img_path};type=image/jpeg'

Use swagger API

1. To use swagger refer to http://localhost:8000/docs where the endpoint parameters are shown and we are able to upload new images and test the api.

• /train/

  • input_example
  • output_example

• /register/

  • input
  • output

• /prediction/

  • input
  • output

insomnia

• /train/
• /register/
• /predict/

mlflow

$ mlflow server -p 5000 -h [HOST_IP]

Reference

• CNN Base model:

• Pipeline:

  • https://mlops-for-mle.github.io/tutorial/docs/intro
  • https://codereader.tistory.com/49
  • https://cow-coding.github.io/posts/day6_torch2/

• MLflow:

  • https://mlflow.org/docs/latest/index.html

• APIs example:

  • https://github.com/afoley587/hosting-yolo-fastapi
  • https://github.com/KiLJ4EdeN/fastapi_tf-keras_example
  • https://github.com/woodywarhol9/hydra-practice

• lightning:

  • https://wikidocs.net/157586
  • https://velog.io/@rapidrabbit76/pytorch-lightning-hooks
  • https://baeseongsu.github.io/posts/pytorch-lightning-introduction/
  • https://lightning.ai/docs/pytorch/stable/deploy/production_advanced.html