MedMNIST with Determined

Example of training a computer vision model on the MedMNIST dataset using Determined AI. Currently configured to train a model on the PathMNIST dataset, which is one of 12 2-D datasets offered by the MedMNIST project (https://medmnist.com/).

Changing the dataset

For training purposes

1. Replace data_flag in config.yaml with the name of a MedMNIST subset. The subset names are listed here, in the INFO variable. For example, pathmnist is one of the subset names.

For inference purposes

1. Copy the example environment file .env_example to .env .
2. Update DATASET_FLAG to the same desired data_flag.

Training the model

After setting up Determined, export DET_MASTER to your desired cluster IP:

  export DET_MASTER=<your_desired_master_ip>

Submit an experiment:

  det e create config.yaml .

Running inference

Install requirements:

  pip install -r inference_requirements.txt

Copy the example environment file .env_example to .env .

  cp .env_example .env

In the .env file, change EXPERIMENT_ID to your experiment ID from your training run:

  EXPERIMENT_ID=<your_experiment_ID>

Generate sample images to run inference on using the test dataset:

  mkdir sample_images
  python3 generate_sample_test_images.py

Deploy the Flask server:

  python3 deploy.py

Run inference on an image (replace filename with desired image name):

  curl -X POST -F file=@sample_images/<filename.jpg> http://localhost:5000/predict