Skip to content

Model Hosting Service: Docker Template

The Model Hosting Service (MHS) enables Techcyte users to run GPU-based batch processing by uploading custom Docker containers as classifiers. This repository provides a template for developing and testing your container locally or in a production-like environment.

We provide template files below that can be built upon and deployed to the Techcyte infrastructure.

Code Features:

  • Modularity: Implement your image processing logic in main.py’s process_image() function.
  • Test locally: Download an image from SCAN_URL, process it, and post results to Techcyte using provided environment variables.
  • GPU Support: Uses NVIDIA CUDA with a simple GPU test via PyTorch.
  • Image Handling: Supports DICOM/SVS/TIFF via pydicom, openslide-python

Environment Variables

Production

  • HOST: Techcyte host (e.g., ci.techcyte.com).
  • TASK_ID: Task identifier.
  • SCAN_ID: Scan identifier.
  • SCAN_URL: Presigned S3 URL for image download.
  • JWT_TOKEN: A task specific jwt token used for authorizing requests to techcyte (not required for local testing)

Local testing

  • API_KEY_ID: Key id used while running locally
  • API_KEY_SECRET: Key secret used while running locally

Step-by-step instructions

1. Clone the Repository:

git clone https://github.com/Techcyte/devkit.git
cd devkit/src/model-hosting-service

2. Customize Your Code

  • Edit main.py, replacing the process_image() function with your classifier logic.
  • Ensure the output matches the required schema (see below).
  • Add dependencies to Dockerfile if needed (e.g., RUN pip3 install <package>).

3. Build the Docker Image

docker build -t my-docker-image:latest .

4. Test locally

  • Provide environment variables: 
    docker run --rm --gpus all \
  -e HOST="ci.techcyte.com" \
  -e TASK_ID="your-task-id" \
  -e SCAN_ID="your-scan-id" \
  -e SCAN_URL="https://example.com/image.svs" \
  -e API_KEY_ID="your-api-key-id" \
  -e API_KEY_SECRET="your-api-key-secret" \
  my-docker-image:latest
  • If running on production HOST="app.techcyte.com", CI HOST="ci.techcyte.com"
  • See Creating a debug request to obtain TASK_ID, SCAN_ID and SCAN_URL values
  • See Creating an API key to obtain the api key id and secret
  • Downloads the image, processes it, and posts results to Techcyte. Prints results if posting fails.

5. Deploy

  • Push your image to the Techcyte container registry. See instructions here.

Example Output (local output)

CUDA is available! Using GPU: Tesla T4
Matrix multiplication (1000x1000) completed in 0.0246 seconds
Result sample: 21.861495971679688
Processing image: /input/image.svs
Generated 4 fake annotations
Visualization saved to /output/result.png
Results JSON: {"caseResults": {"mitosisCount": 4}, "scanResults": [{"scanId": "test_scan", "geojson": {...}}]}

You should have an image with a 2x2 grid or red squares.

Troubleshooting

  • No GPU:
  • Ensure NVIDIA Docker is installed (nvidia-model-toolkit).
  • Verify --gpus all is used.
  • Check CUDA compatibility (image uses CUDA 11.8; ensure your GPU supports it).
  • If error persists, run without --gpus all to test CPU fallback.
  • Image Issues: Ensure the input file is a valid SVS or TIFF. OpenSlide supports both.
  • Posting Errors: Verify environment variables; invalid tokens print results for debugging.
  • Large Images: OpenSlide handles large SVS files efficiently. For other formats (e.g., DICOM), add pip3 install pydicom to the Dockerfile.
  • Memory Issues: Large images may require significant RAM; ensure your Docker host has enough resources.

Posting Results

You will need to modify the caseResults and/or scanResults values to see your results on the Techcyte app. See posting external model results.