Example: Flood Model Training⚓︎

Practical Case: Training a flood detection model.

Context⚓︎

We want to be able to train our Flood Model using the MLOps platform.

The repositories are on the develop cluster GitLab. They are mirrored in the following repositories:

• Flood Model: https://github.com/EOEPCA/flood-model
• Sen1Floods11-Dataset: https://github.com/EOEPCA/Sen1Floods11-Dataset

Scenario (ML Developer)⚓︎

Step 1: Browsing SharingHub for the model⚓︎

• Navigate to https://sharinghub.develop.eoepca.org
• Click on “Models” category
• Filter “flood”
• Click on the model
• Open in GitLab

Step 2: Clone the model⚓︎

• Flood Model: https://gitlab.develop.eoepca.org/sharinghub-test/flood-model

Step 3: Model setup⚓︎

• Install poetry environment
• Setup .env file: credentials for mlflow

Step 4: Dataset setup⚓︎

• Follow the model README “Getting started”
  • Clone dataset: https://gitlab.develop.eoepca.org/sharinghub-test/sen1floods11-dataset
  • Setup credentials for dvc
• Run “dvc pull” in terminal

Step 5: Train Model⚓︎

• Run training session.
• Check metrics in mlflow UI.
• Show modes streaming and no cache.

Step 6: Inference⚓︎

• Run an inference with a tif file 512x512: Pakistan_43105_S1Hand.tif and import the model from .onnx file.
• Show the result stored in predictions/prediction.tif file.

Step 7: Packaging⚓︎

Docker for training⚓︎

• Build the docker image for training.
• Run this image in a docker container.

Docker for inference⚓︎

• Build the docker image for inference and specify the onnx file to use 1 time.
• Run this image with a tif file 512x512: Pakistan_43105_S1Hand.tif as an input.

Scenario (ML User)⚓︎

CWL for inference⚓︎

• Download the onnx from the sharinghub
• Run the cwl file with cwltool. Parameters are saved in run_inference_input.yml file.