This is a SegServe executor provides a minimal working example (MWE) and demonstrates how to add custom executors into the ObiWan-Microbi software.
The general idea of the executors is that we encapsulate the segmentation process into mlflow projects. Thus, all dependencies are collected in a conda.yaml and access interfaces are defined in MLProject.
mlflow is then used to download a specific version of this repository from its github url and executes the segmentation method on a list of images (see test cases).
In order to list all software dependencies please edit the conda.yaml file. It is best practice to fix the software package versions to minimize problems occuring due to new releases.
For deep-learning dependencies, like pytorch and tensorflow please have a look at their installation insturctions and add the dependencies or additional channels to the conda.yaml file. You can also find examples in our exsting executor repositories.
The main segmentation work is performed in the main.py and there in step (2). For every image a segmentation method can be called, e.g. utilizing deep-learning tools like pytorch or tensorflow in order to obtain individual contours of objects.
These instance segmentation objects are then saved into a json format and provided to segServe. Therefore, these executors are fully independent, manage their own dependencies and only need to adhere to the segmentation storing format.
Advanced usages include, e.g., multiple segmentation method implementations in a single executor repository. This can be achieved by completely different implementations or by custom parameters that can be passed. For this you need to edit the MLProject file and add new entrypoints (see mlflow documentation).
Make sure you have anaconda installed and an active environment with mlflow. Then execute
pip install mlflow
mlflow run ./ -e main -P input_images=<path to your local image or image folder (*.png)>and replace <path to your local image or image folder (*.png)> by a path to an image or a folder of images. For a folder, all images need to be present in png format but this behavior can be changed in main.py. For specifying custom parameters please have a look into the mlflow documentation.
The resulting segmentation should be written to output.json and logged as an artifact in the mlflow run.