single-object-detection 0.1.1

Detecting single objects in images based on bounding boxes.

Experimentation Example

Workflow

1. Import required data via versioned repository. dvc import https://git.inf.h-brs.de/mi-project/demo-pipeline/data-annotation-example data/field_A/2023/ -o data/field_A/2023
2. Other added data should be manually versioned using dvc add datapath
3. Reproduce current model training and evaluation using dvc repro.
4. Alter dvc.yaml for new data input paths, or params.yaml for different training configuration.

Tools

DVC

Data Versioning and Reproducible pipelines.

MLFlow

Experiment tracking and model registry.

Pre Commit Hooks

Clear formated and linted code.

Containerization - Docker

Platform independence.

TODO: Create an executable container, which is callable like a script. (Similar to max postprocessing pipeline.) A continuously running service is currently not required.

Insights

• VirtualEnv Always install virtualenv with python -m venv. Else weird behavior can happen. e.g. virtualenv venv

• MLFLow pyfunc Loaded model expects input in numpy format, instead of torch tensors.

• DVC and MLFlow integration MLflow generates random run_ids, which need to be tracked if reproducible pipelines should be set up with dvc. Direct integration was not possible, therefore a workaround with a temporary file had to be implemented.

The output of the training stage is currently not tracked.

• Poetry python version management Poetry natively tries to use the same python versioned used for its installation. Python version management with pyenv requires the experimental parameter virtualenvs.prefer-active-python to be set to true.