Container to run linear model with embeddings input and classification output using json config to load model weights
There are two types of tests: unit tests, which are best run from inside the container, and integration tests that are run from the host and test that the container does the correct thing when passed the parameters.
Run pytest -m docker tests/docker to run tests from the host. This will launch the docker container. You will need the docker daemon running.
This requires that you have pytest installed in your python environment on the host.
Interactively:
Run docker run -ti --rm qutecoacoustics/crane-linear-model-runner sh to launch the container interactively, then python -m pytest -m "not docker" tests/unit to run the tests that are not the docker tests.
From host directly:
Run docker run --rm --entrypoint python lmr:latest -m pytest -m "not docker" tests/unit
To run:
docker run --rm -v input_folder:/mnt/input -v output_folder:/mnt/output -v config_file:/mnt/config/config.json <image_name>
By default, it will classify all parquet files in /mnt/input the input folder, and save the results as csv in /mnt/output/, using /mnt/config.json as the configuration file.
A configuration json file must be supplied. The configuration file contains:
- classifier: the model, including
- the list of classes,
- the beta (weights per class)
- this is am ascii base64 encoded numpy array of shape (num_classes, embedding_size)
- beta_bias
- this is am ascii base64 encoded numpy array of shape (num_classes, )
- any model config saved with the model (not implemented yet but this will be included with the results)
- threshold: the cutoff for the output
Examples can be found in tests/test_data/config