model warm start
closes #800 (closed)
- new dag
train-with-pretrained-weights
which triggers training dags - use PRETRAINED_WEIGHTS env var in clf workflow and nnunet
- minor changes in LocalDagTriggerOperator: adapt to run without any dicom pased to DAGs
- minor changes in clf workflow and nnunet to get the model checkpoints
Known TODOs:
- publication form is not attached for nnunet, because it is not possible to validate required oneOf fields (see #928)
- trigger operator has name "trigger_" because the selected dag id can not be passed to the operator
- trigger operator starts new DAGs as service DAGs (i.e. not possible to interact via Workflow List), becuase the dag runs are not written in the DB by backend