Training worker for segmentation/bbox
Refs https://redmine.teklia.com/issues/5530
We need a new worker to train a YOLO segmenter on datasets. We want the same structure as YOLOTrainCls
. The dataset needs to be prepared differently and we need to support two modes.
User configuration:
-
mode
, enum, choices ="segment", "bbox"
, default"segment"
, - all the parameters as for the classification worker (except
dropout
), -
overlap_mask
(bool) defaults to True, -
mask_ratio
(int), defaults to 4.
The extraction code should be similar to https://gitlab.teklia.com/dla/yolo/-/blob/main/teklia_yolo/extract/bbox.py?ref_type=heads. We need to generate a YAML file which will be stored in self.training_data
(that's what is used for the model.train
call). It's the file saved here.