Key Error showing up in training

[ajenez]

Today I noticed that if I try to export a coco json file, the files that are created have a KeyError: 'segmentation' issue

This is the traceback I'm seeing

Original Traceback (most recent call last): File "/home/src/model_training/detectron-training/venv/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 308, in _worker_loop data = fetcher.fetch(index) File "/home/src/model_training/detectron-training/venv/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 32, in fetch data.append(next(self.dataset_iter)) File "/home/src/model_training/detectron-training/detectron2/detectron2/data/common.py", line 296, in iter yield self.dataset[idx] File "/home/src/model_training/detectron-training/detectron2/detectron2/data/common.py", line 125, in getitem data = self._map_func(self._dataset[cur_idx]) File "/home/src/model_training/detectron-training/detectron2/detectron2/utils/serialize.py", line 26, in call return self._obj(*args, **kwargs) File "/home/src/model_training/detectron-training/detectron2/detectron2/data/dataset_mapper.py", line 189, in call self._transform_annotations(dataset_dict, transforms, image_shape) File "/home/src/model_training/detectron-training/detectron2/detectron2/data/dataset_mapper.py", line 131, in _transform_annotations instances = utils.annotations_to_instances( File "/home/src/model_training/detectron-training/detectron2/detectron2/data/detection_utils.py", line 413, in annotations_to_instances segms = [obj["segmentation"] for obj in annos] File "/home/src/model_training/detectron-training/detectron2/detectron2/data/detection_utils.py", line 413, in segms = [obj["segmentation"] for obj in annos] KeyError: 'segmentation'

[ajenez]

it seems to happen after I change the name of any object

[ryouchinsa]

Thanks for writing the issue.

It looks that "segmentation" is not written in the COCO format because of some reason. We have to reproduce the problem.

Changing the name of any object, this means that on the label dialog, you select another existing object name or add a new object name? After that, exporting to COCO format, the COCO format will be different from the one before changing the object name, such as "segmentation" is missing.

[ryouchinsa]

Recently we wrote tutorials how to train Detectron2 models on Ubuntu 22.04 with GPU.

Train bounding boxes using a Faster R-CNN model on Detectron2 Train polygons using a Mask R-CNN model on Detectron2 Train RLE masks using a Mask R-CNN model on Detectron2 Train keypoints using a Keypoint R-CNN model on Detectron2

We trained tens of times on Detectron2 with GPU, but could not encounter the problem. Currently If you need our support to fix this problem, please let us know.