Fine-tuning: AttributeError: Cannot find field 'gt_masks' in DataLoader with COCO format

Unfortunately, I always run into an error when fine-tuning on my own dataset (in coco format). The detectron2 documentation on datasets was of no help either (maybe I overlooked something though).

After registering the data via register_coco_instances (as seen in train_net_dla.py) the following code

train data (train_data = DatasetCatalog.get("dla_train")
train data[0]

gives me:

{'file_name': 'cyber_train/public_imgs_sampled/file01.png',
 'height': 1754,
 'width': 1241,
 'image_id': 0,
 'annotations': [{'iscrowd': 0,
   'bbox': [78, 600, 977, 33],
   'category_id': 1,
   'segmentation': [[83, 606, 83, 633, 1050, 633, 1050, 606]],
   'bbox_mode': <BoxMode.XYWH_ABS: 1>}
...

which does look good to me telling from the COCO formatting style.

I can even visualize the data using Visualizer:

for d in random.sample(train_data, 1):
    img = cv2.imread(d["file_name"])
    visualizer = Visualizer(img[:, :, :], scale=0.5)
    out = visualizer.draw_dataset_dict(d)
    plt.imshow(out.get_image()[:,:,::-1])

which gives me the expected output:

Creating the trainer is also possible using trainer = DefaultTrainer(cfg) , but calling trainer.train() gives me a stack trace for the AttributeError: Cannot find field 'gt_masks' in the given Instances!' in DataLoader. I am very confused about the error, because ‘segmentation’ is present in the train_data and the specified polygons are also valid telling from the output of Visualizer.

Anyway, thank you very much for your great work @hpanwar08 !