Instance segmentation on COCO-like dataset: RuntimeError: Masks are expected to be MaskFile objects.

[Borda]

🐛 Bug

Trying to run instance segmentation with this fixing PR #1017 on Kaggle competition which provides a dataset in CoCo like format, but getting this strange error and not sure what is the proper "MaskFile objects" as the dataset provides annotation like:

{
 'id': 101439,
 'image_id': 101438,
 'category_id': 1,
 'segmentation': [[281.6, 234.28, 292.24, 236.48, 298.84, 239.04, 305.08, 239.41, 308.38, 239.41, 310.21, 236.11, 310.21, 230.61, 308.75, 226.21, 304.34, 220.71, 302.88, 218.14, 300.68, 214.47, 296.64, 216.3, 288.94, 216.3, 287.11, 219.61, 284.17, 226.94]],
 'area': 492.789049999963,
 'bbox': [281.6, 214.47, 28.609999999999957, 24.939999999999998],
 'iscrowd': 0
}

For sanity adding checking the dataset with

print(f"missing Annot: {[i for i, an in enumerate(data['annotations']) if not an]}")
print(f"missing Segm: {[i for i, an in enumerate(data['annotations']) if not an['segmentation']]}")

which finds nothing suspicious

missing Annot: []
missing Segm: []

To Reproduce

/opt/conda/lib/python3.7/site-packages/flash/core/data/io/input_base.py in _call_load_sample(self, sample)
    149             ),
    150         )
--> 151         return load_sample(copy(sample))
    152 
    153     @staticmethod

/opt/conda/lib/python3.7/site-packages/flash/core/integrations/icevision/data.py in load_sample(self, sample)
     60     def load_sample(self, sample: Dict[str, Any]) -> Dict[str, Any]:
     61         record = sample[DataKeys.INPUT].load()
---> 62         return from_icevision_record(record)
     63 
     64     def predict_load_sample(self, sample: Dict[str, Any]) -> Dict[str, Any]:

/opt/conda/lib/python3.7/site-packages/flash/core/integrations/icevision/transforms.py in from_icevision_record(record)
    214         sample[DataKeys.INPUT] = record.filepath
    215 
--> 216     sample[DataKeys.TARGET] = from_icevision_detection(record)
    217 
    218     if getattr(record.detection, "class_map", None) is not None:

/opt/conda/lib/python3.7/site-packages/flash/core/integrations/icevision/transforms.py in from_icevision_detection(record)
    160                 result["masks"].append(mask.filepath)
    161             else:
--> 162                 raise RuntimeError("Masks are expected to be MaskFile objects.")
    163 
    164     if hasattr(detection, "keypoints"):

RuntimeError: Masks are expected to be MaskFile objects.

Code sample

See the Kaggle kernel: https://www.kaggle.com/jirkaborovec/cell-instance-segm-lightning-flash?scriptVersionId=83933340

Additional context