openvinotoolkit
ResizeTransform bug when dataset contains an RleMask annotation
Attempting to resize a dataset with RleMask annotation types fails.
Root cause seems to be attempting to initialize the "resized" RleMask using an image mask array here. isinstance(ann, Mask) catches RleMask due to type inheritance. isinstance(RleMask, Mask) == True
https://github.com/openvinotoolkit/datumaro/blob/2494e15dfa4db24ad5ea8ec10247f341b70f36f0/src/datumaro/plugins/transforms.py#L1109-L1111
The test for this method excludes an RleMask in the sample dataset https://github.com/openvinotoolkit/datumaro/blob/2494e15dfa4db24ad5ea8ec10247f341b70f36f0/tests/unit/test_transforms.py#L904-L930
Reproduce with
small_dataset = Dataset.from_iterable(
[
DatasetItem(
id=i,
media=Image.from_numpy(data=np.ones((4, 4)) * i),
annotations=[
Label(1),
Bbox(1, 1, 2, 2, label=2),
Polygon([1, 1, 1, 2, 2, 2, 2, 1], label=1),
PolyLine([1, 1, 1, 2, 2, 2, 2, 1], label=2),
Points([1, 1, 1, 2, 2, 2, 2, 1], label=2),
Mask(
np.array(
[
[0, 0, 1, 1],
[1, 0, 0, 1],
[0, 1, 1, 0],
[1, 1, 0, 0],
]
)
),
RleMask(pycocotools.mask.encode(
np.asfortranarray(np.array(
[
[0, 0, 1, 1],
[1, 0, 0, 1],
[0, 1, 1, 0],
[1, 1, 0, 0],
]
).astype(np.uint8))
)),
],
)
for i in range(3)
],
categories=["a", "b", "c"],
)
small_dataset.transform("resize", width=100, height=100)
Full traceback
dset.transform("resize", width=100, height=100)
Out[86]: ---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
File .../lib/python3.10/site-packages/IPython/core/formatters.py:708, in PlainTextFormatter.__call__(self, obj)
701 stream = StringIO()
702 printer = pretty.RepresentationPrinter(stream, self.verbose,
703 self.max_width, self.newline,
704 max_seq_length=self.max_seq_length,
705 singleton_pprinters=self.singleton_printers,
706 type_pprinters=self.type_printers,
707 deferred_pprinters=self.deferred_printers)
--> 708 printer.pretty(obj)
709 printer.flush()
710 return stream.getvalue()
File .../lib/python3.10/site-packages/IPython/lib/pretty.py:410, in RepresentationPrinter.pretty(self, obj)
407 return meth(obj, self, cycle)
408 if cls is not object \
409 and callable(cls.__dict__.get('__repr__')):
--> 410 return _repr_pprint(obj, self, cycle)
412 return _default_pprint(obj, self, cycle)
413 finally:
File .../lib/python3.10/site-packages/IPython/lib/pretty.py:778, in _repr_pprint(obj, p, cycle)
776 """A pprint that just redirects to the normal repr function."""
777 # Find newlines and replace them with p.break_()
--> 778 output = repr(obj)
779 lines = output.splitlines()
780 with p.group():
File .../lib/python3.10/site-packages/datumaro/components/dataset.py:261, in Dataset.__repr__(self)
257 def __repr__(self) -> str:
258 separator = "\t"
259 return (
260 f"Dataset\n"
--> 261 f"\tsize={len(self._data)}\n"
262 f"\tsource_path={self._source_path}\n"
263 f"\tmedia_type={self.media_type()}\n"
264 f"\tannotated_items_count={self.get_annotated_items()}\n"
265 f"\tannotations_count={self.get_annotations()}\n"
266 f"subsets\n"
267 f"\t{separator.join(self.get_subset_info())}"
268 f"infos\n"
269 f"\t{separator.join(self.get_infos())}"
270 f"categories\n"
271 f"\t{separator.join(self.get_categories_info())}"
272 )
File .../lib/python3.10/site-packages/datumaro/components/dataset_storage.py:369, in DatasetStorage.__len__(self)
367 def __len__(self) -> int:
368 if self._length is None:
--> 369 self.init_cache()
370 return self._length
File .../lib/python3.10/site-packages/datumaro/components/dataset_storage.py:178, in DatasetStorage.init_cache(self)
176 def init_cache(self) -> None:
177 if not self.is_cache_initialized():
--> 178 for _ in self._iter_init_cache():
179 pass
File .../lib/python3.10/site-packages/datumaro/components/dataset_storage.py:185, in DatasetStorage._iter_init_cache(self)
181 def _iter_init_cache(self) -> Iterable[DatasetItem]:
182 try:
183 # Can't just return from the method, because it won't add exception handling
184 # It covers cases when we save the null error handler in the source
--> 185 for item in self._iter_init_cache_unchecked():
186 yield item
187 except _ImportFail as e:
File .../lib/python3.10/site-packages/datumaro/components/dataset_storage.py:268, in DatasetStorage._iter_init_cache_unchecked(self)
266 if transform and transform.is_local:
267 old_id = (item.id, item.subset)
--> 268 item = transform.transform_item(item)
270 item_id = (item.id, item.subset) if item else None
272 if item_id in cache:
File .../lib/python3.10/site-packages/datumaro/components/dataset_storage.py:103, in _StackedTransform.transform_item(self, item)
101 if item is None:
102 break
--> 103 item = t.transform_item(item)
104 return item
File .../lib/python3.10/site-packages/datumaro/plugins/transforms.py:1111, in ResizeTransform.transform_item(self, item)
1109 elif isinstance(ann, Mask):
1110 rescaled_mask = self._lazy_resize_mask(ann, new_size)
-> 1111 resized_annotations.append(ann.wrap(image=rescaled_mask))
1112 elif isinstance(ann, (Caption, Label)):
1113 resized_annotations.append(ann)
File .../lib/python3.10/site-packages/datumaro/components/annotation.py:104, in Annotation.wrap(self, **kwargs)
102 def wrap(self, **kwargs):
103 "Returns a modified copy of the object"
--> 104 return attr.evolve(self, **kwargs)
File .../lib/python3.10/site-packages/attr/_funcs.py:419, in evolve(*args, **changes)
416 if init_name not in changes:
417 changes[init_name] = getattr(inst, attr_name)
--> 419 return cls(**changes)
TypeError: RleMask.__init__() got an unexpected keyword argument 'image'
Hi @jlwhelan28, apologies for the delayed response. I've reviewed this and confirmed that the TypeError: RleMask.__init__() got an unexpected keyword argument 'image' occurs when attempting to resize RleMask annotations, as you described. This is indeed due to misinterpretation of type inheritance like RleMask as a Mask. Thank you for providing detailed analysis.
To prevent this issue, I've made modifications to enable separate resizing for RleMask.
https://github.com/openvinotoolkit/datumaro/blob/b29b8c38dcce7a1ca96224580b542667129ac95a/src/datumaro/plugins/transforms.py#L1073-L1080
Additional test code related to this has been added, so please feel free to refer to it if you wish to confirm the issue.
https://github.com/openvinotoolkit/datumaro/blob/b29b8c38dcce7a1ca96224580b542667129ac95a/tests/unit/test_transforms.py#L927-L940
This fix will be included in Datumaro 2.0, scheduled for release at the end of March. If you encounter any issues related to this in the future, please reopen the issue and let us know. Thank you for your interests.
