bug for "RandCropByLabelClassesd" on multiple datapoints for each batch

[DorisxinDU]

Describe the bug
i am trying to use RandCropByLabelClassesd for image and label with following shape, but shows below bug. please can you help me. thanks.

(8, 1, 224, 224, 64) (8, 9, 224, 224, 64)
(8, 1, 224, 224, 64) (8, 9, 224, 224, 64)
(8, 1, 224, 224, 64) (8, 9, 224, 224, 64)

---

ValueError Traceback (most recent call last)
Cell In [110], line 12
8 # data=resizer(data)
9 # data=resizer2(data)
10 # data=reshape(data)
11 print(data['image'].shape,data['label'].shape)
---> 12 result = cropper(data)
13 print(result)

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/dictionary.py:999, in RandCropByLabelClassesd.call(self, data)
996 image = d[self.image_key] if self.image_key else None
997 indices = d.pop(self.indices_key, None) if self.indices_key is not None else None
--> 999 self.randomize(label, indices, image)
1001 # initialize returned list with shallow copy to preserve key ordering
1002 ret: List = [dict(d) for _ in range(self.cropper.num_samples)]

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/dictionary.py:991, in RandCropByLabelClassesd.randomize(self, label, indices, image)
988 def randomize(
989 self, label: torch.Tensor, indices: Optional[List[NdarrayOrTensor]] = None, image: Optional[torch.Tensor] = None
990 ) -> None:
--> 991 self.cropper.randomize(label=label, indices=indices, image=image)

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/array.py:1219, in RandCropByLabelClasses.randomize(self, label, indices, image)
1217 else:
1218 indices_ = indices
-> 1219 self.centers = generate_label_classes_crop_centers(
1220 self.spatial_size, self.num_samples, label.shape[1:], indices_, self.ratios, self.R, self.allow_smaller
1221 )

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/utils.py:558, in generate_label_classes_crop_centers(spatial_size, num_samples, label_spatial_shape, indices, ratios, rand_state, allow_smaller)
556 ratios_: List[Union[float, int]] = ([1] * len(indices)) if ratios is None else ratios
557 if len(ratios_) != len(indices):
--> 558 raise ValueError("random crop ratios must match the number of indices of classes.")
559 if any(i < 0 for i in ratios_):
560 raise ValueError("ratios should not contain negative number.")

ValueError: random crop ratios must match the number of indices of classes.

[KumoLiu]

Hi @DorisxinDU, from the error message, it looks like you didn't specify the length of the ratios or indices correctly.
Could you please share how you instantiate RandCropByLabelClassesd? Thanks!

[DorisxinDU]

Hi @KumoLiu , thanks for reply.
The definition is: cropper=RandCropByLabelClassesd(keys=['image','label'],label_key='label_stacked',num_classes=10,spatial_size=roi_size,num_samples=4)
result = cropper(data)
where data={'image': datapoints['image'],'label':datapoints['label'],"label_stacked":datapoints["label"]}, and the shape of datapoints['image'] is (8, 1, 224, 224, 64), the shape of datapoints["label"] is (8, 9, 224, 224, 64). the order of the shape is (Batch,channel,width,length, num_slides). I tried both using binary one-hot datapoints["label"] and class based datapoints["label"] by changing it to (8, 1, 224, 224, 64). However both cases do not work.

Please see the following error using classes based error:

ValueError Traceback (most recent call last)
Cell In [98], line 20
10 cropper=RandCropByLabelClassesd(keys=['image','label'],label_key='label_stacked',num_classes=10,spatial_size=roi_size,num_samples=4)
11 # data={'image': datapoints['image'][0],'label':datapoints['label'][0]} ##remove [0] will not work
12 # print(data['image'].shape,data['label'].shape)
13 # # reshape=EnsureChannelFirstd(keys=["image"])
(...)
18 # # data=reshape(data)
19 # print(data['image'].shape,data['label'].shape)
---> 20 result = cropper(data)

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/dictionary.py:999, in RandCropByLabelClassesd.call(self, data)
996 image = d[self.image_key] if self.image_key else None
997 indices = d.pop(self.indices_key, None) if self.indices_key is not None else None
--> 999 self.randomize(label, indices, image)
1001 # initialize returned list with shallow copy to preserve key ordering
1002 ret: List = [dict(d) for _ in range(self.cropper.num_samples)]

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/dictionary.py:991, in RandCropByLabelClassesd.randomize(self, label, indices, image)
988 def randomize(
989 self, label: torch.Tensor, indices: Optional[List[NdarrayOrTensor]] = None, image: Optional[torch.Tensor] = None
990 ) -> None:
--> 991 self.cropper.randomize(label=label, indices=indices, image=image)

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/croppad/array.py:1219, in RandCropByLabelClasses.randomize(self, label, indices, image)
1217 else:
1218 indices_ = indices
-> 1219 self.centers = generate_label_classes_crop_centers(
1220 self.spatial_size, self.num_samples, label.shape[1:], indices_, self.ratios, self.R, self.allow_smaller
1221 )

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/utils.py:575, in generate_label_classes_crop_centers(spatial_size, num_samples, label_spatial_shape, indices, ratios, rand_state, allow_smaller)
573 center = unravel_index(indices_to_use[random_int], label_spatial_shape).tolist()
574 # shift center to range of valid centers
--> 575 centers.append(correct_crop_centers(center, spatial_size, label_spatial_shape, allow_smaller))
577 return centers

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/transforms/utils.py:444, in correct_crop_centers(centers, spatial_size, label_spatial_shape, allow_smaller)
426 def correct_crop_centers(
427 centers: List[int],
428 spatial_size: Union[Sequence[int], int],
429 label_spatial_shape: Sequence[int],
430 allow_smaller: bool = False,
431 ):
432 """
433 Utility to correct the crop center if the crop size and centers are not compatible with the image size.
434
(...)
442
443 """
--> 444 spatial_size = fall_back_tuple(spatial_size, default=label_spatial_shape)
445 if any(np.subtract(label_spatial_shape, spatial_size) < 0):
446 if not allow_smaller:

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/utils/misc.py:201, in fall_back_tuple(user_provided, default, func)
164 """
165 Refine user_provided according to the default, and returns as a validated tuple.
166
(...)
198
199 """
200 ndim = len(default)
--> 201 user = ensure_tuple_rep(user_provided, ndim)
202 return tuple( # use the default values if user provided is not valid
203 user_c if func(user_c) else default_c for default_c, user_c in zip(default, user)
204 )

File ~/anaconda3/envs/transform/lib/python3.9/site-packages/monai/utils/misc.py:158, in ensure_tuple_rep(tup, dim)
155 if len(tup) == dim:
156 return tuple(tup)
--> 158 raise ValueError(f"Sequence must have length {dim}, got {len(tup)}.")

ValueError: Sequence must have length 4, got 3.

[KumoLiu]

Hi @DorisxinDU, I see your questions. I think 8 is your batch size. The images and labels should be channel-first.

MONAI/monai/transforms/croppad/array.py

Lines 1236 to 1237 in 11745a6

	img: input data to crop samples from based on the ratios of every class, assumes `img` is a
	channel-first array.

from monai.transforms import RandCropByLabelClassesd
import numpy as np

cropper = RandCropByLabelClassesd(
    keys=["image", "label"],
    label_key="label",
    spatial_size=[3, 3],
    # ratios=[1, 2, 3, 1],   # optional
    num_classes=4,
    num_samples=2,
)
data = {
    "image": np.array([
        [[0.0, 0.3, 0.4, 0.2, 0.0],
        [0.0, 0.1, 0.2, 0.1, 0.4],
        [0.0, 0.3, 0.5, 0.2, 0.0],
        [0.1, 0.2, 0.1, 0.1, 0.0],
        [0.0, 0.1, 0.2, 0.1, 0.0]]
    ])[None],
    "label": np.array([
        [[0, 0, 0, 0, 0],
        [0, 1, 2, 1, 0],
        [0, 1, 3, 0, 0],
        [0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0]]
    ])[None],
}
result = cropper(data)

Hope it can help you, thanks! Just convert it to discussions.