RuntimeError: applying transform RandCropByPosNegLabeld

[dianemarquette]

Hi :)

Thank you for the great package. The detailed documentation is also very much appreciated !

I have successfully run the Spleen 3D Segmentation notebook with the provided data. I'm now trying to get it to work with my custom dataset (CT scans of hands in NIfTI format with 3 classes: background, trapezium & metacarpial bone of the thumb).

My problem
Whereas the first cells execute smoothly, the "typical Pytorch training process" cell gives me the following error:

---

epoch 1/600
1/11, train_loss: 0.8117
2/11, train_loss: 0.8081
3/11, train_loss: 0.8139

ValueError Traceback (most recent call last)
~\anaconda3\envs\kerimedical\lib\site-packages\monai\transforms\utils.py in apply_transform(transform, data, map_items)
360 return [transform(item) for item in data]
--> 361 return transform(data)
362 except Exception as e:

~\anaconda3\envs\kerimedical\lib\site-packages\monai\transforms\croppad\dictionary.py in call(self, data)
534
--> 535 self.randomize(label, fg_indices, bg_indices, image)
536 assert isinstance(self.spatial_size, tuple)

~\anaconda3\envs\kerimedical\lib\site-packages\monai\transforms\croppad\dictionary.py in randomize(self, label, fg_indices, bg_indices, image)
519 if fg_indices is None or bg_indices is None:
--> 520 fg_indices_, bg_indices_ = map_binary_to_indices(label, image, self.image_threshold)
521 else:

~\anaconda3\envs\kerimedical\lib\site-packages\monai\transforms\utils.py in map_binary_to_indices(label, image, image_threshold)
215 img_flat = np.any(image > image_threshold, axis=0).ravel()
--> 216 bg_indices = np.nonzero(np.logical_and(img_flat, ~label_flat))[0]
217 else:

ValueError: operands could not be broadcast together with shapes (778998,) (0,)

The above exception was the direct cause of the following exception:

RuntimeError Traceback (most recent call last)
in
14 epoch_loss = 0
15 step = 0
---> 16 for batch_data in train_loader:
17 step += 1
18 inputs, labels = (

~\anaconda3\envs\kerimedical\lib\site-packages\torch\utils\data\dataloader.py in next(self)
515 if self._sampler_iter is None:
516 self._reset()
--> 517 data = self._next_data()
518 self._num_yielded += 1
519 if self._dataset_kind == _DatasetKind.Iterable and \

~\anaconda3\envs\kerimedical\lib\site-packages\torch\utils\data\dataloader.py in _next_data(self)
555 def _next_data(self):
556 index = self._next_index() # may raise StopIteration
--> 557 data = self._dataset_fetcher.fetch(index) # may raise StopIteration
558 if self._pin_memory:
559 data = _utils.pin_memory.pin_memory(data)

~\anaconda3\envs\kerimedical\lib\site-packages\torch\utils\data_utils\fetch.py in fetch(self, possibly_batched_index)
42 def fetch(self, possibly_batched_index):
43 if self.auto_collation:
---> 44 data = [self.dataset[idx] for idx in possibly_batched_index]
45 else:
46 data = self.dataset[possibly_batched_index]

~\anaconda3\envs\kerimedical\lib\site-packages\torch\utils\data_utils\fetch.py in (.0)
42 def fetch(self, possibly_batched_index):
43 if self.auto_collation:
---> 44 data = [self.dataset[idx] for idx in possibly_batched_index]
45 else:
46 data = self.dataset[possibly_batched_index]

~\anaconda3\envs\kerimedical\lib\site-packages\monai\data\dataset.py in getitem(self, index)
531 if start_run or isinstance(_transform, Randomizable) or not isinstance(_transform, Transform):
532 start_run = True
--> 533 data = apply_transform(_transform, data)
534 return data
535

~\anaconda3\envs\kerimedical\lib\site-packages\monai\transforms\utils.py in apply_transform(transform, data, map_items)
361 return transform(data)
362 except Exception as e:
--> 363 raise RuntimeError(f"applying transform {transform}") from e
364
365

RuntimeError: applying transform <monai.transforms.croppad.dictionary.RandCropByPosNegLabeld object at 0x00000277E4C14F10>

---

What didn't solve the problem
I've tried to play with the spatial_size of the RandCropByPosNegLabeld transform but it didn't help. I don't understand what the "ValueError: operands could not be broadcast together with shapes (778998,) (0,)" implies.

Any help or suggestion would be greatly appreciated :)
Thanks for reading!

[Nic-Ma]

Hi @dianemarquette ,

I feel it's something related to numpy version. Maybe you can update your numpy and try again?
@wyli , dou you have any more suggestion?

Thanks.

[wyli]

could you please share the input parameters? I couldn't replicate this issue -- looks like label_flat's shape is (0,)

[dianemarquette]

Thank you for your suggestions.

@Nic-Ma I'm already using the latest version of numpy (1.20.1). So, it's probably not due to this.

@wyli I'm using the tutorial for the spleen 3D segmentation. I didn't change a lot of code. The only two cells I changed are the ones defining the transforms and the model (modifications are in bold).

Transform cell

train_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
AddChanneld(keys=["image", "label"]),
Spacingd(keys=["image", "label"], pixdim=(
0.4, 0.4, 0.625), mode=("bilinear", "nearest")), # 1.5, 1.5, 2
Orientationd(keys=["image", "label"], axcodes="RAS"),
DataStatsd(keys=['image', 'label']), # added
ScaleIntensityRanged(
keys=["image"], a_min=200, a_max=2000,
b_min=0.0, b_max=1.0, clip=True,
),
CropForegroundd(keys=["image", "label"], source_key="image"),
RandCropByPosNegLabeld(
keys=["image", "label"],
label_key="label",
# spatial_size=(96, 96, 96),
spatial_size=(16,16,16),
pos=1,
neg=1,
num_samples=4,
image_key="image",
image_threshold=0,
),
# user can also add other random transforms
# RandAffined(
# keys=['image', 'label'],
# mode=('bilinear', 'nearest'),
# prob=1.0, spatial_size=(96, 96, 96),
# rotate_range=(0, 0, np.pi/15),
# scale_range=(0.1, 0.1, 0.1)),
ToTensord(keys=["image", "label"]),
]
)
val_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
AddChanneld(keys=["image", "label"]),
Spacingd(keys=["image", "label"], pixdim=(
0.4, 0.4, 0.625), mode=("bilinear", "nearest")), # 1.5, 1.5, 2
Orientationd(keys=["image", "label"], axcodes="RAS"),
DataStatsd(keys=['image', 'label']), # added
ScaleIntensityRanged(
keys=["image"], a_min=200, a_max=2000,
b_min=0.0, b_max=1.0, clip=True,
),
CropForegroundd(keys=["image", "label"], source_key="image"),
ToTensord(keys=["image", "label"]),
]
)

Model cell

device = torch.device("cuda:0")
model = UNet(
dimensions=3,
in_channels=1,
out_channels=3, # 3 classes = background + trapezium + metacarpial
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
norm=Norm.BATCH,
).to(device)
loss_function = DiceLoss(to_onehot_y=True, softmax=True)
optimizer = torch.optim.Adam(model.parameters(), 1e-4)

The output of the DataLoader looks ok to me:

However, I reckon that the fact that label_flat's shape is (0,) is strange. I still didn't find a solution, so any ideas would be extremely helpful !

Thank you for your time.

[dianemarquette]

I ended up understanding the cause of this error. I'm writing it here in case it can help someone else.

The problem didn't come from the code. It was the dataset. The labels of two CT scans must have been incorrectly exported. As a result, the dimensions of the labels didn't match the dimensions of the scans (for scans #6 and #10).

Deleting these mislabeled scans from my dataset solved the problem :).

Cheers,

Diane