HELP with this error: stack expects each tensor to be equal size, but got [1, 32, 32, 16] at entry 0 and [1, 32, 32, 15] at entry 7

[vlgonz]

Hi, I have the title error and I do not know how to solve it. This is MY code:
keys=("image", "label")

xforms = [
LoadImaged(keys, ensure_channel_first=True, image_only=True),
Orientationd(keys, axcodes="LPS"),
CropForegroundd(keys, source_key=keys[0]),
Spacingd(keys, pixdim=(1.0, 1.0, 3.0), mode=("bilinear", "nearest")[: len(keys)]),
ScaleIntensityRanged(keys[0], a_min=-1000.0, a_max=500.0, b_min=0.0, b_max=1.0, clip=True),
SpatialPadd(keys, spatial_size=(32, 32, 16), mode="constant"),
]
if mode == "train":
xforms.extend(
[
RandCropByPosNegLabeld(keys, label_key=keys[1], spatial_size=(32, 32, 16), num_samples=2),
RandGaussianNoised(keys[0], prob=0.15, std=0.01),
RandFlipd(keys, spatial_axis=0, prob=0.5),
RandFlipd(keys, spatial_axis=1, prob=0.5),
RandFlipd(keys, spatial_axis=2, prob=0.5),
]
Why do I get that error if I have added SpatialPadd with a spatial size of (32,32,16) before RandCropByPosNegLabeld? Thank you

[ericspod]

I'm not sure I can duplicate the issue here. What size are your tensors just before SpatialPadd? In your CropForegroundd definition I think you want source_key=keys[1] instead? I have the following which works, see if there's anything you can change to reproduce your problem:

import monai, torch
import monai.transforms as mt

img = torch.rand(1, 32, 32, 17)
lbl = torch.zeros_like(img)
lbl[0, 10:20, 10:20, 4:12] = 1  # one small label
data = {"image": img, "label": lbl}

keys = ("image", "label")
com = mt.Compose(
    [
        mt.CropForegroundd(keys, source_key=keys[1]),
        mt.SpatialPadd(keys, spatial_size=(32, 32, 16), mode="constant"),
        mt.RandCropByPosNegLabeld(keys, label_key=keys[1], spatial_size=(32, 32, 16), num_samples=2),
        mt.RandGaussianNoised(keys[0], prob=0.15, std=0.01),
        mt.RandFlipd(keys, spatial_axis=0, prob=0.5),
        mt.RandFlipd(keys, spatial_axis=1, prob=0.5),
        mt.RandFlipd(keys, spatial_axis=2, prob=0.5),
    ]
)
ds = monai.data.Dataset([data], com)
dl = monai.data.DataLoader(ds)
res = monai.utils.first(dl)
print(res["image"].shape, res["label"].shape)  # prints torch.Size([2, 1, 32, 32, 16]) torch.Size([2, 1, 32, 32, 16])

[vlopezgz]

Hi, thanks for your response. Originally, each tensor had dimensions of (512, 512, z), but after applying CropForeground, each tensor has a different size due to variations in the background of each image. I opted not to use CropForeground with source_key=keys[1] because it creates a patch of the region I intend to segment, which is not my objective. Instead, I aim to eliminate the black areas surrounding the main body, ensuring that the input data to the network does not include a large black portion. To achieve this, I removed CropForeground and applied CenterSpatialCrop with a fixed size using roi_size=(460, 460, -1). This adjustment resolved the error, although it did not eliminate the black parts from all images.

[ericspod]

I'm still not sure how this error is coming up for you. CropForegroundd will remove the black areas around your image which will possibly be resized by Spacingd, but then SpatialPadd should ensure that the image is no smaller than (32,32,16). This is what I get in my example if you play with the input sizes. RandCropByPosNegLabeld should also then get an input that's at least that size and be able to find samples of the right size. I can only suggest using my example code with different combinations of transforms to replicate your setup and with your data to see what reproduces the problem.

[AyaAl1ah]

Hi I think I have a similar issue. I changed a few things like spatial_size=(30, 30, 30) and img_size=(30, 30, 30) in the model as well as in the evaluation function. Also I have two classes and changed it out_channels=2,
img_size=(30, 30, 30), and also the feature size
feature_size=10
the out put:
Training (X / X Steps) (loss=X.X): 0%| | 0/20 [00:00<?, ?it/s]

Image size: torch.Size([4, 1, 30, 30, 30]) # i printed the shapes in the training loop
Label size: torch.Size([4, 1, 30, 30, 30])

Training (X / X Steps) (loss=X.X): 0%| | 0/20 [00:06<?, ?it/s]
RuntimeError Traceback (most recent call last)
in <cell line: 70>()
69 metric_values = []
70 while global_step < max_iterations:
---> 71 global_step, dice_val_best, global_step_best = train(global_step, train_loader, dice_val_best, global_step_best)
72 model.load_state_dict(torch.load(os.path.join(root_dir, "best_metric_model.pth")))

8 frames
/usr/local/lib/python3.10/dist-packages/torch/_tensor.py in torch_function(cls, func, types, args, kwargs)
1441
1442 with _C.DisableTorchFunctionSubclass():
-> 1443 ret = func(*args, **kwargs)
1444 if func in get_default_nowrap_functions():
1445 return ret

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 16 but got size 30 for tensor number 1 in the list. I get this error and nothing seems to fix it. sorry if I am bothering you but I am still learning thanks.