Dimension issues for Swin UNETR implementation

[ghost]

Hello,
I've been trying to implement a Swin UNETR model, using the pretrained weights for BTCV segmentation and the tutorial available on this repository.
However, at the training step, I always run into this error: ValueError: spatial dimensions [2, 3, 4] of input image (spatial shape: torch.Size([73, 73, 34])) must be divisible by 2**5.
The thing is I don't understand why [73 73 34] is the shape of my data, since I used RandCropByPosNegLabeld, with spatial_size = (96, 96, 96)

I tried to resize my data by using Resized in the transforms, but it makes my environment crash...
I then try to resize them later on, in the training process with the line: x = resize_transform(x), with resize_transform = Resized(spatial_dim=(64, 64, 32). This gets me the an error saying that the dimensions of spatial _dim and the one of my image do not match.
Is there a way to fix this?

[KumoLiu]

Hi @bintadkr, could you please paste your transform chain?
And I would suggest you check the data shape after each transform to ensure that you know what shape you get from that transform.

Hope it helps, thanks.

[ghost]

Hi @KumoLiu, thank you for your help :)
Here is my transform chain:
num_samples = 4

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

train_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys=["image", "label"]),
CropForegroundd(keys=["image", "label"], source_key="image"),
RandCropByPosNegLabeld(
keys=["image", "label"],
label_key="label",
spatial_size=(96, 96, 96),
pos=1,
neg=1,
num_samples=4,
image_key="image",
image_threshold=0,
),
Spacingd(
keys=["image", "label"],
pixdim=(1.5, 1.5, 2.0),
mode=("bilinear", "nearest")
),
EnsureTyped(keys=["image", "label"], device=device, track_meta=True),
]
)
output_shape = (1, 407, 407, 68)
val_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys=["image", "label"]),
#Resized(keys=["label"], spatial_size = [407, 407, 68]),
CropForegroundd(keys=["image", "label"], source_key="image"),
Spacingd(
keys=["image", "label"],
pixdim=(1.5, 1.5, 2.0),
mode=("bilinear", "nearest")
),
EnsureTyped(keys=["image", "label"], device=device, track_meta=True),
]
)

resize_transforms = Resize(spatial_size = (64, 64, 32))

[ericspod]

CropForegroundd could be cropping much too tightly for some instances. You can add SpatialPadd right after it with (96,96,96) for the spatial size to ensure this isn't happening.