RuntimeError: Sizes of tensors must match except in dimension 1. Got 47 and 48 in dimension 4 (The offending index is 1)

[talhaanwarch]

Here is the model
from monai.networks.nets import UNet
from monai.networks.layers import Norm

model = UNet(
            dimensions=3,
            in_channels=1,
            out_channels=5,
            channels=(16, 32, 64, 128, 256),
            strides=(2, 2, 2, 2),
            num_res_units=2,
            norm=Norm.BATCH)

When I pass the data

import torch
x=torch.randn(1, 1, 256, 256,94)
out=model(x)
out.shape

I got this error
RuntimeError: Sizes of tensors must match except in dimension 1. Got 47 and 48 in dimension 4 (The offending index is 1)

I don't think padding 94 to 96 is a solution. because I want output shape to be (1, 1, 256, 256,94)

[Nic-Ma]

Hi @talhaanwarch ,

Thanks for your interest here.
Although you don't want to pad 94 to 96, it is the root cause.
Because here you set channels=(16, 32, 64, 128, 256) in UNet, so the input size will be divided 4 times by 2 in the net, so at least the size should be N x 16.
@ericspod @wyli please feel free to correct me if I misunderstand the root cause.

Thanks.

[talhaanwarch]

i want to do the same as in the spline tutorial, but I want to keep the predicted mask (segment) shape as the input shape, let say (512,512,58)
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/spleen_segmentation_3d_lightning.ipynb

[rijobro]

You shouldn't need to invert to be able to calculate the loss. During training you can pad both the input image and the ground truth segmentation. Since they both get padded by the same amount, the loss function can be calculated directly using the padded ground truth segmentation and the output of your network. If you decide you want to save your results to file, for example, then you can think about doing the inverse at that point.

[rijobro]

I haven't spent much time looking at it, but I imagine the notebook you just linked would work just fine if we added a DivisiblePad to the list of transforms.

[talhaanwarch]

thanks @rijobro i am confused what should be the value of K
DivisiblePadd(keys=["image", "label"],k=...),

[rijobro]

The stride of your convolutions will determine the step of downsampling at each layer. If your network definition has strides=(2, 2, 2, 2), then your images need to be divisible by 2^4=16. Hope that helps.