unet error with (1,3,256,256) input

[3bobo]

I'm try to input a tensor with the size of (1, 3, 256, 256), after many conv2d and maxpool, the centre feature size is (1, 256, 8, 8), and the covn4 is (1, 128, 24, 24), after up_layer, the feature size is (1, 128, 12, 12), and after up is (1,64,24,24), but the conv3 is (1, 64, 57, 57), and if this size goes with cropping, it will be (1,64,23,23), which is different from (1,64,24,24), and makes an error.

[Qirui-Y]

Do you solve this problem? Can you share yourself config file with Unet

[hurricane2018]

Same Problem

[flyking1994]

Do you solve this problem? I have the same problem.

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 24 and 23 in dimension 2 at /pytorch/aten/src/THC/generic/THCTensorMath.cu:71

[brucemuller]

Do you solve this problem? I have the same problem.

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 24 and 23 in dimension 2 at /pytorch/aten/src/THC/generic/THCTensorMath.cu:71

Are you applying unet to a square input or rectangular? I had to alter the model to work for validating on rectangular input

[flyking1994]

Do you solve this problem? I have the same problem. RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 24 and 23 in dimension 2 at /pytorch/aten/src/THC/generic/THCTensorMath.cu:71

Are you applying unet to a square input or rectangular? I had to alter the model to work for validating on rectangular input

My image size is (256,256,3), its a square input

[flyking1994]

def forward(self, inputs1, inputs2):
    outputs2 = self.up(inputs2)
    offset1 = outputs2.size()[2] - inputs1.size()[2]
    offset2 = inputs1.size()[2] - outputs2.size()[2]
    padding = 2 * [offset1//2, -(offset2//2)]  
    outputs1 = F.pad(inputs1, padding)
    return self.conv(torch.cat([outputs1, outputs2], 1))

I changed the calculation method of padding. It seems that this problem has been solved.

[svanbodegraven]

Thanks @flyking1994, +1 here for Cityscapes datasets gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip, cropped and resized to (224, 224). Will try a run later with the original size of (2048, 1024)