Should the aspect ratio of test set same as train set?

vision
1

Hi, I trained my UNET model with images with size 320*320, I want to test(predict for new dataset) my model using a dataset with different image sizes, I heard that is possible but I got an error for mismatching the dimension sizes.
I have used model.eval() mode.
error:

torch.Size([1, 3, 813, 1024])
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
 in 
     21         img.cuda()
     22         print(img.shape)
---> 23         out = model(img)
     24         temp = torch.Tensor.cpu(x).detach().np()
     25         print(type(temp), temp.shape)

C:\Anaconda3\envs\torchgpu\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs)
    491             result = self._slow_forward(*input, **kwargs)
    492         else:
--> 493             result = self.forward(*input, **kwargs)
    494         for hook in self._forward_hooks.values():
    495             hook_result = hook(self, input, result)

 in forward(self, inputs)
     48 
     49         center = self.center(maxpool4)       # 256*32*32
---> 50         up4 = self.up_concat4(center,conv4)  # 128*64*64
     51         up3 = self.up_concat3(up4,conv3)     # 64*128*128
     52         up2 = self.up_concat2(up3,conv2)     # 32*256*256

C:\Anaconda3\envs\torchgpu\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs)
    491             result = self._slow_forward(*input, **kwargs)
    492         else:
--> 493             result = self.forward(*input, **kwargs)
    494         for hook in self._forward_hooks.values():
    495             hook_result = hook(self, input, result)

 in forward(self, high_feature, *low_feature)
     53         outputs0 = self.up(high_feature)
     54         for feature in low_feature:
---> 55             outputs0 = torch.cat([outputs0, feature], 1)
     56         return self.conv(outputs0)

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 100 and 101 in dimension 2 at ..\aten\src\TH/generic/THTensor.cpp:711

Please help, Thanks

2

Apparently your code has some restrictions on the input shapes, as the height of your current input is 100 and 101 for center and conv4.

Usually it is easier to deal with shapes of powers of two instead of completely random input shapes.
If you really want to use all possible input shapes, you could add some shape checks manually and use cropping or padding to create matching shapes for your concat layers.

3

@ptrblck What do you mean by deal with shapes of powers of two? can you explain please?
Thanks for your response.
Note that my work is image resizing so maintaining the aspect ratio is critical.