Inconsistent size between output and target

[yftzz]

Hi, congrats on your great work.

I was trying to experimenting with your proposed code on a dataset other than cityscape, where I set the input image shape to be 512x512. But I see that with your default settings for the network, the output has a shape of 128x128, so do I have to add the code for upsampling manually based on your implementation?
I might me dumb somehow, but I don't see where to adjust the output shape.

Regards.

[Fly2flies]

I have the same confusion !
Before the Bottleneck, 2 convolution layers with stride 2 make the input down sample 4x. Should I up sample 4x after the final layer to recover the input size or change the stride of the first two convolution layer into 1 ?

[yftzz]

I guess you might wanna check out other issues, the similar questions have been asked for a few times. But, my question is, put the bilinear upsampling layer before the final layer or after? cuz I believe if the final output comes from a bilinear upsampling layer, the output has to be a bit coarse.

[DragoniteDD]

Based on the author's reply here, it seems to be bilinear interpolation upsampling after the output
#15 (comment)

[Fly2flies]

Thanks for your reply. I've seen the issue. But it seems to be a little coarse to bilinear 4x to predict directly. I want to know whether you have some solutions to address it?

[sidml]

If anyone else is wondering, how to train the model if the model output size is not equal to label size, check out this

[NickLucche]

Hey, another place where you might want to take a look at is the initial strided convolution "block"

HRNet-Semantic-Segmentation/lib/models/seg_hrnet.py

Lines 414 to 419 in 77044e6

	x = self.conv1(x)
	x = self.bn1(x)
	x = self.relu(x)
	x = self.conv2(x)
	x = self.bn2(x)
	x = self.relu(x)

which effectively downscales the original input resolution; although editing it should be considered an architectural change, you can get your input resolution to stay constant.