Train with different image size

[puhuk]

Hi, Thanks for the great work!

I'm trying to train with different image size and change relevant code (input size, input_nc for initializing FeatureRegression in class GMM.
But the result seems like the picture.

Could you help me what should I need to correct for different image size.

• Left image is input and right image is warped from GMM

[thaithanhtuan]

The GMM network is like: input --> correlation --> transformation parameters TPS
Then output = TPS transformation (input, transformation parameters TPS )
One easy way is you can increase the resolution of 'transformation parameters TPS" so that it can transform the high-resolution input. In this way, you don't need to change the network input size.
I've tried changing the image size of the input, but --> calculated correlation is broken out of memory.

[anzwolf]

Hi can you please point out the lines of code where to make those changes as per your line that you mentioned below
"One easy way is you can increase the resolution of 'transformation parameters TPS" so that it can transform the high-resolution input. In this way, you don't need to change the network input size."
Detailed explanation will be welcomed

[sudip550]

The GMM network is like: input --> correlation --> transformation parameters TPS Then output = TPS transformation (input, transformation parameters TPS ) One easy way is you can increase the resolution of 'transformation parameters TPS" so that it can transform the high-resolution input. In this way, you don't need to change the network input size. I've tried changing the image size of the input, but --> calculated correlation is broken out of memory.

hii sir, did you have pretrained model with more than 2,00,000 steps. please share if you have.