Using your model on multiple image pair batch will cause drastically different output

[Animadversio]

Input the image in a 2 image batches


Input the same image in a single image batch

Result in very different flow outputs.
Using the model chairs-things.pytorch and sintel.pytorch. I didn't find a module that introduce interaction between batch in your model...
Do you understand why that is the case?

[Animadversio]

I detailed debug the code on 2 computers and it seems it's about the default behavior of a pytorch function torch.nn.functional.grid_sample before 1.4.0 the default is to align_corner=True, at 1.4.0 this behavior is changed to be align_corner=False. so making the option explicit will retain the model's performance, which is trained with the older option.

See this pytorch/pytorch#23923

[Animadversio]

This phenomenon is different from machine to machine, seems in one GPU the correlation layer ouput is wrong for Batch size >1 but on the other one it's correct.
Seems it's related to the compiled cuda code