Performance Drop in Stereo Matching Network After PyTorch Version Update

I have been training the same stereo matching network with different versions of PyTorch and noticed a significant performance decrease with PyTorch 1.13 compared to PyTorch 1.2. According to the original author’s response, this is due to differences in the implementation of the interpolate function between versions. Specifically, my network contains an operation like this:


F.interpolate(inputdata, [d, h, w], mode='trilinear', align_corners=False)

Where inputdata is a five-dimensional tensor. I would like to implement the interpolate function from PyTorch 1.2 myself, but upon inspecting the code, I found that torch._c has been encapsulated and is inaccessible.

Does anyone have advice on how I could proceed to either access this function or an alternative way to replicate the PyTorch 1.2 behavior? Your assistance would be greatly appreciated. Thank you!