Hi there everyone,
Me and my project partners are working on modifying existing pytorch functionalities.
We definitely got stuck at one point, which I will mention concisely.
When we call Conv2d() function in python script, it basically calls a class in conv.py and returns an object. Then, we create a torch.tensor in python code and feed it again to created Conv2d object and in some magical way, convolution is done. We were not able to find the exact interface between cpp codes and the python script. Tracing from conv.py, we are able to see that def forward() function for Conv2d calls F.conv2d() function, which resides in functionality.py. But then, we cannot see how it exactly calls cpp functions from convolution.cpp.
Can you please help us at that point? Since we need to modify the code, we have to be able to pass variables throughout whole pipeline, this is why we need this information.
P.S We are also aware of .yaml files but did not show us the concrete interface.