Overflow Exception creating simple nn module

[peteryang1]

Hello,

When I use the latest code to build the LibTorch and Torchsharp, a overflow exception is raised when creating a simple Conv2d module.

My simple code:

var convTmp = torch.nn.Conv2d(16, 16, 3);

error:


It's weird that the size of the parameters here returned a very big number.
call stack:

After digging into the call stack, seems the "named_parameters" of default module created by libtorch returned a very large map and the size of this map raised a overflow exception.

I was running the code on windows with cuda11.1 and libtorch 1.10.0.

Hope to get any help on this error, thanks!

[peteryang1]

Seems it's because I compile the LibTorch with latest download libtorch libraries, but the runtime is downloaded from nuget.

But it's weird the I downloaded the same version as the libtorch described in nuget. But it's fine to fix this issue.