I saw some transfer learning Libtorch codes using pretrained torchscript models and all of them is only training additional layers added in Libtorch codes.
(e.g) Transfer-Learning-Dogs-Cats-Libtorch/main.cpp at master · krshrimali/Transfer-Learning-Dogs-Cats-Libtorch · GitHub)
My question is, can we train the torchscript model without defining network model in Libtorch?
I think torchscript model has already model’s weights and structure is similar to torch::nn::Module(actually it is instance of torch::jit::Module object) so I assume it’s not impossible…but still can’t find the way.
Best Regards