I’m doing neuroevolution. Since I would only be doing inference with the model, and manipulating the models’ weight “manually”, do I need to do something else but declaring torch.autograd.set_grad_enabled(False) to avoid unnecessary computations? Is there anything else that I can do to improve performance?
Another point is that currently, I’m using Python, for convenience. I’m also a C++ developer so believe that in this situation (reinforcement learning scenarios) C++ would provide much better performance since it is required much computation out of the Pytorch framework. Is there anyone out there with experience (or material) to share with me? I never used Pytorch API with C++ btw.