Setting this flag allows the built-in auto-tuner cuDNN efficient algorithm to automatically find the most suitable for the current configuration, to solve the problem to optimize operating efficiency.
- If the input data on the type or dimension of the network is not changed, provided torch.backends.cudnn.benchmark = true operating efficiency can be increased;
- If the input data networks in each iteration are changes, it will lead to cuDNN every time again to find the optimal configuration, but this will reduce the operating efficiency.