How the loss back propagate when using multiple gpus for extract task

[gaozhiguang]

Hi, i read your code, and i have some questions, how the loss back propagate when using multiple gpus for extract task,here is your code:

Here, the loss back propagation is just the loss back propagation in a process, and the distributed module is not used to add the loss of each process running on different gpu, it's a little diferent from your code in abs task.