How to ensure gradient passing of subnetwork after using the subnetwork training result as filters parameters for calculation

[mxteaw]

I tried to make the CNN_PP module using pytorch, but since the subnet and yolo connection need to use the results of the subnetwork as parameters of the filters, it seems to causes the gradient propagation to break and the program reports an error, can someone handle the problem?

[mxteaw]

I tried to make the CNN_PP module using pytorch, but since the subnet and yolo connection need to use the results of the subnetwork as parameters of the filters, it seems to causes the gradient propagation to break and the program reports an error, can someone handle the problem?

Since the formula is derivable, now I know that it does not affect the passing of the gradient, but the running program will report the error: "RuntimeError: Only Tensors created explicitly by the user (graph leaves) support the deepcopy protocol at the moment", how should I solve it?