Policy Gradient Reinforcement Learning with Knet

[etotheipluspi]

I'm trying to implement a policy gradient method using Knet. To do this, I need to multiply the loss gradients from by a scalar before doing backprop on the rest of the net. Is there a recommended way to do that through the Knet API? If there is a way to run a part of the computational graph that would do the job as well.

[rejuvyesh]

Is gscale in update! not enough?

[etotheipluspi]

That should work thanks!