can i solve cuda memory problem by setting densenet121.grad to be zero?

[attense]

I read about the paper and try to reproduce the simulation but cuda memory always haunts me.
I try to add

   for p1 in self.push_color_trunk.parameters():
            p1.requires_grad = False
        for p2 in self.push_depth_trunk.parameters():
            p2.requires_grad = False
        for p3 in self.grasp_color_trunk.parameters():
            p3.requires_grad = False
        for p4 in self.grasp_depth_trunk.parameters():
            p4.requires_grad = False

to reduce calculating
this actually works, the code can run now. But can somebody tell me whether this is right? cuz I'm not sure if this modification will destroy the training process