428516056abe41f135133e732a8d44af6ce9a234,rllib/policy/torch_policy.py,TorchPolicy,compute_gradients,#TorchPolicy#Any#,264

Before Change

        loss_out = self._loss(self, self.model, self.dist_class, train_batch)
        self._optimizer.zero_grad()
        loss_out.backward()

        grad_process_info = self.extra_grad_process()

        // Note that return values are just references;

After Change

        grad_process_info = {}
        grads = []
        for i, opt in enumerate(self._optimizers):
            opt.zero_grad()
            loss_out[i].backward()
            grad_process_info = self.extra_grad_process(opt, loss_out[i])

            // Note that return values are just references;
            // calling zero_grad will modify the values
            for param_group in opt.param_groups:
                for p in param_group["params"]:
                    if p.grad is not None:
                        grads.append(p.grad.data.cpu().numpy())
                    else:
                        grads.append(None)

        grad_info = self.extra_grad_info(train_batch)
        grad_info.update(grad_process_info)
        return grads, {LEARNER_STATS_KEY: grad_info}

Instances