73b2221b5eb4fd21802e6bf41e21d5df8ef9bf2c,src/sdk/pynni/nni/nas/pytorch/darts/trainer.py,DartsTrainer,_compute_hessian,#DartsTrainer#Any#Any#Any#Any#,119

Before Change

                    p += eps * d

            self.mutator.reset()
            loss = self.loss(self.model(trn_X), trn_y)
            if e > 0:
                dalpha_pos = torch.autograd.grad(loss, self.mutator.parameters())  // dalpha { L_trn(w+) }
            elif e < 0:
                dalpha_neg = torch.autograd.grad(loss, self.mutator.parameters())  // dalpha { L_trn(w-) }

After Change

            _, loss = self._logits_and_loss(trn_X, trn_y)
            dalphas.append(torch.autograd.grad(loss, self.mutator.parameters()))

        dalpha_pos, dalpha_neg = dalphas  // dalpha { L_trn(w+) }, // dalpha { L_trn(w-) }
        hessian = [(p - n) / 2. * eps for p, n in zip(dalpha_pos, dalpha_neg)]
        return hessian

Instances