def _update()

in higher/optim.py [0:0]

• 28 lines of code
• 7 McCabe index (conditional complexity)

    def _update(self, grouped_grads: _GroupedGradsType, **kwargs) -> None:

        zipped = zip(self.param_groups, grouped_grads)
        for group_idx, (group, grads) in enumerate(zipped):
            rho, eps = group['rho'], group['eps']

            for p_idx, (p, g) in enumerate(zip(group['params'], grads)):
                if g is None:
                    continue

                if g.data.is_sparse:
                    raise RuntimeError(
                        'Adadelta does not support sparse gradients'
                    )
                state = self.state[group_idx][p_idx]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    state['square_avg'] = _torch.zeros_like(p.data)
                    state['acc_delta'] = _torch.zeros_like(p.data)

                square_avg, acc_delta = state['square_avg'], state['acc_delta']
                state['step'] += 1

                if group['weight_decay'] != 0:
                    g = _add(g, group['weight_decay'], p)

                square_avg = _addcmul(square_avg.mul(rho), 1 - rho, g, g)
                state['square_avg'] = square_avg
                std = _add(square_avg, eps).sqrt()
                delta = _add(acc_delta, eps).sqrt().div(std).mul(g)
                state['acc_delta'] = _addcmul(
                    acc_delta.mul(rho), 1 - rho, delta, delta
                )
                group['params'][p_idx] = _add(p, -group['lr'], delta)