def _update()

in higher/optim.py [0:0]

• 32 lines of code
• 8 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):
            for p_idx, (p, g) in enumerate(zip(group['params'], grads)):
                if g is None:
                    continue

                if g.is_sparse:
                    raise RuntimeError(
                        'Rprop does not support sparse gradients'
                    )

                state = self.state[group_idx][p_idx]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    state['prev'] = _torch.zeros_like(p.data)
                    state['step_size'] = g.new().resize_as_(g).fill_(
                        group['lr']
                    )

                etaminus, etaplus = group['etas']
                step_size_min, step_size_max = group['step_sizes']
                step_size = state['step_size']

                state['step'] += 1

                sign = g.mul(state['prev']).sign()
                sign[sign.gt(0)] = etaplus
                sign[sign.lt(0)] = etaminus
                sign[sign.eq(0)] = 1

                # update stepsizes with step size updates
                step_size = step_size.mul(sign).clamp(
                    step_size_min, step_size_max
                )
                state['step_size'] = step_size

                # for dir<0, dfdx=0
                # for dir>=0 dfdx=dfdx
                g = _torch.where(sign.eq(etaminus), _torch.zeros_like(g), g)

                # update parameters
                group['params'][p_idx] = _addcmul(p, -1, g.sign(), step_size)

                state['prev'] = g.clone()