def step()

in timm/optim/lars.py [0:0]

• 42 lines of code
• 13 McCabe index (conditional complexity)

    def step(self, closure=None):
        """Performs a single optimization step.

        Args:
            closure (callable, optional): A closure that reevaluates the model and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            weight_decay = group['weight_decay']
            momentum = group['momentum']
            dampening = group['dampening']
            nesterov = group['nesterov']
            trust_coeff = group['trust_coeff']
            eps = group['eps']

            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad

                # apply LARS LR adaptation, LARC clipping, weight decay
                # ref: https://github.com/NVIDIA/apex/blob/master/apex/parallel/LARC.py
                if weight_decay != 0 or group['always_adapt']:
                    w_norm = p.norm(2.0)
                    g_norm = grad.norm(2.0)
                    trust_ratio = trust_coeff * w_norm / (g_norm + w_norm * weight_decay + eps)
                    # FIXME nested where required since logical and/or not working in PT XLA
                    # Set the ratio to 1.0 (no change) if either weight norm or grad norm is zero
                    trust_ratio = torch.where(
                        w_norm > 0,
                        torch.where(g_norm > 0, trust_ratio, 1.0),
                        1.0,
                    )
                    if group['trust_clip']:
                        trust_ratio = torch.clamp(trust_ratio / group['lr'], max=1.0)
                    grad.add_(p, alpha=weight_decay)
                    grad.mul_(trust_ratio)

                # apply SGD update https://github.com/pytorch/pytorch/blob/1.7/torch/optim/sgd.py#L100
                if momentum != 0:
                    param_state = self.state[p]
                    if 'momentum_buffer' not in param_state:
                        buf = param_state['momentum_buffer'] = torch.clone(grad).detach()
                    else:
                        buf = param_state['momentum_buffer']
                        buf.mul_(momentum).add_(grad, alpha=1. - dampening)
                    if nesterov:
                        grad = grad.add(buf, alpha=momentum)
                    else:
                        grad = buf

                p.add_(grad, alpha=-group['lr'])

        return loss