Class LARS

utils.py:533–571

Almost copy-paste from https://github.com/facebookresearch/barlowtwins/blob/main/main.py

class LARS(torch.optim.Optimizer):
    """
    Almost copy-paste from https://github.com/facebookresearch/barlowtwins/blob/main/main.py
    """
    def __init__(self, params, lr=0, weight_decay=0, momentum=0.9, eta=0.001,
                 weight_decay_filter=None, lars_adaptation_filter=None):
        defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum,
                        eta=eta, weight_decay_filter=weight_decay_filter,
                        lars_adaptation_filter=lars_adaptation_filter)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self):
        for g in self.param_groups:
            for p in g['params']:
                dp = p.grad

                if dp is None:
                    continue

                if p.ndim != 1:
                    dp = dp.add(p, alpha=g['weight_decay'])

                if p.ndim != 1:
                    param_norm = torch.norm(p)
                    update_norm = torch.norm(dp)
                    one = torch.ones_like(param_norm)
                    q = torch.where(param_norm > 0.,
                                    torch.where(update_norm > 0,
                                                (g['eta'] * param_norm / update_norm), one), one)
                    dp = dp.mul(q)

                param_state = self.state[p]
                if 'mu' not in param_state:
                    param_state['mu'] = torch.zeros_like(p)
                mu = param_state['mu']
                mu.mul_(g['momentum']).add_(dp)

                p.add_(mu, alpha=-g['lr'])


class MultiCropWrapper(nn.Module):