Function construct_optimizer

slowfast/models/optimizer_backbone_special.py:14–141

Construct a stochastic gradient descent or ADAM optimizer with momentum. Details can be found in: Herbert Robbins, and Sutton Monro. "A stochastic approximation method." and Diederik P.Kingma, and Jimmy Ba. "Adam: A Method for Stochastic Optimization." Args: mod

(model, cfg)

def construct_optimizer(model, cfg):
    """
    Construct a stochastic gradient descent or ADAM optimizer with momentum.
    Details can be found in:
    Herbert Robbins, and Sutton Monro. "A stochastic approximation method."
    and
    Diederik P.Kingma, and Jimmy Ba.
    "Adam: A Method for Stochastic Optimization."

    Args:
        model (model): model to perform stochastic gradient descent
        optimization or ADAM optimization.
        cfg (config): configs of hyper-parameters of SGD or ADAM, includes base
        learning rate,  momentum, weight_decay, dampening, and etc.
    """
    bn_parameters = []
    non_bn_parameters = []
    zero_parameters = []
    special_bn_parameters = []
    special_non_bn_parameters = []
    special_zero_parameters = []
    skip = {}
    if hasattr(model, "no_weight_decay"):
        skip = model.no_weight_decay()

    special_list = cfg.SOLVER.SPECIAL_LIST
    special_ration = cfg.SOLVER.SPECIAL_RATIO
    logger.info(f'Special parameter list: {special_list}')
    logger.info(f'LR Ration for special parameter is {special_ration}')

    total_num = 0
    for name, m in model.named_modules():
        is_bn = isinstance(m, torch.nn.modules.batchnorm._NormBase)
        for p in m.parameters(recurse=False):
            if not p.requires_grad:
                continue
            total_num += 1
            if is_bn:
                flag = False
                for s in special_list:
                    if s in name:
                        special_bn_parameters.append(p)
                        flag = True
                        break
                if not flag:
                    bn_parameters.append(p)
            elif name in skip or (
                (len(p.shape) == 1 or name.endswith(".bias"))
                and cfg.SOLVER.ZERO_WD_1D_PARAM
            ):
                flag = False
                for s in special_list:
                    if s in name:
                        special_zero_parameters.append(p)
                        flag = True
                        break
                if not flag:
                    zero_parameters.append(p)