torch.nn.utils.clip_grad_norm_()函数详解

torch.nn.utils.clip_grad_norm_()函数详解

torch.nn.utils.clip_grad_norm(parameters, max_norm, norm_type=2)

• 这个函数的主要目的是对$parameters$里的所有参数的梯度进行规范化;
• $parameters$为要规范的参数，$max\_norm$为所有参数的梯度的范数的上界，$norm\_type$为范数的类型，当$norm\_type=inf$时，代表无穷范数;
• 设$parameters$里所有参数的梯度的范数为$total\_norm$，
  若$max\_norm>total\_norm，$$parameters$里面的参数的梯度不做改变;
  若$max\_norm<total\_norm，$$parameters$里面的参数的梯度都要乘以一个系数$clip\_coef$，$$clip\_coef=max\_norm/(total\_norm+1e-6)$$
• 附上源码：（注意源码中的这句话"The norm is computed over all gradients together"——计算所有梯度的范数）

def clip_grad_norm_(parameters, max_norm, norm_type=2):
    r"""Clips gradient norm of an iterable of parameters.
    The norm is computed over all gradients together, as if they were
    concatenated into a single vector. Gradients are modified in-place.
    Arguments:
        parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a
            single Tensor that will have gradients normalized
        max_norm (float or int): max norm of the gradients
        norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for
            infinity norm.
    Returns:
        Total norm of the parameters (viewed as a single vector).
    """
    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]
    #第一步
    parameters = list(filter(lambda p: p.grad is not None, parameters))
    max_norm = float(max_norm)
    norm_type = float(norm_type)
    if norm_type == inf:
        total_norm = max(p.grad.data.abs().max() for p in parameters)
    else:
        total_norm = 0
        for p in parameters:
        	#第二步
            param_norm = p.grad.data.norm(norm_type)
            #第三步
            total_norm += param_norm.item() ** norm_type
        total_norm = total_norm ** (1. / norm_type)
    clip_coef = max_norm / (total_norm + 1e-6)
    if clip_coef < 1:
        for p in parameters:
            p.grad.data.mul_(clip_coef)
    return total_norm