ColossalAI/colossalai/nn/optimizer/lamb.py

"""
Adapted from the pytorch-lamb library at https://github.com/cybertronai/pytorch-lamb
"""

import torch
from torch.optim import Optimizer


class Lamb(Optimizer):
    r"""Implements Lamb algorithm.
    It has been proposed in `Large Batch Optimization for Deep Learning: Training BERT in 76 minutes`_.

    Arguments:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float, optional): learning rate (default: 1e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its square (default: (0.9, 0.999))
        eps (float, optional): term added to the denominator to improve
            numerical stability (default: 1e-6)
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        adam (bool, optional): always use trust ratio = 1, which turns this into
            Adam. Useful for comparison purposes.

    .. _Large Batch Optimization for Deep Learning\: Training BERT in 76 minutes:
        https://arxiv.org/abs/1904.00962
    """

    def __init__(
        self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6, weight_decay=0, adam=False, bias_correction=False
    ):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, bias_correction=bias_correction)
        self.adam = adam
        super(Lamb, self).__init__(params, defaults)

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

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

        for group in self.param_groups:
            for p in group["params"]:
                if p.grad is None:
                    continue
                grad = p.grad.data
                if grad.is_sparse:
                    raise RuntimeError("Lamb does not support sparse gradients, consider SparseAdam instead.")

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state["step"] = 0
                    # Exponential moving average of gradient values
                    state["exp_avg"] = torch.zeros_like(p)
                    # Exponential moving average of squared gradient values
                    state["exp_avg_sq"] = torch.zeros_like(p)

                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
                beta1, beta2 = group["betas"]

                state["step"] += 1

                # Decay the first and second moment running average coefficient
                # m_t
                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
                # v_t
                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)

                # NOTE: Paper v3 does not use debiasing.
                scaled_lr = group["lr"]
                if group["bias_correction"]:
                    bias_correction1 = 1 - beta1 ** state["step"]
                    bias_correction2 = 1 - beta2 ** state["step"]
                    # Apply debiasing to lr to avoid broadcast
                    scaled_lr *= (bias_correction2**0.5) / bias_correction1
                    # exp_avg.div_(bias_correction1)
                    # exp_avg_sq.div_(bias_correction2)

                weight_norm = p.data.pow(2).sum().sqrt()

                adam_step = exp_avg / exp_avg_sq.sqrt().add(group["eps"])
                if group["weight_decay"] != 0:
                    adam_step.add_(p.data, alpha=group["weight_decay"])

                adam_norm = adam_step.pow(2).sum().sqrt()
                if weight_norm == 0 or adam_norm == 0:
                    trust_ratio = 1
                else:
                    trust_ratio = weight_norm / adam_norm

                if self.adam:
                    trust_ratio = 1

                p.data.add_(adam_step, alpha=-scaled_lr * trust_ratio)

        return loss