hivemind/examples/albert/optim.py

import math

import torch
from torch.optim.optimizer import Optimizer


__all__ = ('Lamb',)


class Lamb(Optimizer):

    def __init__(
        self,
        params,
        lr: float = 1e-3,
        betas = (0.9, 0.999),
        eps: float = 1e-6,
        weight_decay: float = 0,
        clamp_value: float = 10,
        adam: bool = False,
        debias: bool = False,
    ) -> None:
        if lr <= 0.0:
            raise ValueError('Invalid learning rate: {}'.format(lr))
        if eps < 0.0:
            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])
            )
        if weight_decay < 0:
            raise ValueError(
                'Invalid weight_decay value: {}'.format(weight_decay)
            )
        if clamp_value < 0.0:
            raise ValueError('Invalid clamp value: {}'.format(clamp_value))

        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
        self.clamp_value = clamp_value
        self.adam = adam
        self.debias = debias

        super(Lamb, self).__init__(params, defaults)

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

        Arguments:
            closure: 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:
                    msg = (
                        'Lamb does not support sparse gradients, '
                        'please consider SparseAdam instead'
                    )
                    raise RuntimeError(msg)

                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)

                # Paper v3 does not use debiasing.
                if self.debias:
                    bias_correction = math.sqrt(1 - beta2 ** state['step'])
                    bias_correction /= 1 - beta1 ** state['step']
                else:
                    bias_correction = 1

                # Apply bias to lr to avoid broadcast.
                step_size = group['lr'] * bias_correction

                weight_norm = torch.norm(p.data).clamp(0, self.clamp_value)

                adam_step = exp_avg / exp_avg_sq.sqrt().add(group['eps'])
                adam_step.add_(p.data, alpha=group['weight_decay'])

                adam_norm = torch.norm(adam_step).clamp_min(0.001)
                trust_ratio = weight_norm / adam_norm
                state['weight_norm'] = weight_norm
                state['adam_norm'] = adam_norm
                state['trust_ratio'] = trust_ratio
                if self.adam:
                    trust_ratio = 1

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

        return loss