optimizer.py

# --------------------------------------------------------
# Swin Transformer
# Copyright (c) 2021 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ze Liu
# --------------------------------------------------------
import torch
from torch import optim as optim
from torch.optim import Optimizer

class Optimizers(object):
    """Handles a list of optimizers."""

    def __init__(self, *ops):
        self.optimizers = ops

    def zero_grad(self):
        for op in self.optimizers:
            op.zero_grad()

    def step(self):
        """Makes all optimizers update their params."""
        for optimizer in self.optimizers:
            optimizer.step()

def build_optimizer(config, model):
    """
    Build optimizer, set weight decay of normalization to 0 by default.
    """
    skip = {}
    skip_keywords = {}
    if hasattr(model, 'no_weight_decay'):
        skip = model.no_weight_decay()
    if hasattr(model, 'no_weight_decay_keywords'):
        skip_keywords = model.no_weight_decay_keywords()
    parameters = set_weight_decay(model, skip, skip_keywords)

    opt_lower = config.TRAIN.OPTIMIZER.NAME.lower()
    optimizer = None
    if opt_lower == 'sgd':
        optimizer = optim.SGD(parameters, momentum=config.TRAIN.OPTIMIZER.MOMENTUM, nesterov=True,
                              lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
    elif opt_lower == 'adamw':
        optimizer = optim.AdamW(parameters, eps=config.TRAIN.OPTIMIZER.EPS, betas=config.TRAIN.OPTIMIZER.BETAS,
                                lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
    elif opt_lower == "adamax":
        optimizer = AdamaxW(parameters, lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)

    return optimizer


def set_weight_decay(model, skip_list=(), skip_keywords=()):
    has_decay = []
    no_decay = []

    for name, param in model.named_parameters():
        if not param.requires_grad:
            continue  # frozen weights
        if len(param.shape) == 1 or name.endswith(".bias") or (name in skip_list) or \
                check_keywords_in_name(name, skip_keywords):
            no_decay.append(param)
            # print(f"{name} has no weight decay")
        else:
            has_decay.append(param)
    return [{'params': has_decay},
            {'params': no_decay, 'weight_decay': 0.}]


def check_keywords_in_name(name, keywords=()):
    isin = False
    for keyword in keywords:
        if keyword in name:
            isin = True
    return isin

class AdamaxW(Optimizer):
    r"""Implements AdamaxW algorithm.
    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
    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-8)
        weight_decay (float, optional): weight decay coefficient (default: 1e-2)
        amsgrad (boolean, optional): whether to use the AMSGrad variant of this
            algorithm from the paper `On the Convergence of Adam and Beyond`_
            (default: False)
    .. _Adamax\: A Method for Stochastic Optimization:
        https://arxiv.org/abs/1412.6980
    .. _Decoupled Weight Decay Regularization:
        https://arxiv.org/abs/1711.05101
    .. _On the Convergence of Adam and Beyond:
        https://openreview.net/forum?id=ryQu7f-RZ
    """

    # Learning rate/Weight decay are different for Swin
    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
                 weight_decay=1e-2, amsgrad=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, amsgrad=amsgrad)
        super(AdamaxW, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(AdamaxW, self).__setstate__(state)
        for group in self.param_groups:
            group.setdefault('amsgrad', False)

    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

                # Perform step-weight decay
                p.data.mul_(1 - group['lr'] * group['weight_decay'])

                # Perform optimization step
                grad = p.grad.data
                if grad.is_sparse:
                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
                amsgrad = group['amsgrad']

                state = self.state[p]
                eps = group['eps']

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    # Exponential moving average of gradient values
                    state['exp_avg'] = torch.zeros_like(p.data)
                    # Exponential moving average of inf gradient values
                    state['exp_inf'] = torch.zeros_like(p.data)

                exp_avg, exp_inf = state['exp_avg'], state['exp_inf']
                beta1, beta2 = group['betas']
                state['step'] += 1
                # Decay the first and inf moment running average coefficient
                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)

                norm_buf = torch.cat([
                    exp_inf.mul_(beta2).unsqueeze(0),
                    grad.abs().add_(eps).unsqueeze_(0)
                ], 0)
                torch.max(norm_buf, 0, keepdim=False, out=(exp_inf, exp_inf.new().long()))

                bias_correction = 1 - beta1 ** state['step']
                clr = group['lr'] / bias_correction
                p.data.addcdiv_(exp_avg, exp_inf, value=-clr)

        return loss