deep-person-reid/torchreid/optim/optimizer.py

from __future__ import absolute_import
from __future__ import print_function

import warnings

import torch
import torch.nn as nn


AVAI_OPTIMS = ['adam', 'amsgrad', 'sgd', 'rmsprop']


def build_optimizer(
        model,
        optim='adam',
        lr=0.0003,
        weight_decay=5e-04,
        momentum=0.9,
        sgd_dampening=0,
        sgd_nesterov=False,
        rmsprop_alpha=0.99,
        adam_beta1=0.9,
        adam_beta2=0.99,
        staged_lr=False,
        new_layers='',
        base_lr_mult=0.1
    ):
    """A function wrapper for building an optimizer.

    Args:
        model (nn.Module): model.
        optim (str, optional): optimizer. Default is "adam".
        lr (float, optional): learning rate. Default is 0.0003.
        weight_decay (float, optional): weight decay (L2 penalty). Default is 5e-04.
        momentum (float, optional): momentum factor in sgd. Default is 0.9,
        sgd_dampening (float, optional): dampening for momentum. Default is 0.
        sgd_nesterov (bool, optional): enables Nesterov momentum. Default is False.
        rmsprop_alpha (float, optional): smoothing constant for rmsprop. Default is 0.99.
        adam_beta1 (float, optional): beta-1 value in adam. Default is 0.9.
        adam_beta2 (float, optional): beta-2 value in adam. Default is 0.99,
        staged_lr (bool, optional): uses different learning rates for base and new layers. Base
            layers are pretrained layers while new layers are randomly initialized, e.g. the
            identity classification layer. Enabling ``staged_lr`` can allow the base layers to
            be trained with a smaller learning rate determined by ``base_lr_mult``, while the new
            layers will take the ``lr``. Default is False.
        new_layers (str or list): attribute names in ``model``. Default is empty.
        base_lr_mult (float, optional): learning rate multiplier for base layers. Default is 0.1.

    Examples::
        >>> # A normal optimizer can be built by
        >>> optimizer = torchreid.optim.build_optimizer(model, optim='sgd', lr=0.01)
        >>> # If you want to use a smaller learning rate for pretrained layers
        >>> # and the attribute name for the randomly initialized layer is 'classifier',
        >>> # you can do
        >>> optimizer = torchreid.optim.build_optimizer(
        >>>     model, optim='sgd', lr=0.01, staged_lr=True,
        >>>     new_layers='classifier', base_lr_mult=0.1
        >>> )
        >>> # Now the `classifier` has learning rate 0.01 but the base layers
        >>> # have learning rate 0.01 * 0.1.
        >>> # new_layers can also take multiple attribute names. Say the new layers
        >>> # are 'fc' and 'classifier', you can do
        >>> optimizer = torchreid.optim.build_optimizer(
        >>>     model, optim='sgd', lr=0.01, staged_lr=True,
        >>>     new_layers=['fc', 'classifier'], base_lr_mult=0.1
        >>> )
    """
    if optim not in AVAI_OPTIMS:
        raise ValueError('Unsupported optim: {}. Must be one of {}'.format(optim, AVAI_OPTIMS))
    
    if not isinstance(model, nn.Module):
        raise TypeError('model given to build_optimizer must be an instance of nn.Module')

    if staged_lr:
        if isinstance(new_layers, str):
            if new_layers is None:
                warnings.warn('new_layers is empty, therefore, staged_lr is useless')
            new_layers = [new_layers]
        
        if isinstance(model, nn.DataParallel):
            model = model.module

        base_params = []
        base_layers = []
        new_params = []
        
        for name, module in model.named_children():
            if name in new_layers:
                new_params += [p for p in module.parameters()]
            else:
                base_params += [p for p in module.parameters()]
                base_layers.append(name)
        
        param_groups = [
            {'params': base_params, 'lr': lr * base_lr_mult},
            {'params': new_params},
        ]

    else:
        param_groups = model.parameters()

    if optim == 'adam':
        optimizer = torch.optim.Adam(
            param_groups,
            lr=lr,
            weight_decay=weight_decay,
            betas=(adam_beta1, adam_beta2),
        )

    elif optim == 'amsgrad':
        optimizer = torch.optim.Adam(
            param_groups,
            lr=lr,
            weight_decay=weight_decay,
            betas=(adam_beta1, adam_beta2),
            amsgrad=True,
        )

    elif optim == 'sgd':
        optimizer = torch.optim.SGD(
            param_groups,
            lr=lr,
            momentum=momentum,
            weight_decay=weight_decay,
            dampening=sgd_dampening,
            nesterov=sgd_nesterov,
        )

    elif optim == 'rmsprop':
        optimizer = torch.optim.RMSprop(
            param_groups,
            lr=lr,
            momentum=momentum,
            weight_decay=weight_decay,
            alpha=rmsprop_alpha,
        )

    return optimizer