mmengine.optim.optimizer.builder 源代码

# Copyright (c) OpenMMLab. All rights reserved.
import copy
import inspect
from typing import List, Union

import torch
import torch.nn as nn

from mmengine.config import Config, ConfigDict
from mmengine.device import is_npu_available, is_npu_support_full_precision
from mmengine.registry import OPTIM_WRAPPER_CONSTRUCTORS, OPTIMIZERS
from .optimizer_wrapper import OptimWrapper


def register_torch_optimizers() -> List[str]:
    """Register optimizers in ``torch.optim`` to the ``OPTIMIZERS`` registry.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    torch_optimizers = []
    for module_name in dir(torch.optim):
        if module_name.startswith('__'):
            continue
        _optim = getattr(torch.optim, module_name)
        if inspect.isclass(_optim) and issubclass(_optim,
                                                  torch.optim.Optimizer):
            if module_name == 'Adafactor':
                OPTIMIZERS.register_module(
                    name='TorchAdafactor', module=_optim)
            else:
                OPTIMIZERS.register_module(module=_optim)
            torch_optimizers.append(module_name)
    return torch_optimizers


TORCH_OPTIMIZERS = register_torch_optimizers()


def register_torch_npu_optimizers() -> List[str]:
    """Register optimizers in ``torch npu`` to the ``OPTIMIZERS`` registry.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    if not is_npu_available():
        return []

    import torch_npu
    if not hasattr(torch_npu, 'optim'):
        return []

    torch_npu_optimizers = []
    for module_name in dir(torch_npu.optim):
        if module_name.startswith('__') or module_name in OPTIMIZERS:
            continue
        _optim = getattr(torch_npu.optim, module_name)
        if inspect.isclass(_optim) and issubclass(_optim,
                                                  torch.optim.Optimizer):
            OPTIMIZERS.register_module(module=_optim)
            torch_npu_optimizers.append(module_name)
    return torch_npu_optimizers


NPU_OPTIMIZERS = register_torch_npu_optimizers()


def register_dadaptation_optimizers() -> List[str]:
    """Register optimizers in ``dadaptation`` to the ``OPTIMIZERS`` registry.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    dadaptation_optimizers = []
    try:
        import dadaptation
    except ImportError:
        pass
    else:
        for module_name in ['DAdaptAdaGrad', 'DAdaptAdam', 'DAdaptSGD']:
            _optim = getattr(dadaptation, module_name)
            if inspect.isclass(_optim) and issubclass(_optim,
                                                      torch.optim.Optimizer):
                OPTIMIZERS.register_module(module=_optim)
                dadaptation_optimizers.append(module_name)
    return dadaptation_optimizers


DADAPTATION_OPTIMIZERS = register_dadaptation_optimizers()


def register_lion_optimizers() -> List[str]:
    """Register Lion optimizer to the ``OPTIMIZERS`` registry.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    optimizers = []
    try:
        from lion_pytorch import Lion
    except ImportError:
        pass
    else:
        OPTIMIZERS.register_module(module=Lion)
        optimizers.append('Lion')
    return optimizers


LION_OPTIMIZERS = register_lion_optimizers()


def register_sophia_optimizers() -> List[str]:
    """Register Sophia optimizer to the ``OPTIMIZERS`` registry.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    optimizers = []
    try:
        import Sophia
    except ImportError:
        pass
    else:
        for module_name in dir(Sophia):
            _optim = getattr(Sophia, module_name)
            if inspect.isclass(_optim) and issubclass(_optim,
                                                      torch.optim.Optimizer):
                OPTIMIZERS.register_module(module=_optim)
                optimizers.append(module_name)
    return optimizers


SOPHIA_OPTIMIZERS = register_sophia_optimizers()


def register_bitsandbytes_optimizers() -> List[str]:
    """Register optimizers in ``bitsandbytes`` to the ``OPTIMIZERS`` registry.

    In the `bitsandbytes` library, optimizers that have the same name as the
    default optimizers in PyTorch are prefixed with ``bnb_``. For example,
    ``bnb_Adagrad``.

    Returns:
        List[str]: A list of registered optimizers' name.
    """
    dadaptation_optimizers = []
    try:
        import bitsandbytes as bnb
    except ImportError:
        pass
    else:
        optim_classes = inspect.getmembers(
            bnb.optim, lambda _optim: (inspect.isclass(_optim) and issubclass(
                _optim, torch.optim.Optimizer)))
        for name, optim_cls in optim_classes:
            if name in OPTIMIZERS:
                name = f'bnb_{name}'
            OPTIMIZERS.register_module(module=optim_cls, name=name)
            dadaptation_optimizers.append(name)
    return dadaptation_optimizers


BITSANDBYTES_OPTIMIZERS = register_bitsandbytes_optimizers()


def register_transformers_optimizers():
    transformer_optimizers = []
    try:
        from transformers import Adafactor
    except ImportError:
        pass
    else:
        OPTIMIZERS.register_module(name='Adafactor', module=Adafactor)
        transformer_optimizers.append('Adafactor')
    return transformer_optimizers


TRANSFORMERS_OPTIMIZERS = register_transformers_optimizers()


def build_optim_wrapper(model: nn.Module,
                        cfg: Union[dict, Config, ConfigDict]) -> OptimWrapper:
    """Build function of OptimWrapper.

    If ``constructor`` is set in the ``cfg``, this method will build an
    optimizer wrapper constructor, and use optimizer wrapper constructor to
    build the optimizer wrapper. If ``constructor`` is not set, the
    ``DefaultOptimWrapperConstructor`` will be used by default.

    Args:
        model (nn.Module): Model to be optimized.
        cfg (dict): Config of optimizer wrapper, optimizer constructor and
            optimizer.

    Returns:
        OptimWrapper: The built optimizer wrapper.
    """
    optim_wrapper_cfg = copy.deepcopy(cfg)
    constructor_type = optim_wrapper_cfg.pop('constructor',
                                             'DefaultOptimWrapperConstructor')
    paramwise_cfg = optim_wrapper_cfg.pop('paramwise_cfg', None)

    # Since the current generation of NPU(Ascend 910) only supports
    # mixed precision training, here we turn on mixed precision
    # to make the training normal
    if is_npu_available() and not is_npu_support_full_precision():
        optim_wrapper_cfg['type'] = 'AmpOptimWrapper'

    optim_wrapper_constructor = OPTIM_WRAPPER_CONSTRUCTORS.build(
        dict(
            type=constructor_type,
            optim_wrapper_cfg=optim_wrapper_cfg,
            paramwise_cfg=paramwise_cfg))
    optim_wrapper = optim_wrapper_constructor(model)
    return optim_wrapper