当前位置: 首页>>代码示例>>Python>>正文


Python torch.broadcast_optimizer_state方法代码示例

本文整理汇总了Python中horovod.torch.broadcast_optimizer_state方法的典型用法代码示例。如果您正苦于以下问题:Python torch.broadcast_optimizer_state方法的具体用法?Python torch.broadcast_optimizer_state怎么用?Python torch.broadcast_optimizer_state使用的例子?那么恭喜您, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在horovod.torch的用法示例。


在下文中一共展示了torch.broadcast_optimizer_state方法的3个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。

示例1: _start_sync

# 需要导入模块: from horovod import torch [as 别名]
# 或者: from horovod.torch import broadcast_optimizer_state [as 别名]
def _start_sync(self):
        hvd.broadcast_parameters(self.model.state_dict(), root_rank=0)
        hvd.broadcast_optimizer_state(self.optimizer, root_rank=0) 
开发者ID:ddkang,项目名称:advex-uar,代码行数:5,代码来源:trainer.py

示例2: setup_horovod

# 需要导入模块: from horovod import torch [as 别名]
# 或者: from horovod.torch import broadcast_optimizer_state [as 别名]
def setup_horovod(model, learning_rate):
    """ Setup for Horovod usage.

    Args:
        model(MultitaskModel): The MultitaskModel object.
        learning_rate(float): Learning rate for the model.

    Returns: hvd.DistributedOptimizer: Optimizer to use for computing
    gradients and applying updates.

    """
    # Horovod: scale learning rate by the number of GPUs.
    optimizer = optim.Adam(model.parameters(), lr=learning_rate * hvd.size())

    # Horovod: broadcast parameters & optimizer state.
    hvd.broadcast_parameters(model.state_dict(), root_rank=0)
    hvd.broadcast_optimizer_state(optimizer, root_rank=0)

    # Horovod: (optional) compression algorithm.
    compression = hvd.Compression.fp16

    # Horovod: wrap optimizer with DistributedOptimizer.
    optimizer = hvd.DistributedOptimizer(
        optimizer,
        named_parameters=model.named_parameters(),
        compression=compression,
    )

    return optimizer 
开发者ID:microsoft,项目名称:nlp-recipes,代码行数:31,代码来源:gensen_train.py

示例3: horovod_train

# 需要导入模块: from horovod import torch [as 别名]
# 或者: from horovod.torch import broadcast_optimizer_state [as 别名]
def horovod_train(self, model):
        # call setup after the ddp process has connected
        self.setup('fit')
        if self.is_function_implemented('setup', model):
            model.setup('fit')

        if torch.cuda.is_available() and self.on_gpu:
            # Horovod: pin GPU to local rank
            assert self.root_gpu == hvd.local_rank()
            torch.cuda.set_device(self.root_gpu)
            model.cuda(self.root_gpu)

        # avoid duplicating progress bar
        if hvd.rank() != 0 and self.progress_bar_callback is not None:
            self.progress_bar_callback.disable()

        # CHOOSE OPTIMIZER
        # allow for lr schedulers as well
        self.optimizers, self.lr_schedulers, self.optimizer_frequencies = self.init_optimizers(model)

        # Horovod: scale the learning rate by the number of workers to account for
        # increased total batch size
        for optimizer in self.optimizers:
            for param_group in optimizer.param_groups:
                param_group['lr'] *= hvd.size()

        if self.use_amp:
            # An example
            model, optimizers = model.configure_apex(amp, model, self.optimizers, self.amp_level)
            self.optimizers = optimizers
            self.reinit_scheduler_properties(self.optimizers, self.lr_schedulers)

        # Horovod: broadcast parameters & optimizer state to ensure consistent initialization
        hvd.broadcast_parameters(model.state_dict(), root_rank=0)
        for optimizer in self.optimizers:
            hvd.broadcast_optimizer_state(optimizer, root_rank=0)

        def filter_named_parameters(model, optimizer):
            opt_params = set([p for group in optimizer.param_groups for p in group.get('params', [])])
            return [(name, p) for name, p in model.named_parameters() if p in opt_params]

        # Horovod: wrap optimizers to perform gradient aggregation via allreduce
        self.optimizers = [
            hvd.DistributedOptimizer(optimizer, named_parameters=filter_named_parameters(model, optimizer))
            for optimizer in self.optimizers
        ]

        # Update logger rank info from Horovod to avoid race conditions from  different ranks
        # creating directories / writing files in the same locations.
        self.global_rank = hvd.rank()
        rank_zero_only.rank = self.global_rank

        with ExitStack() as stack:
            for optimizer in self.optimizers:
                # Synchronization will be performed explicitly following backward()
                stack.enter_context(optimizer.skip_synchronize())

            self.run_pretrain_routine(model)

        # Make sure all workers have finished training before returning to the user
        hvd.join() 
开发者ID:PyTorchLightning,项目名称:pytorch-lightning,代码行数:63,代码来源:distrib_parts.py


注:本文中的horovod.torch.broadcast_optimizer_state方法示例由纯净天空整理自Github/MSDocs等开源代码及文档管理平台,相关代码片段筛选自各路编程大神贡献的开源项目,源码版权归原作者所有,传播和使用请参考对应项目的License;未经允许,请勿转载。