415be78cc0d1275a29d0ceda550d0d7a7a5224ea,python/ray/util/sgd/torch/torch_trainer.py,TorchTrainer,__init__,#,171

Before Change

                 "For more information, see "
                 "https://github.com/pytorch/examples/issues/467."))

        if not (callable(model_creator) and callable(optimizer_creator)):
            raise ValueError(
                "Must provide a callable model_creator and optimizer_creator.")

        if num_replicas is not None:
            raise DeprecationWarning(
                "num_replicas is deprecated. Use num_workers instead.")

        if batch_size is not None:
            raise DeprecationWarning(
                "batch_size is deprecated. Use config={"batch_size": N} "
                "specify a batch size for each worker or "
                "config={ray.util.sgd.utils.BATCH_SIZE: N} to specify a "
                "batch size to be used across all workers.")

        if data_loader_args:
            raise ValueError(
                "data_loader_args is deprecated. You can return a "
                "torch.utils.data.DataLoader in data_creator. Ray will "
                "automatically set a DistributedSampler if a DataLoader is "
                "returned and num_workers > 1.")

        self.model_creator = model_creator
        self.optimizer_creator = optimizer_creator
        self.loss_creator = loss_creator
        self.data_creator = data_creator
        self.scheduler_creator = scheduler_creator
        self.training_operator_cls = training_operator_cls

        if not training_operator_cls and not loss_creator:
            raise ValueError("If a loss_creator is not provided, you must "
                             "provide a custom training operator.")

        self.initialization_hook = initialization_hook
        self.config = {} if config is None else config
        if use_gpu == "auto":
            use_gpu = torch.cuda.is_available()

After Change

            serialize_data_creation=None,
            data_loader_args=None,
    ):
        if (model_creator or data_creator or optimizer_creator
                or scheduler_creator or loss_creator):
            raise DeprecationWarning(
                "Creator functions are deprecated. You should create a "
                "custom TrainingOperator, override setup, and register all "
                "training state there. See TrainingOperator for more info. "
                "If you would still like to use creator functions, you can "
                "do CustomOperator = TrainingOperator.from_creators("
                "model_creator, ...) and pass in CustomOperator into "
                "TorchTrainer.")

        if num_workers > 1 and not dist.is_available():
            raise ValueError(
                ("Distributed PyTorch is not supported on macOS. "
                 "To run without distributed PyTorch, set "num_workers=1". "
                 "For more information, see "
                 "https://github.com/pytorch/examples/issues/467."))

        if num_replicas is not None:
            raise DeprecationWarning(
                "num_replicas is deprecated. Use num_workers instead.")

        if batch_size is not None:
            raise DeprecationWarning(
                "batch_size is deprecated. Use config={"batch_size": N} "
                "specify a batch size for each worker or "
                "config={ray.util.sgd.utils.BATCH_SIZE: N} to specify a "
                "batch size to be used across all workers.")

        if serialize_data_creation is True:
            if log_once("serialize_data_creation"):
                logging.warning(
                    "serialize_data_creation is deprecated and will be "
                    "ignored. If you require serialized data loading you "
                    "should implement this in TrainingOperator.setup. "
                    "You may find FileLock useful here.")

        if data_loader_args:
            raise DeprecationWarning(
                "data_loader_args is deprecated. You can return a "

Instances