分布式训练¶

from snowflake.ml.modeling.distributors.xgboost import XGBEstimator, XGBScalingConfig

# Training parameters
params = {
    'objective': 'reg:squarederror',
    'max_depth': 6,
    'learning_rate': 0.1
}

# Automatic scaling (recommended)
estimator = XGBEstimator(
    params=params
)

# Call with custom GPU scaling
gpu_estimator = XGBEstimator(
    params=params,
    scaling_config=XGBScalingConfig(use_gpu=True) # optional - available resources will be used automatically
)

# Train and evaluate
booster = estimator.fit(
    dataset=train_connector,
    input_cols=['age', 'income', 'credit_score'],
    label_col='default_risk',
    eval_set=eval_connector,
    verbose_eval=10
)

# Access results
booster = estimator.get_booster() # If you forgot to save the output of fit, get the booster from the estimator
feature_importance = booster.get_score(importance_type='gain')

# Create ShardedDataConnector for data ingestion
from snowflake.ml.data.sharded_data_connector import ShardedDataConnector

example_snowpark_dataframe = session.table("EXAMPLE_TRAINING_DATA")

data_connector = ShardedDataConnector.from_dataframe(example_snowpark_dataframe)

# Import necessary PyTorch libraries
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader

# Define a simple neural network
class SimpleNet(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(SimpleNet, self).__init__()
        self.fc1 = nn.Linear(input_size, hidden_size)
        self.relu = nn.ReLU()
        self.fc2 = nn.Linear(hidden_size, output_size)

    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        return x

# Define the training function
def train_func():
    import torch.distributed as dist
    from torch.nn.parallel import DistributedDataParallel as DDP
    from snowflake.ml.modeling.distributors.pytorch import get_context

    # Use the Snowflake context to get the necessary methods to manage and retrieve information about the distributed training environment
    context = get_context()
    rank = context.get_rank()
    dist.init_process_group(backend='gloo')
    device = torch.device(f"cuda:{context.get_local_rank()}"
                         if torch.cuda.is_available() else "cpu")

    # Initialize model, loss function, and optimizer
    model = SimpleNet(input_size=len(input_cols), hidden_size=32, output_size=1).to(device)
    model = DDP(model)
    criterion = nn.MSELoss()
    optimizer = optim.Adam(model.parameters(), lr=0.001)

    # Retrieve training data
    dataset_map = context.get_dataset_map()
    torch_dataset = dataset_map['train'].get_shard().to_torch_dataset(batch_size=1024)
    dataloader = DataLoader(torch_dataset)

    # Training loop
    for epoch in range(10):
        for batch_dict in dataloader:
            features = torch.cat([batch_dict[col].T for col in input_cols], dim=1).float().to(device)
            labels = batch_dict[label_col].T.squeeze(0).float().to(device)
            output = model(features)
            loss = criterion(output, labels.unsqueeze(1))

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        print(f'Epoch [{epoch+1}/10], Loss: {loss.item():.4f}')

    # Save the model to the model directory provided by the context
    if context.get_rank() == 0:
        torch.save(
            model.module.state_dict(), os.path.join(context.get_model_dir(), "model.pt")
        )

# Set up PyTorchDistributor for distributed training
from snowflake.ml.modeling.distributors.pytorch import PyTorchDistributor, PyTorchScalingConfig, WorkerResourceConfig

pytorch_trainer = PyTorchDistributor(
    train_func=train_func,
    # Optional Scaling Configuration, for single node multi-GPU training.
    scaling_config=PyTorchScalingConfig(
        num_nodes=1,
        num_workers_per_node=1,
        resource_requirements_per_worker=WorkerResourceConfig(num_cpus=0, num_gpus=4)
    )
)

# Run the training process
pytorch_trainer.run(dataset_map={'train': data_connector})

response = pytorch_trainer.run(
        dataset_map={'train': data_connector},
        artifact_stage_location="DB_NAME.SCHEMA_NAME.STAGE_NAME",
    )

stage_location = response.get_model_dir()