多节点集群 ML 上的容器运行时¶

from snowflake.ml.modeling.distributors.xgboost import XGBEstimator, XGBScalingConfig
from snowflake.ml.data.data_connector import DataConnector
from implementations.ray_data_ingester import RayDataIngester
table_name = "MULTINODE_SAMPLE_TRAIN_DS"

# Use code like the following to generate example data
"""
# Create a table in current database/schema and store data there
def generate_dataset_sql(db, schema, table_name) -> str:
    sql_script = f"CREATE TABLE IF NOT EXISTS {db}.{schema}.{table_name} AS \n"
    sql_script += f"select \n"
    for i in range(1, 10):
        sql_script += f"uniform(0::float, 10::float, random()) AS FT_{i}, \n"
    sql_script += f"FT_1 + FT_2 AS TARGET, \n"
    sql_script += f"from TABLE(generator(rowcount=>({10000})));"
    return sql_script
session.sql(generate_dataset_sql(session.get_current_database(), session.get_current_schema(), table_name)).collect()
"""

sample_train_df = session.table(table_name)
INPUT_COLS = list(sample_train_df.columns)
LABEL_COL = "TARGET"
INPUT_COLS.remove(LABEL_COL)

params = {
    "eta": 0.1,
    "max_depth": 8,
    "min_child_weight": 100,
    "tree_method": "hist",
}

scaling_config = XGBScalingConfig(
    use_gpu=False
)

estimator = XGBEstimator(
    n_estimators=50,
    objective="reg:squarederror",
    params=params,
    scaling_config=scaling_config,
)
data_connector = DataConnector.from_dataframe(
    sample_train_df, ingestor_class=RayDataIngester
)

xgb_model = estimator.fit(
    data_connector, input_cols=INPUT_COLS, label_col=LABEL_COL
)

dataset_map = context.get_dataset_map()
train_dataset = DecodedDataset(dataset_map["train"].get_shard().to_torch_dataset())
test_dataset = DecodedDataset(dataset_map["test"].to_torch_dataset())

hyper_parms = context.get_hyper_params()
num_epochs = int(hyper_parms['num_epochs'])

def train_func():
    import io
    import base64
    import time
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    import torch.optim as optim
    import torch.distributed as dist
    from torch.nn.parallel import DistributedDataParallel as DDP
    from torchvision import transforms
    from torch.utils.data import IterableDataset
    from torch.optim.lr_scheduler import StepLR
    from PIL import Image
    from snowflake.ml.modeling.distributors.pytorch import get_context

    class Net(nn.Module):
        def __init__(self):
            super(Net, self).__init__()
            self.conv1 = nn.Conv2d(1, 32, 3, 1)
            self.conv2 = nn.Conv2d(32, 64, 3, 1)
            self.dropout1 = nn.Dropout(0.25)
            self.dropout2 = nn.Dropout(0.5)
            self.fc1 = nn.Linear(9216, 128)
            self.fc2 = nn.Linear(128, 10)

        def forward(self, x):
            x = self.conv1(x)
            x = F.relu(x)
            x = self.conv2(x)
            x = F.relu(x)
            x = F.max_pool2d(x, 2)
            x = self.dropout1(x)
            x = torch.flatten(x, 1)
            x = self.fc1(x)
            x = F.relu(x)
            x = self.dropout2(x)
            x = self.fc2(x)
            output = F.log_softmax(x, dim=1)
            return output

    class DecodedDataset(IterableDataset):
        def __init__(self, source_dataset):
            self.source_dataset = source_dataset
            self.transforms = transforms.ToTensor()  # Ensure we apply ToTensor transform

        def __iter__(self):
            for row in self.source_dataset:
                base64_image = row['IMAGE']
                image = Image.open(io.BytesIO(base64.b64decode(base64_image)))
                # Convert the image to a tensor
                image = self.transforms(image)  # Converts PIL image to tensor

                labels = row['LABEL']
                yield image, int(labels)

    def train(model, device, train_loader, optimizer, epoch):
        model.train()
        batch_idx = 1
        for data, target in train_loader:
            # print(f"data : {data} \n target: {target}")
            # raise RuntimeError("test")
            data, target = data.to(device), target.to(device)
            optimizer.zero_grad()
            output = model(data)
            loss = F.nll_loss(output, target)
            loss.backward()
            optimizer.step()
            if batch_idx % 100 == 0:
                print('Train Epoch: {} [Processed {} images]\tLoss: {:.6f}'.format(epoch, batch_idx * len(data), loss.item()))
            batch_idx += 1

    context = get_context()
    rank = context.get_local_rank()
    device = f"cuda:{rank}"
    is_distributed = context.get_world_size() > 1
    if is_distributed:
        dist.init_process_group(backend="nccl")
    print(f"Worker Rank : {context.get_rank()}, world_size: {context.get_world_size()}")

    dataset_map = context.get_dataset_map()
    train_dataset = DecodedDataset(dataset_map["train"].get_shard().to_torch_dataset())
    test_dataset = DecodedDataset(dataset_map["test"].to_torch_dataset())

    batch_size = 64
    train_loader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=batch_size,
        pin_memory=True,
        pin_memory_device=f"cuda:{rank}"
    )
    test_loader = torch.utils.data.DataLoader(
        test_dataset,
        batch_size=batch_size,
        pin_memory=True,
        pin_memory_device=f"cuda:{rank}"
    )

    model = Net().to(device)
    if is_distributed:
        model = DDP(model)
    optimizer = optim.Adadelta(model.parameters())
    scheduler = StepLR(optimizer, step_size=1)

    hyper_parms = context.get_hyper_params()
    num_epochs = int(hyper_parms['num_epochs'])
    start_time = time.time()
    for epoch in range(num_epochs):
        train(model, device, train_loader, optimizer, epoch+1)
        scheduler.step()
    now = time.time()
    context.get_metrics_reporter().log_metrics({"train_func_train_time": int(now-start_time)})
    test(model, device, test_loader, context)

# Set up PyTorchDistributor
from snowflake.ml.modeling.distributors.pytorch import PyTorchDistributor, PyTorchScalingConfig, WorkerResourceConfig
from snowflake.ml.data.sharded_data_connector import ShardedDataConnector
from snowflake.ml.data.data_connector import DataConnector

df = session.table("MNIST_60K")

train_df, test_df = df.random_split([0.99, 0.01], 0)

# Create data connectors for training and test data
train_data = ShardedDataConnector.from_dataframe(train_df)
test_data = DataConnector.from_dataframe(test_df)

pytorch_trainer = PyTorchDistributor(
    train_func=train_func,
    scaling_config=PyTorchScalingConfig(  # scaling configuration
        num_nodes=2,
        num_workers_per_node=1,
        resource_requirements_per_worker=WorkerResourceConfig(num_cpus=0, num_gpus=1),
    )
)

# Run the trainer.
results = pytorch_trainer.run(  # accepts context values as parameters
    dataset_map={"train": train_data, "test": test_data},
    hyper_params={"num_epochs": "1"}
)