Modello di Deep Learning end-to-end con PyTorch

Questa esercitazione illustra il ciclo di vita completo della sperimentazione, del training, dell'ottimizzazione, della registrazione, della valutazione e della distribuzione per un progetto di modellazione di Deep Learning. Illustra come usare MLflow per tenere traccia di ogni aspetto dei processi di sviluppo e distribuzione del modello.

Il notebook usa PyTorch, un pacchetto di Python che fornisce il calcolo accelerato della GPU e funzionalità di alto livello per la creazione di reti di Deep Learning. Quando si è pronti, è possibile distribuire il modello usando Mosaic AI Model Serving.

In questa esercitazione dettagliata si apprenderà come:

  • Generare e visualizzare i dati: Creare dati sintetici per simulare scenari reali e visualizzare le relazioni tra funzionalità.
  • Progettare ed eseguire il training di una rete neurale PyTorch: Creare un modello di Deep Learning flessibile personalizzato per le attività di regressione.
  • Tenere traccia degli esperimenti con MLflow: Registrare metriche, parametri e artefatti per la riproducibilità completa.
  • Automatizzare l'ottimizzazione degli iperparametri: Usare Optuna per cercare in modo efficiente configurazioni di modelli ottimali.
  • Registrare e gestire i modelli: Usare il Registro modelli MLflow integrato con Unity Catalog per la governance sicura e organizzata dei modelli.
  • Distribuire e prevedere: Caricare i modelli registrati per eseguire predizioni localmente o su larga scala usando le Spark UDFs.
%pip install -Uqqq mlflow pytorch-lightning optuna skorch uv optuna-integration[pytorch_lightning]
%restart_python
from typing import Tuple, Optional, Dict, List, Any

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score


import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset

import pytorch_lightning as pl
from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

import mlflow
from mlflow.models import infer_signature
from mlflow.tracking import MlflowClient
from mlflow.entities import Metric, Param

import optuna
from optuna.integration import PyTorchLightningPruningCallback

import time

0. Configurare il Registro modelli con il catalogo Unity

Uno dei vantaggi principali dell'uso di MLflow in Databricks è l'integrazione senza problemi con Unity Catalog. Questa integrazione semplifica la gestione e la governance dei modelli, garantendo che ogni modello sviluppato venga monitorato, sottoposto a controllo delle versioni e sicuro. Per altre informazioni sul catalogo unity, vedere (AWS | Azure | GCP).

Impostare l'URI del Registro di sistema

La cella seguente configura MLflow per l'uso del catalogo Unity per la registrazione del modello.

mlflow.set_registry_uri("databricks-uc")

1. Creare un set di dati di regressione sintetica

La cella successiva definisce la create_regression_data funzione . Questa funzione genera dati sintetici per la regressione. Il set di dati risultante include relazioni lineari e non lineari tra le caratteristiche e la destinazione, il rumore e le funzionalità con importanza variabile. Queste funzionalità sono progettate per simulare scenari di dati reali.

def create_regression_data(
    n_samples: int,
    n_features: int,
    seed: int = 1994,
    noise_level: float = 0.3,
    nonlinear: bool = True
) -> Tuple[pd.DataFrame, pd.Series]:
    """Generates synthetic regression data with interesting correlations for MLflow and PyTorch demonstrations.

    This function creates a DataFrame of continuous features and computes a target variable with nonlinear
    relationships and interactions between features. The data is designed to be complex enough to demonstrate
    the capabilities of deep learning, but not so complex that a reasonable model can't be learned.

    Args:
        n_samples (int): Number of samples (rows) to generate.
        n_features (int): Number of feature columns.
        seed (int, optional): Random seed for reproducibility. Defaults to 1994.
        noise_level (float, optional): Level of Gaussian noise to add to the target. Defaults to 0.3.
        nonlinear (bool, optional): Whether to add nonlinear feature transformations. Defaults to True.

    Returns:
        Tuple[pd.DataFrame, pd.Series]:
            - pd.DataFrame: DataFrame containing the synthetic features.
            - pd.Series: Series containing the target labels.

    Example:
        >>> df, target = create_regression_data(n_samples=1000, n_features=10)
    """
    rng = np.random.RandomState(seed)

    # Generate random continuous features
    X = rng.uniform(-5, 5, size=(n_samples, n_features))

    # Create feature DataFrame with meaningful names
    columns = [f"feature_{i}" for i in range(n_features)]
    df = pd.DataFrame(X, columns=columns)

    # Generate base target variable with linear relationship to a subset of features
    # Use only the first n_features//2 features to create some irrelevant features
    weights = rng.uniform(-2, 2, size=n_features//2)
    target = np.dot(X[:, :n_features//2], weights)

    # Add some nonlinear transformations if requested
    if nonlinear:
        # Add square term for first feature
        target += 0.5 * X[:, 0]**2

        # Add interaction between the second and third features
        if n_features >= 3:
            target += 1.5 * X[:, 1] * X[:, 2]

        # Add sine transformation of fourth feature
        if n_features >= 4:
            target += 2 * np.sin(X[:, 3])

        # Add exponential of fifth feature, scaled down
        if n_features >= 5:
            target += 0.1 * np.exp(X[:, 4] / 2)

        # Add threshold effect for sixth feature
        if n_features >= 6:
            target += 3 * (X[:, 5] > 1.5).astype(float)

    # Add Gaussian noise
    noise = rng.normal(0, noise_level * target.std(), size=n_samples)
    target += noise

    # Add a few more interesting features to the DataFrame

    # Add a correlated feature (but not used in target calculation)
    if n_features >= 7:
        df['feature_correlated'] = df['feature_0'] * 0.8 + rng.normal(0, 0.2, size=n_samples)

    # Add a cyclical feature
    df['feature_cyclical'] = np.sin(np.linspace(0, 4*np.pi, n_samples))

    # Add a feature with outliers
    df['feature_with_outliers'] = rng.normal(0, 1, size=n_samples)
    # Add outliers to ~1% of samples
    outlier_idx = rng.choice(n_samples, size=n_samples//100, replace=False)
    df.loc[outlier_idx, 'feature_with_outliers'] = rng.uniform(10, 15, size=len(outlier_idx))

    return df, pd.Series(target, name='target')

2. Visualizzazioni esplorative dei dati

Le visualizzazioni consentono di comprendere i dati. Il codice nella cella seguente crea 6 funzioni, ognuna delle quali genera un tracciato diverso per consentire di esaminare visivamente il set di dati.

È possibile usare MLflow per registrare le visualizzazioni come artefatti, rendendo la sperimentazione completamente riproducibile.

def plot_feature_distributions(X: pd.DataFrame, y: pd.Series, n_cols: int = 3) -> plt.Figure:
    """
    Creates a grid of histograms for each feature in the dataset.

    Args:
        X (pd.DataFrame): DataFrame containing features.
        y (pd.Series): Series containing the target variable.
        n_cols (int): Number of columns in the grid layout.

    Returns:
        plt.Figure: The matplotlib Figure object containing the distribution plots.
    """
    features = X.columns
    n_features = len(features)
    n_rows = (n_features + n_cols - 1) // n_cols

    fig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 4 * n_rows))
    axes = axes.flatten() if n_rows * n_cols > 1 else [axes]

    for i, feature in enumerate(features):
        if i < len(axes):
            ax = axes[i]
            sns.histplot(X[feature], ax=ax, kde=True, color='skyblue')
            ax.set_title(f'Distribution of {feature}')

    # Hide any unused subplots
    for i in range(n_features, len(axes)):
        axes[i].set_visible(False)

    plt.tight_layout()
    fig.suptitle('Feature Distributions', y=1.02, fontsize=16)
    plt.close(fig)
    return fig

def plot_correlation_heatmap(X: pd.DataFrame, y: pd.Series) -> plt.Figure:
    """
    Creates a correlation heatmap of all features and the target variable.

    Args:
        X (pd.DataFrame): DataFrame containing features.
        y (pd.Series): Series containing the target variable.

    Returns:
        plt.Figure: The matplotlib Figure object containing the heatmap.
    """
    # Combine features and target into one DataFrame
    data = X.copy()
    data['target'] = y

    # Calculate correlation matrix
    corr_matrix = data.corr()

    # Set up the figure
    fig, ax = plt.subplots(figsize=(12, 10))

    # Draw the heatmap with a color bar
    cmap = sns.diverging_palette(220, 10, as_cmap=True)
    sns.heatmap(corr_matrix, annot=True, fmt='.2f', cmap=cmap,
                center=0, square=True, linewidths=0.5, ax=ax)

    ax.set_title('Feature Correlation Heatmap', fontsize=16)
    plt.close(fig)
    return fig

def plot_feature_target_relationships(X: pd.DataFrame, y: pd.Series, n_cols: int = 3) -> plt.Figure:
    """
    Creates a grid of scatter plots showing the relationship between each feature and the target.

    Args:
        X (pd.DataFrame): DataFrame containing features.
        y (pd.Series): Series containing the target variable.
        n_cols (int): Number of columns in the grid layout.

    Returns:
        plt.Figure: The matplotlib Figure object containing the relationship plots.
    """
    features = X.columns
    n_features = len(features)
    n_rows = (n_features + n_cols - 1) // n_cols

    fig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 4 * n_rows))
    axes = axes.flatten() if n_rows * n_cols > 1 else [axes]

    for i, feature in enumerate(features):
        if i < len(axes):
            ax = axes[i]
            # Scatter plot with regression line
            sns.regplot(x=X[feature], y=y, ax=ax,
                       scatter_kws={'alpha': 0.5, 'color': 'blue'},
                       line_kws={'color': 'red'})
            ax.set_title(f'{feature} vs Target')

    for i in range(n_features, len(axes)):
        axes[i].set_visible(False)

    plt.tight_layout()
    fig.suptitle('Feature vs Target Relationships', y=1.02, fontsize=16)
    plt.close(fig)
    return fig

def plot_pairwise_relationships(X: pd.DataFrame, y: pd.Series, features: list[str]) -> plt.Figure:
    """
    Creates a pairplot showing relationships between selected features and the target.

    Args:
        X (pd.DataFrame): DataFrame containing features.
        y (pd.Series): Series containing the target variable.
        features (List[str]): List of feature names to include in the plot.

    Returns:
        plt.Figure: The matplotlib Figure object containing the pairplot.
    """
    # Ensure features exist in the DataFrame
    valid_features = [f for f in features if f in X.columns]

    if not valid_features:
        fig, ax = plt.subplots()
        ax.text(0.5, 0.5, "No valid features provided", ha='center', va='center')
        return fig

    # Combine selected features and target
    data = X[valid_features].copy()
    data['target'] = y

    # Create pairplot
    pairgrid = sns.pairplot(data, diag_kind="kde",
                          plot_kws={"alpha": 0.6, "s": 50},
                          corner=True)

    pairgrid.fig.suptitle("Pairwise Feature Relationships", y=1.02, fontsize=16)
    plt.close(pairgrid.fig)
    return pairgrid.fig

def plot_outliers(X: pd.DataFrame, n_cols: int = 3) -> plt.Figure:
    """
    Creates a grid of box plots to detect outliers in each feature.

    Args:
        X (pd.DataFrame): DataFrame containing features.
        n_cols (int): Number of columns in the grid layout.

    Returns:
        plt.Figure: The matplotlib Figure object containing the outlier plots.
    """
    features = X.columns
    n_features = len(features)
    n_rows = (n_features + n_cols - 1) // n_cols

    fig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 4 * n_rows))
    axes = axes.flatten() if n_rows * n_cols > 1 else [axes]

    for i, feature in enumerate(features):
        if i < len(axes):
            ax = axes[i]
            # Box plot to detect outliers
            sns.boxplot(x=X[feature], ax=ax, color='skyblue')
            ax.set_title(f'Outlier Detection for {feature}')
            ax.set_xlabel(feature)

    # Hide any unused subplots
    for i in range(n_features, len(axes)):
        axes[i].set_visible(False)

    plt.tight_layout()
    fig.suptitle('Outlier Detection for Features', y=1.02, fontsize=16)
    plt.close(fig)
    return fig

def plot_residuals(y_true: pd.Series, y_pred: np.ndarray) -> plt.Figure:
    """
    Creates a residual plot to analyze model prediction errors.

    Args:
        y_true (pd.Series): True target values.
        y_pred (np.ndarray): Predicted target values.

    Returns:
        plt.Figure: The matplotlib Figure object containing the residual plot.
    """
    residuals = y_true - y_pred

    fig, ax = plt.subplots(figsize=(10, 6))

    # Scatter plot of predicted values vs residuals
    ax.scatter(y_pred, residuals, alpha=0.5)
    ax.axhline(y=0, color='r', linestyle='-')

    ax.set_xlabel('Predicted Values')
    ax.set_ylabel('Residuals')
    ax.set_title('Residual Plot')

    plt.tight_layout()
    plt.close(fig)
    return fig

3. Progettare una rete neurale PyTorch per la regressione

Il codice nella cella seguente definisce l'architettura del modello PyTorch. Crea una rete neurale flessibile con le caratteristiche seguenti:

  • Architettura configurabile: Numero regolabile e dimensioni dei livelli nascosti.
  • Funzioni di attivazione: RiLU per i livelli nascosti, lineare per l'output.
  • Regolarizzazione: Eliminazione facoltativa per impedire l'overfitting.
  • Normalizzazione dei livelli: Per stabilizzare la formazione e accelerare la convergenza.

Per illustrare diversi approcci, le celle seguenti illustrano come creare prima la rete neurale usando un modulo PyTorch standard e quindi un modulo PyTorch Lightning.

class RegressionNN(nn.Module):
    """
    A flexible feedforward neural network for regression tasks.

    Attributes:
        input_dim (int): Number of input features.
        hidden_dims (List[int]): List of hidden layer dimensions.
        dropout_rate (float): Dropout probability for regularization.
        use_layer_norm (bool): Whether to use layer normalization.
    """

    def __init__(
        self,
        input_dim: int,
        hidden_dims: List[int] = [64, 32],
        dropout_rate: float = 0.1,
        use_layer_norm: bool = True
    ):
        """
        Initialize the neural network.

        Args:
            input_dim (int): Number of input features.
            hidden_dims (List[int]): List of hidden layer dimensions.
            dropout_rate (float): Dropout probability for regularization.
            use_layer_norm (bool): Whether to use layer normalization.
        """
        super().__init__()

        self.input_dim = input_dim
        self.hidden_dims = hidden_dims
        self.dropout_rate = dropout_rate
        self.use_layer_norm = use_layer_norm

        # Build layers dynamically based on hidden_dims
        layers = []

        # Input layer
        prev_dim = input_dim

        # Hidden layers
        for dim in hidden_dims:
            layers.append(nn.Linear(prev_dim, dim))

            if use_layer_norm:
                layers.append(nn.LayerNorm(dim))

            layers.append(nn.ReLU())

            if dropout_rate > 0:
                layers.append(nn.Dropout(dropout_rate))

            prev_dim = dim

        # Output layer (single output for regression)
        layers.append(nn.Linear(prev_dim, 1))

        # Combine all layers
        self.model = nn.Sequential(*layers)

    def forward(self, x):
        """Forward pass through the network."""
        return self.model(x).squeeze()

    def get_params(self) -> Dict[str, Any]:
        """Return model parameters as a dictionary for MLflow logging."""
        return {
            "input_dim": self.input_dim,
            "hidden_dims": self.hidden_dims,
            "dropout_rate": self.dropout_rate,
            "use_layer_norm": self.use_layer_norm
        }
class RegressionLightningModule(pl.LightningModule):
    """
    PyTorch Lightning module for regression tasks.

    This class wraps the RegressionNN model and adds training, validation,
    and testing logic using the PyTorch Lightning framework.
    """

    def __init__(
        self,
        input_dim: int,
        hidden_dims: List[int] = [64, 32],
        dropout_rate: float = 0.1,
        use_layer_norm: bool = True,
        learning_rate: float = 1e-3,
        weight_decay: float = 1e-5
    ):
        """
        Initialize the Lightning module.

        Args:
            input_dim (int): Number of input features.
            hidden_dims (List[int]): List of hidden layer dimensions.
            dropout_rate (float): Dropout probability for regularization.
            use_layer_norm (bool): Whether to use layer normalization.
            learning_rate (float): Learning rate for the optimizer.
            weight_decay (float): Weight decay for L2 regularization.
        """
        super().__init__()

        # Save hyperparameters
        self.save_hyperparameters()

        # Create the model
        self.model = RegressionNN(
            input_dim=input_dim,
            hidden_dims=hidden_dims,
            dropout_rate=dropout_rate,
            use_layer_norm=use_layer_norm
        )

        # Loss function
        self.loss_fn = nn.MSELoss()

    def forward(self, x):
        """Forward pass through the network."""
        return self.model(x)

    def configure_optimizers(self):
        """Configure the optimizer for training."""
        optimizer = torch.optim.Adam(
            self.parameters(),
            lr=self.hparams.learning_rate,
            weight_decay=self.hparams.weight_decay
        )
        return optimizer

    def training_step(self, batch, batch_idx):
        """Perform a training step."""
        x, y = batch
        y_pred = self(x)
        loss = self.loss_fn(y_pred, y)
        self.log('train_loss', loss, prog_bar=True)
        return loss

    def validation_step(self, batch, batch_idx):
        """Perform a validation step."""
        x, y = batch
        y_pred = self(x)
        loss = self.loss_fn(y_pred, y)
        self.log('val_loss', loss, prog_bar=True)

        # Calculate additional metrics
        rmse = torch.sqrt(loss)
        mae = torch.mean(torch.abs(y_pred - y))

        self.log('val_rmse', rmse, prog_bar=True)
        self.log('val_mae', mae, prog_bar=True)

        return loss

    def test_step(self, batch, batch_idx):
        """Perform a test step."""
        x, y = batch
        y_pred = self(x)
        loss = self.loss_fn(y_pred, y)

        # Calculate metrics for test set
        rmse = torch.sqrt(loss)
        mae = torch.mean(torch.abs(y_pred - y))

        self.log('test_loss', loss)
        self.log('test_rmse', rmse)
        self.log('test_mae', mae)

        return loss

    def get_params(self) -> Dict[str, Any]:
        """Return model parameters as a dictionary for MLflow logging."""
        return {
            "input_dim": self.hparams.input_dim,
            "hidden_dims": self.hparams.hidden_dims,
            "dropout_rate": self.hparams.dropout_rate,
            "use_layer_norm": self.hparams.use_layer_norm,
            "learning_rate": self.hparams.learning_rate,
            "weight_decay": self.hparams.weight_decay
        }
def prepare_dataloader(
    X_train, y_train, X_val, y_val, X_test, y_test, batch_size: int = 32
):
    """
    Create PyTorch DataLoaders for training, validation, and testing.

    Args:
        X_train, y_train: Training data and labels.
        X_val, y_val: Validation data and labels.
        X_test, y_test: Test data and labels.
        batch_size (int): Batch size for the DataLoaders.

    Returns:
        Tuple of (train_loader, val_loader, test_loader, scaler)
    """
    # Initialize a scaler
    scaler = StandardScaler()

    # Fit and transform the training data
    X_train_scaled = scaler.fit_transform(X_train)
    X_val_scaled = scaler.transform(X_val)
    X_test_scaled = scaler.transform(X_test)

    # Convert to PyTorch tensors - explicitly set dtype to float32
    X_train_tensor = torch.tensor(X_train_scaled, dtype=torch.float32)
    y_train_tensor = torch.tensor(y_train.values, dtype=torch.float32)

    X_val_tensor = torch.tensor(X_val_scaled, dtype=torch.float32)
    y_val_tensor = torch.tensor(y_val.values, dtype=torch.float32)

    X_test_tensor = torch.tensor(X_test_scaled, dtype=torch.float32)
    y_test_tensor = torch.tensor(y_test.values, dtype=torch.float32)

    # Create TensorDatasets
    train_dataset = TensorDataset(X_train_tensor, y_train_tensor)
    val_dataset = TensorDataset(X_val_tensor, y_val_tensor)
    test_dataset = TensorDataset(X_test_tensor, y_test_tensor)

    # Create DataLoaders
    train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
    val_loader = DataLoader(val_dataset, batch_size=batch_size)
    test_loader = DataLoader(test_dataset, batch_size=batch_size)

    return train_loader, val_loader, test_loader, scaler

4. Flusso di lavoro di modellazione standard

Il codice nella cella successiva implementa un flusso di lavoro di modellazione PyTorch standard con l'integrazione MLflow, seguendo questa procedura:

  1. Generare ed esplorare i dati sintetici.
  2. Suddividere i dati in insiemi di training, convalida e test.
  3. Ridimensionare i dati e creare PyTorch DataLoaders.
  4. Definire ed eseguire il training di un modello di rete neurale.
  5. Valutare le prestazioni del modello.
  6. Registrare metriche, parametri e artefatti in MLflow.

Questo flusso di lavoro standard fornisce un modello di base. È quindi possibile usare l'ottimizzazione degli iperparametri per migliorare il modello.

# Create the regression dataset
n_samples = 1000
n_features = 10
X, y = create_regression_data(n_samples=n_samples, n_features=n_features, nonlinear=True)

# Create EDA plots
dist_plot = plot_feature_distributions(X, y)
corr_plot = plot_correlation_heatmap(X, y)
scatter_plot = plot_feature_target_relationships(X, y)
corr_with_target = X.corrwith(y).abs().sort_values(ascending=False)
top_features = corr_with_target.head(4).index.tolist()
pairwise_plot = plot_pairwise_relationships(X, y, top_features)
outlier_plot = plot_outliers(X)

# Split the data into train, validation, and test sets
X_train, X_temp, y_train, y_temp = train_test_split(X, y, test_size=0.3, random_state=42)
X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.5, random_state=42)

# Prepare DataLoaders
batch_size = 32
train_loader, val_loader, test_loader, scaler = prepare_dataloader(
    X_train, y_train, X_val, y_val, X_test, y_test, batch_size=batch_size)

# Define model parameters
input_dim = X_train.shape[1]
hidden_dims = [64, 32]
dropout_rate = 0.1
use_layer_norm = True
learning_rate = 1e-3
weight_decay = 1e-5

# Create the PyTorch Lightning model
model = RegressionLightningModule(
    input_dim=input_dim,
    hidden_dims=hidden_dims,
    dropout_rate=dropout_rate,
    use_layer_norm=use_layer_norm,
    learning_rate=learning_rate,
    weight_decay=weight_decay
)

# Define early stopping and model checkpoint callbacks
early_stopping = EarlyStopping(
    monitor='val_loss',
    patience=10,
    mode='min'
)

checkpoint_callback = ModelCheckpoint(
    monitor='val_loss',
    dirpath='./checkpoints',
    filename='pytorch-regression-{epoch:02d}-{val_loss:.4f}',
    save_top_k=1,
    mode='min'
)

# Define trainer
trainer = pl.Trainer(
    max_epochs=100,
    callbacks=[early_stopping, checkpoint_callback],
    enable_progress_bar=True,
    log_every_n_steps=5
)

# Train the model
trainer.fit(model, train_loader, val_loader)

# Test the model
test_results = trainer.test(model, test_loader)

# Make predictions on the test set for evaluation
model.eval()
test_preds = []
true_values = []

with torch.no_grad():
    for batch in test_loader:
        x, y = batch
        y_pred = model(x)
        test_preds.extend(y_pred.numpy())
        true_values.extend(y.numpy())

test_preds = np.array(test_preds)
true_values = np.array(true_values)

# Calculate metrics
rmse = np.sqrt(mean_squared_error(true_values, test_preds))
mae = mean_absolute_error(true_values, test_preds)
r2 = r2_score(true_values, test_preds)

# Create residual plot
residual_plot = plot_residuals(pd.Series(true_values), test_preds)

5. Registrare il modello usando MLflow

Quando si registra un modello usando MLflow in Databricks, vengono acquisiti elementi e metadati importanti. Ciò garantisce che il modello non solo sia riproducibile, ma anche pronto per la distribuzione con tutte le dipendenze necessarie e contratti API chiari. Per informazioni dettagliate su ciò che viene registrato, vedere la documentazione di MLflow.

Il codice nella cella successiva avvia un'esecuzione MLflow usando with mlflow.start_run():. Inizializza la gestione del contesto MLflow per l'esecuzione e racchiude l'esecuzione in un blocco di codice. Al termine del blocco di codice, vengono salvate tutte le metriche, i parametri e gli artefatti registrati e l'esecuzione di MLflow viene terminata automaticamente.

# Log the model and training results with MLflow
with mlflow.start_run() as run:
    # Create MLflow client for batch logging
    mlflow_client = MlflowClient()
    run_id = run.info.run_id

    # Extract metrics
    final_train_loss = trainer.callback_metrics.get("train_loss").item() if "train_loss" in trainer.callback_metrics else None
    final_val_loss = trainer.callback_metrics.get("val_loss").item() if "val_loss" in trainer.callback_metrics else None

    # Extract parameters for logging
    model_params = model.get_params()

    # Create a list to store all metrics for batch logging
    all_metrics = []

    # Add each metric to the list
    if final_train_loss is not None:
        all_metrics.append(Metric(key="train_loss", value=final_train_loss, timestamp=0, step=0))
    if final_val_loss is not None:
        all_metrics.append(Metric(key="val_loss", value=final_val_loss, timestamp=0, step=0))

    # Add test metrics
    all_metrics.append(Metric(key="test_rmse", value=rmse, timestamp=0, step=0))
    all_metrics.append(Metric(key="test_mae", value=mae, timestamp=0, step=0))
    all_metrics.append(Metric(key="test_r2", value=r2, timestamp=0, step=0))

    # Collect all parameters to log
    # Note: The code uses log_params for model_params since there could be many parameters,
    # but converts the individual param calls to batch
    from mlflow.entities import Param
    all_params = [
        Param(key="batch_size", value=str(batch_size)),
        Param(key="early_stopping_patience", value=str(early_stopping.patience)),
        Param(key="max_epochs", value=str(trainer.max_epochs)),
        Param(key="actual_epochs", value=str(trainer.current_epoch))
    ]

    # Generate a model signature using the infer signature utility in MLflow
    input_example = X_train.iloc[[0]].values.astype(np.float32)  # Ensure float32 type
    input_example_scaled = scaler.transform(input_example).astype(np.float32)

    model.eval()
    with torch.no_grad():
        # Ensure tensor is float32
        tensor_input = torch.tensor(input_example_scaled, dtype=torch.float32)
        signature_preds = model(tensor_input)

    # Ensure prediction is also float32
    signature = infer_signature(input_example, signature_preds.numpy().reshape(-1).astype(np.float32))

    # Log model parameters first (since these could be numerous)
    mlflow.log_params(model_params)

    # Log all metrics and remaining parameters in a single batch operation
    mlflow_client.log_batch(
        run_id=run_id,
        metrics=all_metrics,
        params=all_params
    )

    # Log the model to MLflow and register the model to Unity Catalog
    model_info = mlflow.pytorch.log_model(
        model,
        artifact_path="model",
        input_example=input_example,
        signature=signature,
        registered_model_name="demo.pytorch_regression_model",
    )

    # Log feature analysis plots
    mlflow.log_figure(dist_plot, "feature_distributions.png")
    mlflow.log_figure(corr_plot, "correlation_heatmap.png")
    mlflow.log_figure(scatter_plot, "feature_target_relationships.png")
    mlflow.log_figure(pairwise_plot, "pairwise_relationships.png")
    mlflow.log_figure(outlier_plot, "outlier_detection.png")
    mlflow.log_figure(residual_plot, "residual_plot.png")

    # Run MLflow evaluation to generate additional metrics without having to implement them
    evaluation_data = X_test.copy()
    evaluation_data["label"] = y_test

    # Skip mlflow.evaluate for now to avoid type mismatch issues
    # Instead, log the metrics directly
    print(f"Model logged: {model_info.model_uri}")
    print(f"Test RMSE: {rmse:.4f}")
    print(f"Test MAE: {mae:.4f}")
    print(f"Test R²: {r2:.4f}")

6. Ottimizzazione degli iperparametri

Questa sezione illustra come automatizzare l'ottimizzazione degli iperparametri usando Optuna e le esecuzioni annidate in MLflow. In questo modo è possibile esplorare una gamma di configurazioni dei parametri e acquisire tutti i dettagli sperimentali.

Il codice nella cella successiva esegue le operazioni seguenti:

  1. Usa la create_regression_data funzione definita in precedenza per generare un set di dati di regressione sintetica.

  2. Suddivide il set di dati in set di dati di training e test separati e salva una copia del set di dati di test per la valutazione.

  3. Crea una funzione obiettivo per il processo di ottimizzazione degli iperparametri. La funzione obiettivo definisce lo spazio di ricerca per gli iperparametri del modello PyTorch, ad esempio il numero di livelli, dimensioni nascoste, frequenza di rilascio, frequenza di apprendimento e parametri di regolarizzazione. Optuna campiona dinamicamente questi valori, garantendo che ogni trial testi una combinazione diversa di parametri.

  4. Avvia un'esecuzione di MLflow annidata all'interno della funzione objective. Questa esecuzione annidata acquisisce e registra automaticamente tutti i dettagli specifici della prova dell'iperparametro corrente. Isolando ogni prova nella propria esecuzione nidificata, è possibile mantenere un record ben organizzato di ogni configurazione e le metriche di prestazione corrispondenti. L'esecuzione annidata registra quanto segue:

    • Iperparametri specifici usati per quell'esperimento.
    • La metrica delle prestazioni (in questo caso, perdita di convalida) calcolata nel set di test.
    • L'istanza del modello addestrato viene archiviata anche come parte dei metadati dell'esperimento. In questo modo è più semplice recuperare il modello con prestazioni migliori in un secondo momento.

    Il codice non registra ogni modello in MLflow. Durante l'ottimizzazione degli iperparametri, non è garantito che ogni iterazione sia particolarmente buona, quindi non esiste alcun motivo per registrare l'artefatto del modello per ognuno di essi.

  5. Crea un run padre di MLflow. Questa esecuzione avvia uno studio Optuna progettato per identificare il set ottimale di iperparametri (il set che produce la perdita di convalida più bassa). Optuna esegue una serie di versioni di valutazione in cui ogni versione di valutazione usa una combinazione univoca di iperparametri. Durante ogni versione di valutazione, l'esecuzione annidata di MLflow acquisisce tutti i dettagli dell'esperimento, in modo da poter tenere traccia e confrontare le prestazioni di ogni configurazione del modello.

  6. Lo studio identifica la prova migliore in base alla perdita di convalida più bassa. Il codice estrae il modello migliore e i valori dei parametri ottimali. Il codice usa infer_signature per salvare una firma del modello, che specifica gli schemi di input e output previsti ed è importante per la distribuzione e l'integrazione coerenti con sistemi come Unity Catalog. Infine, il modello migliore viene registrato e registrato in Unity Catalog. Vengono registrati anche elementi aggiuntivi, ad esempio tracciati EDA e grafici di importanza delle caratteristiche.

# Create a custom pruning callback as a fallback
class PyTorchLightningPruningCallback(pl.Callback):
    """PyTorch Lightning callback to prune unpromising trials.

    This is a simplified version for use when the optuna-integration package isn't available.
    """

    def __init__(self, trial, monitor):
        super().__init__()
        self._trial = trial
        self.monitor = monitor

    def on_validation_end(self, trainer, pl_module):
        # Report the validation metric to Optuna
        metrics = trainer.callback_metrics
        current_score = metrics.get(self.monitor)

        if current_score is not None:
            self._trial.report(current_score.item(), trainer.current_epoch)

            # If trial should be pruned based on current value,
            # stop the training
            if self._trial.should_prune():
                message = "Trial was pruned at epoch {}.".format(trainer.current_epoch)
                raise optuna.TrialPruned(message)

# Generate a larger dataset for hyperparameter tuning
n_samples = 2000
n_features = 10

X, y = create_regression_data(n_samples=n_samples, n_features=n_features, nonlinear=True)

# Split the data
X_train, X_temp, y_train, y_temp = train_test_split(X, y, test_size=0.3, random_state=42)
X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.5, random_state=42)

# Prepare the evaluation data
evaluation_data = X_test.copy()
evaluation_data["label"] = y_test

# Create the data loaders
batch_size = 32
train_loader, val_loader, test_loader, scaler = prepare_dataloader(
    X_train, y_train, X_val, y_val, X_test, y_test, batch_size=batch_size)

def objective(trial):
    """Optuna objective function to minimize validation loss."""

    # Define the hyperparameter search space
    n_layers = trial.suggest_int("n_layers", 1, 3)

    # Create hidden dimensions based on number of layers
    hidden_dims = []
    for i in range(n_layers):
        hidden_dims.append(trial.suggest_int(f"hidden_dim_{i}", 16, 128))

    # Other hyperparameters
    dropout_rate = trial.suggest_float("dropout_rate", 0.0, 0.5)
    learning_rate = trial.suggest_float("learning_rate", 1e-4, 1e-2, log=True)
    weight_decay = trial.suggest_float("weight_decay", 1e-6, 1e-3, log=True)
    use_layer_norm = trial.suggest_categorical("use_layer_norm", [True, False])

    # Start a nested MLflow run for this trial
    with mlflow.start_run(nested=True) as child_run:
        # Create MLflow client for batch logging
        mlflow_client = MlflowClient()
        run_id = child_run.info.run_id

        # Prepare parameters for batch logging
        params_list = []
        param_dict = {
            "n_layers": n_layers,
            "hidden_dims": str(hidden_dims),  # Convert list to string
            "dropout_rate": dropout_rate,
            "learning_rate": learning_rate,
            "weight_decay": weight_decay,
            "use_layer_norm": use_layer_norm,
            "batch_size": batch_size
        }

        # Convert parameters to Param objects
        for key, value in param_dict.items():
            params_list.append(Param(key, str(value)))

        # Create the model with these hyperparameters
        model = RegressionLightningModule(
            input_dim=X_train.shape[1],
            hidden_dims=hidden_dims,
            dropout_rate=dropout_rate,
            use_layer_norm=use_layer_norm,
            learning_rate=learning_rate,
            weight_decay=weight_decay
        )

        # Callbacks
        early_stopping = EarlyStopping(
            monitor='val_loss',
            patience=5,
            mode='min'
        )

        pruning_callback = PyTorchLightningPruningCallback(
            trial, monitor="val_loss"
        )

        # Define trainer with early stopping and pruning
        trainer = pl.Trainer(
            max_epochs=50,
            callbacks=[early_stopping, pruning_callback],
            enable_progress_bar=False,
            log_every_n_steps=10
        )

        # Train and validate the model
        trainer.fit(model, train_loader, val_loader)

        # Get the best validation loss
        best_val_loss = trainer.callback_metrics.get("val_loss").item()
        val_rmse = np.sqrt(best_val_loss)

        # Prepare metrics for batch logging
        current_time = int(time.time() * 1000)  # Current time in milliseconds
        metrics_list = [
            Metric("val_loss", best_val_loss, current_time, 0),
            Metric("val_rmse", val_rmse, current_time, 0)
        ]

        # Use log_batch through the client for efficient logging
        mlflow_client.log_batch(run_id, metrics=metrics_list, params=params_list)

    # Store the model in the trial's user attributes
    trial.set_user_attr("model", model)

    # Return the value to minimize (validation loss)
    return best_val_loss

best_model_version = None
# The parent run stores the best iteration from the hyperparameter tuning execution
with mlflow.start_run() as run:
    # Create MLflow client for batch logging
    mlflow_client = MlflowClient()
    run_id = run.info.run_id

    study = optuna.create_study(direction="minimize")
    study.optimize(objective, n_trials=20)

    best_trial = study.best_trial
    best_model = best_trial.user_attrs["model"]

    # Test the best model
    trainer = pl.Trainer(
        enable_progress_bar=True,
        log_every_n_steps=5
    )
    test_results = trainer.test(best_model, test_loader)

    # Make predictions on the test set for evaluation
    best_model.eval()
    test_preds = []
    true_values = []

    with torch.no_grad():
        for batch in test_loader:
            x, y = batch
            y_pred = best_model(x)
            test_preds.extend(y_pred.numpy())
            true_values.extend(y.numpy())

    test_preds = np.array(test_preds)
    true_values = np.array(true_values)

    # Calculate metrics
    rmse = np.sqrt(mean_squared_error(true_values, test_preds))
    mae = mean_absolute_error(true_values, test_preds)
    r2 = r2_score(true_values, test_preds)

    # Prepare parameters for batch logging
    best_params_list = []
    for key, value in best_trial.params.items():
        best_params_list.append(Param(f"best_{key}", str(value)))

    # Prepare metrics for batch logging
    current_time = int(time.time() * 1000)  # Current time in milliseconds
    metrics_list = [
        Metric("best_val_loss", best_trial.value, current_time, 0),
        Metric("test_rmse", rmse, current_time, 0),
        Metric("test_mae", mae, current_time, 0),
        Metric("test_r2", r2, current_time, 0)
    ]

    # Log metrics and parameters in a single batch call
    mlflow_client.log_batch(run_id, metrics=metrics_list, params=best_params_list)

    # Generate model signature - ensure consistent float32 types
    input_example = X_train.iloc[[0]].values.astype(np.float32)
    input_example_scaled = scaler.transform(input_example).astype(np.float32)

    best_model.eval()
    with torch.no_grad():
        tensor_input = torch.tensor(input_example_scaled, dtype=torch.float32)
        signature_preds = best_model(tensor_input)

    signature = infer_signature(input_example, signature_preds.numpy().reshape(-1).astype(np.float32))

    # Log and register the PyTorch model
    model_info = mlflow.pytorch.log_model(
        best_model,
        artifact_path="model",
        input_example=input_example,
        signature=signature,
        registered_model_name="demo.pytorch_regression_optimized",
    )

    # Create residual plot
    residual_plot = plot_residuals(pd.Series(true_values), test_preds)

    # Log figures (no batch equivalent for figures)
    mlflow.log_figure(dist_plot, "feature_distributions.png")
    mlflow.log_figure(corr_plot, "correlation_heatmap.png")
    mlflow.log_figure(scatter_plot, "feature_target_relationships.png")
    mlflow.log_figure(pairwise_plot, "pairwise_relationships.png")
    mlflow.log_figure(outlier_plot, "outlier_detection.png")
    mlflow.log_figure(residual_plot, "residual_plot.png")

    # Skip mlflow.evaluate for now to avoid type mismatch issues
    # Instead, log the metrics directly
    print(f"Best model logged: {model_info.model_uri}")
    print(f"Best parameters: {best_trial.params}")
    print(f"Test RMSE: {rmse:.4f}")
    print(f"Test MAE: {mae:.4f}")
    print(f"Test R²: {r2:.4f}")

    best_model_version = model_info.registered_model_version
from mlflow import MlflowClient

# Initialize MLflow client
client = MlflowClient()

# Set a human-readable alias for the best model version
# This makes it easier to reference specific model versions programmatically
client.set_registered_model_alias("demo.pytorch_regression_optimized", "best", int(best_model_version))

7. Convalida preliminare della distribuzione

MLflow fornisce l'utilità mlflow.models.predict per simulare un ambiente simile alla produzione e verificare che il modello sia configurato correttamente.

# Reference the model by its alias
model_uri = "models:/demo.pytorch_regression_optimized@best"

# Validate the model's deployment readiness
mlflow.models.predict(model_uri=model_uri, input_data=X_test, env_manager="local")

8. Caricare il modello registrato ed eseguire stime

Il codice in questa sezione illustra come caricare il modello registrato da MLflow e usarlo per eseguire stime in locale. Ciò è utile per i test o per gli scenari di inferenza batch.

# Convert the data type of X_test to float32
X_test = X_test.astype('float32')

# Load the model using the pyfunc interface (recommended for deployment)
loaded_model = mlflow.pyfunc.load_model(model_uri=model_uri)

# Make predictions with the loaded model
predictions = loaded_model.predict(X_test)

print(f"Shape of predictions: {predictions.shape}")
print(f"First 5 predictions: {predictions[:5]}")
print(f"First 5 actual values: {y_test.values[:5]}")

9. Predizione batch usando UDF (User Defined Functions) di Spark in MLflow

Per le stime su larga scala, è possibile convertire il modello in una funzione definita dall'utente (UDF) di Spark e applicarlo a un dataframe Spark, per abilitare l'inferenza distribuita.

from pyspark.sql.functions import array, col

# Convert the test data to a Spark DataFrame
X_spark = spark.createDataFrame(X_test)

# Create an array of all feature columns
# This step is necessary because:
# 1. The PyTorch model expects an input tensor with shape [-1, 13]
# 2. The model_udf needs to receive each row as a single array of 13 values
# 3. Without this array transformation, 13 separate columns would be passed to the model
#    which wouldn't match the expected tensor structure
X_spark_with_array = X_spark.withColumn(
    "features_array",
    array(*[col(c) for c in X_spark.columns])
)

# Create a Spark UDF from the registered model
model_udf = mlflow.pyfunc.spark_udf(spark, model_uri=model_uri)

# Apply MLflow UDF to the array column
# Pass the single array column to the model, which matches the expected tensor format
X_spark_with_predictions = X_spark_with_array.withColumn(
    "prediction",
    model_udf("features_array")
)

display(X_spark_with_predictions.limit(5))

Notebook di esempio

Modello di Deep Learning end-to-end con PyTorch

Ottieni il notebook