Capítulo 2: Weights and Biases

Introdução ao W&B

Weights & Biases (W&B) é uma poderosa plataforma de MLOps para rastreamento de experimentos, versionamento de conjuntos de dados e gerenciamento de modelos, com excelentes recursos de visualização.

Configuração e Instalação

# Install
pip install wandb

# Login
wandb login

# Or set API key
export WANDB_API_KEY="your-api-key"

Rastreamento Básico de Experimentos

import wandb
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, roc_auc_score

# Initialize a run
wandb.init(
    project="customer-churn",
    name="rf-baseline",
    config={
        "n_estimators": 100,
        "max_depth": 10,
        "learning_rate": 0.1,
        "dataset": "v2.0"
    }
)

# Access config
config = wandb.config

# Train model
model = RandomForestClassifier(
    n_estimators=config.n_estimators,
    max_depth=config.max_depth
)
model.fit(X_train, y_train)

# Log metrics
accuracy = accuracy_score(y_test, model.predict(X_test))
auc = roc_auc_score(y_test, model.predict_proba(X_test)[:, 1])

wandb.log({
    "accuracy": accuracy,
    "auc_roc": auc,
    "epoch": 1
})

# Finish run
wandb.finish()

Sweeps de Hiperparâmetros

# Define sweep configuration
sweep_config = {
    "method": "bayes",  # grid, random, bayes
    "metric": {"name": "auc_roc", "goal": "maximize"},
    "parameters": {
        "n_estimators": {"values": [100, 200, 300]},
        "max_depth": {"min": 5, "max": 20},
        "min_samples_split": {"values": [2, 5, 10]},
        "learning_rate": {"distribution": "log_uniform_values", "min": 0.001, "max": 0.1}
    }
}

# Create sweep
sweep_id = wandb.sweep(sweep_config, project="customer-churn")

# Training function
def train():
    wandb.init()
    config = wandb.config
    
    model = GradientBoostingClassifier(
        n_estimators=config.n_estimators,
        max_depth=config.max_depth,
        learning_rate=config.learning_rate
    )
    model.fit(X_train, y_train)
    
    auc = roc_auc_score(y_test, model.predict_proba(X_test)[:, 1])
    wandb.log({"auc_roc": auc})

# Run sweep
wandb.agent(sweep_id, train, count=50)

Artifacts do W&B

# Log dataset as artifact
artifact = wandb.Artifact("training-data", type="dataset")
artifact.add_file("data/train.csv")
artifact.add_dir("data/features/")
wandb.log_artifact(artifact)

# Log model as artifact
model_artifact = wandb.Artifact("churn-model", type="model")
model_artifact.add_file("models/model.pkl")
wandb.log_artifact(model_artifact)

# Use artifact in another run
wandb.init(project="customer-churn")
artifact = wandb.use_artifact("training-data:latest")
artifact_dir = artifact.download()

W&B Tables para Visualização de Dados

import wandb
import pandas as pd

wandb.init(project="customer-churn")

# Create a W&B Table
table = wandb.Table(dataframe=df.head(100))
wandb.log({"data_sample": table})

# Log predictions table
predictions_df = pd.DataFrame({
    "customer_id": test_ids,
    "actual": y_test,
    "predicted": predictions,
    "probability": probabilities
})
wandb.log({"predictions": wandb.Table(dataframe=predictions_df)})

Principais Pontos

Summary

1. W&B se destaca em visualização com dashboards interativos
2. Sweeps automatizam a otimização de hiperparâmetros com métodos bayesianos
3. Artifacts rastreiam conjuntos de dados e modelos com linhagem
4. Tables permitem exploração rica de dados na interface
5. Reports facilitam a colaboração em equipe

---

Próximo: Capítulo 3: Comparando Experimentos