SGD training snippet for Medium post

sgd_training.py

def sgd(x_train, y_train, x_valid, y_valid, variance_threshold=0.1):    
    threshold = VarianceThreshold(variance_threshold)
        
    sgd_classifier = SGDClassifier(
        alpha=1./len(x_train),
        class_weight='balanced',
        loss='log', penalty='elasticnet', 
        fit_intercept=False, tol=0.001, n_jobs=-1)
    
    bagging = BaggingClassifier(
        base_estimator=sgd_classifier,
        bootstrap_features=True,
        n_jobs=-1, max_samples=0.5, max_features=0.5)    
    
    x_thresh = threshold.fit_transform(x_train)
    bagging.fit(x_thresh, y_train)
    train_metrics = build_metrics(bagging, x_thresh, y_train)

    x_thresh = threshold.transform(x_valid)
    valid_metrics = build_metrics(bagging, x_thresh, y_valid)

    return bagging, train_metrics, valid_metrics    


def build_metrics(model, X, y):
    probs = model.predict_proba(X)
    preds = np.argmax(probs, axis=1)
    metrics = dict(
        probs=probs,
        preds=preds,
        loss=log_loss(y, probs),
        accuracy=np.mean(preds == y))
    return namedtuple('Predictions', metrics.keys())(**metrics)