File size: 8,750 Bytes
431bcf6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
"""
Evaluate ALL models (Ridge, XGBoost, Random Forest, Fine-tuned BERT)
on the test set. Generates full comparison table + chart.

Usage:
    python -m src.evaluation.evaluate_all_models
"""
import os, json
import numpy as np
import pandas as pd
import joblib
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score

BASE_DIR   = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
DATA_DIR   = os.path.join(BASE_DIR, "data")
MODELS_DIR = os.path.join(BASE_DIR, "models")


def rmse(y, p):
    return float(np.sqrt(mean_squared_error(y, p)))


def evaluate_classical(name: str, model_path: str, X_test, y_test) -> dict:
    model = joblib.load(model_path)
    preds = model.predict(X_test)
    return {
        "model": name,
        "RMSE":  round(rmse(y_test, preds), 4),
        "MAE":   round(float(mean_absolute_error(y_test, preds)), 4),
        "R²":    round(float(r2_score(y_test, preds)), 4),
    }


def evaluate_bert_regressor(model_pt_path: str, X_test_bert, y_test) -> dict:
    """
    Load the fine-tuned BERT regressor and evaluate.
    Falls back gracefully if model not yet downloaded from Colab.
    """
    if not os.path.exists(model_pt_path):
        print(f"  ⚠️  Fine-tuned BERT model not found at: {model_pt_path}")
        print("  →  Please train on Colab and download bert_regressor_best.pt")
        return {"model": "Fine-tuned BERT", "RMSE": "N/A", "MAE": "N/A", "R²": "N/A"}

    import torch
    from src.models.bert_regressor import BertRegressorModel

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model  = BertRegressorModel()
    model.load_state_dict(torch.load(model_pt_path, map_location=device))
    model.eval().to(device)

    # X_test_bert shape: (N, 768) – raw BERT embeddings
    # For the fine-tuned model we need tokenised text, so we skip re-tokenisation
    # and instead use the frozen embeddings as a proxy (quick evaluation mode).
    # Full text-based evaluation requires raw text → use Colab metrics JSON if available.
    bert_metrics_path = os.path.join(MODELS_DIR, "bert_test_metrics.json")
    if os.path.exists(bert_metrics_path):
        with open(bert_metrics_path) as f:
            m = json.load(f)
        return {
            "model": "Fine-tuned BERT",
            "RMSE":  round(m["test_rmse"], 4),
            "MAE":   round(m["test_mae"],  4),
            "R²":    round(m["test_r2"],   4),
        }

    print("  ⚠️  bert_test_metrics.json not found. Using embedding-proxy evaluation.")
    # Proxy: pass frozen embeddings through the regressor head only
    with torch.no_grad():
        t = torch.tensor(X_test_bert, dtype=torch.float32).to(device)
        # Direct head pass (no BERT backbone) — approximate
        preds = (model.regressor(t).squeeze(-1) * 99.0).cpu().numpy()
    return {
        "model": "Fine-tuned BERT (proxy)",
        "RMSE":  round(rmse(y_test, preds), 4),
        "MAE":   round(float(mean_absolute_error(y_test, preds)), 4),
        "R²":    round(float(r2_score(y_test, preds)), 4),
    }


def plot_comparison(results: list[dict], save_path: str):
    models = [r["model"] for r in results]
    rmse_v = [r["RMSE"] if isinstance(r["RMSE"], float) else 0 for r in results]
    mae_v  = [r["MAE"]  if isinstance(r["MAE"],  float) else 0 for r in results]
    r2_v   = [r["R²"]   if isinstance(r["R²"],   float) else 0 for r in results]

    x      = np.arange(len(models))
    width  = 0.25

    fig, axes = plt.subplots(1, 2, figsize=(14, 6))
    fig.patch.set_facecolor('#0A192F')
    for ax in axes:
        ax.set_facecolor('#112240')

    # Left: RMSE + MAE
    ax1 = axes[0]
    bars1 = ax1.bar(x - width/2, rmse_v, width, label='RMSE', color='#FF6B6B', alpha=0.85)
    bars2 = ax1.bar(x + width/2, mae_v,  width, label='MAE',  color='#FFB347', alpha=0.85)
    ax1.set_xticks(x); ax1.set_xticklabels(models, rotation=15, ha='right', color='#80d8ff', fontsize=9)
    ax1.set_ylabel('Error (Lower is Better)', color='#80d8ff')
    ax1.set_title('RMSE & MAE Comparison', color='#00e5ff', fontweight='bold')
    ax1.tick_params(colors='#80d8ff')
    ax1.legend(facecolor='#0A192F', labelcolor='#E6F1FF')
    for bar in list(bars1) + list(bars2):
        h = bar.get_height()
        if h > 0:
            ax1.text(bar.get_x() + bar.get_width()/2, h + 0.3, f'{h:.2f}',
                     ha='center', va='bottom', fontsize=8, color='#E6F1FF')

    # Right: R²
    ax2 = axes[1]
    colors = ['#546e7a', '#00b0ff', '#00e5ff', '#ce93d8']
    bars3  = ax2.bar(x, r2_v, 0.5, color=colors[:len(models)], alpha=0.85)
    ax2.set_xticks(x); ax2.set_xticklabels(models, rotation=15, ha='right', color='#80d8ff', fontsize=9)
    ax2.set_ylabel('R² Score (Higher is Better)', color='#80d8ff')
    ax2.set_title('R² Score Comparison', color='#00e5ff', fontweight='bold')
    ax2.tick_params(colors='#80d8ff')
    for bar in bars3:
        h = bar.get_height()
        if h > 0:
            ax2.text(bar.get_x() + bar.get_width()/2, h + 0.002, f'{h:.3f}',
                     ha='center', va='bottom', fontsize=9, color='#E6F1FF', fontweight='bold')

    plt.suptitle('All Model Performance Comparison — Pandora System', color='#00e5ff',
                 fontsize=13, fontweight='bold', y=1.01)
    plt.tight_layout()
    plt.savefig(save_path, dpi=200, bbox_inches='tight', facecolor='#0A192F')
    plt.close()
    print(f"Comparison chart saved -> {save_path}")


def plot_training_curve(log_csv: str, save_path: str):
    if not os.path.exists(log_csv):
        print(f"Training log not found: {log_csv}")
        return
    df  = pd.read_csv(log_csv)
    fig, axes = plt.subplots(1, 2, figsize=(12, 4))
    fig.patch.set_facecolor('#0A192F')
    for ax in axes:
        ax.set_facecolor('#112240')

    # Training loss
    axes[0].plot(df["epoch"], df["train_loss"], marker='o', color='#00e5ff', linewidth=2)
    axes[0].set_title('Training Loss per Epoch', color='#00e5ff', fontweight='bold')
    axes[0].set_xlabel('Epoch', color='#80d8ff')
    axes[0].set_ylabel('MSE Loss', color='#80d8ff')
    axes[0].tick_params(colors='#80d8ff')

    # Val RMSE
    axes[1].plot(df["epoch"], df["val_rmse"], marker='s', color='#FF6B6B', linewidth=2, label='RMSE')
    axes[1].plot(df["epoch"], df["val_r2"],   marker='^', color='#69F0AE', linewidth=2, label='R²')
    axes[1].set_title('Validation Metrics per Epoch', color='#00e5ff', fontweight='bold')
    axes[1].set_xlabel('Epoch', color='#80d8ff')
    axes[1].tick_params(colors='#80d8ff')
    axes[1].legend(facecolor='#0A192F', labelcolor='#E6F1FF')

    plt.suptitle('Fine-Tuned BERT — Training Curve', color='#00e5ff', fontsize=12, fontweight='bold')
    plt.tight_layout()
    plt.savefig(save_path, dpi=200, bbox_inches='tight', facecolor='#0A192F')
    plt.close()
    print(f"Training curve saved -> {save_path}")


def main():
    # Load test features
    print("Loading test data...")
    test_df  = pd.read_csv(os.path.join(DATA_DIR, "test_features.csv"))
    y_test   = test_df["extraversion"].values
    X_test_c = test_df.drop(columns=["extraversion"]).values
    X_test_b = np.load(os.path.join(DATA_DIR, "test_bert_embeddings.npy"))
    X_test   = np.hstack((X_test_c, X_test_b))

    results = []
    print("\nEvaluating models...")

    results.append(evaluate_classical(
        "Ridge (Baseline)",
        os.path.join(MODELS_DIR, "classical_ridge_model.pkl"),
        X_test_c[:, :1018], y_test
    ))
    results.append(evaluate_classical(
        "XGBoost",
        os.path.join(MODELS_DIR, "advanced_xgboost_model_local.pkl"),
        X_test, y_test
    ))

    # Try Random Forest
    rf_path = os.path.join(MODELS_DIR, "random_forest_model_local.pkl")
    if os.path.exists(rf_path):
        results.append(evaluate_classical("Random Forest", rf_path, X_test, y_test))

    results.append(evaluate_bert_regressor(
        os.path.join(MODELS_DIR, "bert_regressor_best.pt"),
        X_test_b, y_test
    ))

    # Print table
    print("\n" + "="*60)
    print(f"{'Model':<25} {'RMSE':>8} {'MAE':>8} {'R²':>8}")
    print("-"*60)
    for r in results:
        print(f"{r['model']:<25} {str(r['RMSE']):>8} {str(r['MAE']):>8} {str(r['R²']):>8}")
    print("="*60)

    # Save CSV
    pd.DataFrame(results).to_csv(os.path.join(MODELS_DIR, "full_model_comparison.csv"), index=False)
    print("\nFull comparison CSV saved.")

    # Charts
    plot_comparison(results, os.path.join(MODELS_DIR, "full_model_comparison.png"))
    plot_training_curve(
        os.path.join(MODELS_DIR, "bert_training_log.csv"),
        os.path.join(MODELS_DIR, "bert_training_curve.png")
    )


if __name__ == "__main__":
    main()