Spaces:
Sleeping
Sleeping
errors
Browse files- SVM/multiclass_svm.py +2 -202
SVM/multiclass_svm.py
CHANGED
|
@@ -1,20 +1,11 @@
|
|
| 1 |
import numpy as np
|
| 2 |
-
import matplotlib.pyplot as plt
|
| 3 |
-
from sklearn.datasets import fetch_openml
|
| 4 |
-
from sklearn.model_selection import train_test_split
|
| 5 |
-
from sklearn.preprocessing import StandardScaler
|
| 6 |
-
from sklearn.metrics import (accuracy_score, precision_score, recall_score,
|
| 7 |
-
f1_score, classification_report,
|
| 8 |
-
confusion_matrix, ConfusionMatrixDisplay)
|
| 9 |
|
|
|
|
| 10 |
try:
|
| 11 |
from linear_svm import LinearSVMScartch
|
| 12 |
except ModuleNotFoundError:
|
| 13 |
from SVM.linear_svm import LinearSVMScartch
|
| 14 |
|
| 15 |
-
RANDOM_STATE = 42
|
| 16 |
-
|
| 17 |
-
|
| 18 |
# ---------------------------------------------------------------------------
|
| 19 |
# One-vs-Rest multiclass wrapper around LinearSVMScartch
|
| 20 |
# ---------------------------------------------------------------------------
|
|
@@ -59,195 +50,4 @@ class MultiClassSVMScratch:
|
|
| 59 |
def predict(self, X):
|
| 60 |
# Pick the class whose binary SVM is most confident (+highest score)
|
| 61 |
scores = self.decision_scores(X)
|
| 62 |
-
return self.classes_[np.argmax(scores, axis=1)]
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
# ---------------------------------------------------------------------------
|
| 66 |
-
# Data loading
|
| 67 |
-
# ---------------------------------------------------------------------------
|
| 68 |
-
|
| 69 |
-
def load_data(n_dev=5000):
|
| 70 |
-
print("Loading MNIST...")
|
| 71 |
-
mnist = fetch_openml('mnist_784', version=1, as_frame=False,
|
| 72 |
-
parser='auto', data_home='d:/sklearn_data')
|
| 73 |
-
X, y = mnist.data, mnist.target.astype(int)
|
| 74 |
-
|
| 75 |
-
# Normalize pixels to [0, 1]
|
| 76 |
-
X = X / 255.0
|
| 77 |
-
|
| 78 |
-
# 60 / 20 / 20 stratified split
|
| 79 |
-
X_trainval, X_test, y_trainval, y_test = train_test_split(
|
| 80 |
-
X, y, test_size=0.20, stratify=y, random_state=RANDOM_STATE)
|
| 81 |
-
X_train, X_val, y_train, y_val = train_test_split(
|
| 82 |
-
X_trainval, y_trainval, test_size=0.25, stratify=y_trainval,
|
| 83 |
-
random_state=RANDOM_STATE)
|
| 84 |
-
|
| 85 |
-
print(f"Train: {len(y_train)} Val: {len(y_val)} Test: {len(y_test)}")
|
| 86 |
-
|
| 87 |
-
# Standardize — fit on train only to avoid leakage
|
| 88 |
-
scaler = StandardScaler()
|
| 89 |
-
X_train_s = scaler.fit_transform(X_train)
|
| 90 |
-
X_val_s = scaler.transform(X_val)
|
| 91 |
-
X_test_s = scaler.transform(X_test)
|
| 92 |
-
|
| 93 |
-
# Small dev subset for fast hyperparameter search
|
| 94 |
-
rng = np.random.RandomState(RANDOM_STATE)
|
| 95 |
-
dev_idx = rng.choice(len(X_train_s), size=n_dev, replace=False)
|
| 96 |
-
X_dev, y_dev = X_train_s[dev_idx], y_train[dev_idx]
|
| 97 |
-
print(f"Dev subset: {len(y_dev)} samples")
|
| 98 |
-
|
| 99 |
-
return X_train_s, X_val_s, X_test_s, y_train, y_val, y_test, X_dev, y_dev, scaler
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
# ---------------------------------------------------------------------------
|
| 103 |
-
# Manual grid search
|
| 104 |
-
# ---------------------------------------------------------------------------
|
| 105 |
-
|
| 106 |
-
def manual_grid_search(X_dev, y_dev, X_val, y_val):
|
| 107 |
-
C_values = [0.1, 1.0, 10.0]
|
| 108 |
-
lr_values = [0.001, 0.0005, 0.0001]
|
| 109 |
-
|
| 110 |
-
results = []
|
| 111 |
-
print("\n--- Hyperparameter Grid Search ---")
|
| 112 |
-
print(f"{'C':>6} {'LR':>8} {'Val Acc':>9}")
|
| 113 |
-
print("-" * 30)
|
| 114 |
-
|
| 115 |
-
for C in C_values:
|
| 116 |
-
for lr in lr_values:
|
| 117 |
-
model = MultiClassSVMScratch(C=C, learning_rate=lr,
|
| 118 |
-
n_epochs=30, random_state=RANDOM_STATE)
|
| 119 |
-
model.fit(X_dev, y_dev)
|
| 120 |
-
acc = accuracy_score(y_val, model.predict(X_val))
|
| 121 |
-
results.append((C, lr, acc))
|
| 122 |
-
print(f"{C:>6} {lr:>8} {acc:>9.4f}")
|
| 123 |
-
|
| 124 |
-
best = max(results, key=lambda x: x[2])
|
| 125 |
-
print(f"\nBest: C={best[0]}, lr={best[1]}, Val Acc={best[2]:.4f}")
|
| 126 |
-
return best[0], best[1]
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
# ---------------------------------------------------------------------------
|
| 130 |
-
# Plotting helpers
|
| 131 |
-
# ---------------------------------------------------------------------------
|
| 132 |
-
|
| 133 |
-
def plot_loss_curves(model):
|
| 134 |
-
fig, ax = plt.subplots(figsize=(10, 5))
|
| 135 |
-
for k in model.classes_:
|
| 136 |
-
ax.plot(model.classifiers[k].history, label=f"Digit {k}")
|
| 137 |
-
ax.set_xlabel("Epoch")
|
| 138 |
-
ax.set_ylabel("Loss")
|
| 139 |
-
ax.set_title("Training Loss per Binary SVM (OvR)")
|
| 140 |
-
ax.legend(ncol=5, fontsize=8)
|
| 141 |
-
ax.grid(True)
|
| 142 |
-
plt.tight_layout()
|
| 143 |
-
plt.show()
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
def plot_confusion_matrix(y_true, y_pred):
|
| 147 |
-
cm = confusion_matrix(y_true, y_pred)
|
| 148 |
-
disp = ConfusionMatrixDisplay(cm, display_labels=list(range(10)))
|
| 149 |
-
fig, ax = plt.subplots(figsize=(9, 9))
|
| 150 |
-
disp.plot(ax=ax, cmap="Blues", colorbar=False)
|
| 151 |
-
ax.set_title("Confusion Matrix — Multiclass SVM (Scratch)")
|
| 152 |
-
plt.tight_layout()
|
| 153 |
-
plt.show()
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
def plot_decision_scores(model, X_sample, true_label):
|
| 157 |
-
scores = model.decision_scores(X_sample.reshape(1, -1))[0]
|
| 158 |
-
colors = ["steelblue"] * 10
|
| 159 |
-
colors[true_label] = "tomato"
|
| 160 |
-
fig, ax = plt.subplots(figsize=(8, 4))
|
| 161 |
-
ax.bar(model.classes_, scores, color=colors)
|
| 162 |
-
ax.set_xlabel("Digit class")
|
| 163 |
-
ax.set_ylabel("Decision score (w·x + b)")
|
| 164 |
-
ax.set_title(f"Decision scores for one sample (true label = {true_label})")
|
| 165 |
-
ax.axhline(0, color="black", linewidth=0.8, linestyle="--")
|
| 166 |
-
plt.tight_layout()
|
| 167 |
-
plt.show()
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
def plot_per_class_metrics(y_true, y_pred):
|
| 171 |
-
"""Bar chart + table showing accuracy, precision, recall, F1 per digit class."""
|
| 172 |
-
classes = np.arange(10)
|
| 173 |
-
cm = confusion_matrix(y_true, y_pred, labels=classes)
|
| 174 |
-
per_class_acc = cm.diagonal() / cm.sum(axis=1)
|
| 175 |
-
precision = precision_score(y_true, y_pred, average=None, labels=classes, zero_division=0)
|
| 176 |
-
recall = recall_score(y_true, y_pred, average=None, labels=classes, zero_division=0)
|
| 177 |
-
f1 = f1_score(y_true, y_pred, average=None, labels=classes, zero_division=0)
|
| 178 |
-
|
| 179 |
-
print("\n=== PER-CLASS METRICS ===")
|
| 180 |
-
print(f"{'Digit':>6} {'Accuracy':>9} {'Precision':>10} {'Recall':>8} {'F1':>8}")
|
| 181 |
-
print("-" * 50)
|
| 182 |
-
for c in classes:
|
| 183 |
-
print(f"{c:>6} {per_class_acc[c]:>9.4f} {precision[c]:>10.4f} {recall[c]:>8.4f} {f1[c]:>8.4f}")
|
| 184 |
-
|
| 185 |
-
x = np.arange(10)
|
| 186 |
-
width = 0.2
|
| 187 |
-
fig, ax = plt.subplots(figsize=(12, 5))
|
| 188 |
-
ax.bar(x - 1.5*width, per_class_acc, width, label="Accuracy")
|
| 189 |
-
ax.bar(x - 0.5*width, precision, width, label="Precision")
|
| 190 |
-
ax.bar(x + 0.5*width, recall, width, label="Recall")
|
| 191 |
-
ax.bar(x + 1.5*width, f1, width, label="F1")
|
| 192 |
-
ax.set_xticks(x)
|
| 193 |
-
ax.set_xticklabels([f"Digit {c}" for c in classes], rotation=45)
|
| 194 |
-
ax.set_ylabel("Score")
|
| 195 |
-
ax.set_title("Per-Class Metrics — Multiclass SVM (Scratch)")
|
| 196 |
-
ax.set_ylim(0, 1.05)
|
| 197 |
-
ax.legend()
|
| 198 |
-
ax.grid(axis="y", linestyle="--", alpha=0.5)
|
| 199 |
-
plt.tight_layout()
|
| 200 |
-
plt.show()
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
def test_single_digit(model, X_test, y_test, digit=7):
|
| 204 |
-
"""Find one test image of `digit`, run it through the model, print prediction."""
|
| 205 |
-
idx = np.where(y_test == digit)[0][0]
|
| 206 |
-
X_sample = X_test[idx]
|
| 207 |
-
predicted = model.predict(X_sample.reshape(1, -1))[0]
|
| 208 |
-
print(f"Input: {digit} → Predicted: {predicted} ({'CORRECT' if predicted == digit else 'WRONG'})")
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
# ---------------------------------------------------------------------------
|
| 212 |
-
# Main
|
| 213 |
-
# ---------------------------------------------------------------------------
|
| 214 |
-
|
| 215 |
-
if __name__ == "__main__":
|
| 216 |
-
|
| 217 |
-
# 1. Load data
|
| 218 |
-
X_train, X_val, X_test, y_train, y_val, y_test, X_dev, y_dev, scaler = load_data()
|
| 219 |
-
|
| 220 |
-
# 2. Hyperparameter tuning
|
| 221 |
-
best_C, best_lr = manual_grid_search(X_dev, y_dev, X_val, y_val)
|
| 222 |
-
|
| 223 |
-
# 3. Train final model on full training set
|
| 224 |
-
print("\nTraining final model on full training set...")
|
| 225 |
-
final_model = MultiClassSVMScratch(C=best_C, learning_rate=best_lr,
|
| 226 |
-
n_epochs=50, random_state=RANDOM_STATE)
|
| 227 |
-
final_model.fit(X_train, y_train)
|
| 228 |
-
print("Done.")
|
| 229 |
-
|
| 230 |
-
# 4. Evaluate on test set
|
| 231 |
-
y_pred = final_model.predict(X_test)
|
| 232 |
-
|
| 233 |
-
print("\n=== TEST SET RESULTS (Scratch SVM) ===")
|
| 234 |
-
print(f"Accuracy : {accuracy_score(y_test, y_pred):.4f}")
|
| 235 |
-
print(f"Precision (macro) : {precision_score(y_test, y_pred, average='macro'):.4f}")
|
| 236 |
-
print(f"Recall (macro) : {recall_score(y_test, y_pred, average='macro'):.4f}")
|
| 237 |
-
print(f"F1 (macro) : {f1_score(y_test, y_pred, average='macro'):.4f}")
|
| 238 |
-
print(f"F1 (weighted): {f1_score(y_test, y_pred, average='weighted'):.4f}")
|
| 239 |
-
print("\nClassification Report:")
|
| 240 |
-
print(classification_report(y_test, y_pred))
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
# 6. Plots
|
| 244 |
-
plot_loss_curves(final_model)
|
| 245 |
-
plot_confusion_matrix(y_test, y_pred)
|
| 246 |
-
plot_per_class_metrics(y_test, y_pred)
|
| 247 |
-
|
| 248 |
-
# Decision score bar chart for one sample (pick first test sample)
|
| 249 |
-
sample_idx = 0
|
| 250 |
-
plot_decision_scores(final_model, X_test[sample_idx], y_test[sample_idx])
|
| 251 |
-
|
| 252 |
-
# 7. Single image test — change digit= to any digit 0-9
|
| 253 |
-
test_single_digit(final_model, X_test, y_test, digit=4)
|
|
|
|
| 1 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
+
# We keep this because your Multiclass model uses the binary model inside it!
|
| 4 |
try:
|
| 5 |
from linear_svm import LinearSVMScartch
|
| 6 |
except ModuleNotFoundError:
|
| 7 |
from SVM.linear_svm import LinearSVMScartch
|
| 8 |
|
|
|
|
|
|
|
|
|
|
| 9 |
# ---------------------------------------------------------------------------
|
| 10 |
# One-vs-Rest multiclass wrapper around LinearSVMScartch
|
| 11 |
# ---------------------------------------------------------------------------
|
|
|
|
| 50 |
def predict(self, X):
|
| 51 |
# Pick the class whose binary SVM is most confident (+highest score)
|
| 52 |
scores = self.decision_scores(X)
|
| 53 |
+
return self.classes_[np.argmax(scores, axis=1)]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|