Upload perceptron/perceptron.py with huggingface_hub
Browse files- perceptron/perceptron.py +131 -0
perceptron/perceptron.py
ADDED
|
@@ -0,0 +1,131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Perceptron Learning Algorithm — CS401
|
| 3 |
+
From scratch, no ML libraries.
|
| 4 |
+
Matches the lecture: linear neuron with step activation, weights update only on misclassification.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
# ---- Training Data ----
|
| 8 |
+
# (Hours studied, Hours slept) → Passed (1) / Failed (0)
|
| 9 |
+
DATA = [
|
| 10 |
+
(1.0, 5.0, 0),
|
| 11 |
+
(1.5, 6.0, 0),
|
| 12 |
+
(2.0, 5.0, 0),
|
| 13 |
+
(2.5, 4.0, 0),
|
| 14 |
+
(3.0, 5.0, 0),
|
| 15 |
+
(4.0, 5.0, 1),
|
| 16 |
+
(5.0, 6.0, 1),
|
| 17 |
+
(6.0, 5.0, 1),
|
| 18 |
+
(6.5, 7.0, 1),
|
| 19 |
+
(7.0, 6.0, 1),
|
| 20 |
+
]
|
| 21 |
+
|
| 22 |
+
n = len(DATA)
|
| 23 |
+
POINTS = [(d[0], d[1]) for d in DATA]
|
| 24 |
+
LABELS = [d[2] for d in DATA]
|
| 25 |
+
|
| 26 |
+
# Standardise features so the learning rate behaves consistently
|
| 27 |
+
means = [sum(p[i] for p in POINTS) / n for i in range(2)]
|
| 28 |
+
stds = [(sum((p[i] - means[i]) ** 2 for p in POINTS) / n) ** 0.5 for i in range(2)]
|
| 29 |
+
POINTS_S = [((p[0] - means[0]) / stds[0], (p[1] - means[1]) / stds[1]) for p in POINTS]
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# ---- Neuron: step activation ----
|
| 33 |
+
def neuron(x, w, b):
|
| 34 |
+
"""output = 1 if w.x + b >= 0 else 0"""
|
| 35 |
+
z = w[0] * x[0] + w[1] * x[1] + b
|
| 36 |
+
return 1 if z >= 0 else 0
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
# ---- Perceptron Learning Rule ----
|
| 40 |
+
def train_perceptron(lr=0.1, epochs=50):
|
| 41 |
+
w = [0.0, 0.0]
|
| 42 |
+
b = 0.0
|
| 43 |
+
|
| 44 |
+
for epoch in range(epochs):
|
| 45 |
+
errors = 0
|
| 46 |
+
for x, y in zip(POINTS_S, LABELS):
|
| 47 |
+
y_pred = neuron(x, w, b)
|
| 48 |
+
if y_pred != y: # update ONLY on mistake
|
| 49 |
+
errors += 1
|
| 50 |
+
w[0] += lr * (y - y_pred) * x[0]
|
| 51 |
+
w[1] += lr * (y - y_pred) * x[1]
|
| 52 |
+
b += lr * (y - y_pred)
|
| 53 |
+
if epoch % 10 == 0 or errors == 0:
|
| 54 |
+
print(f"Epoch {epoch:3d} | errors: {errors} | w: [{w[0]:+.4f}, {w[1]:+.4f}] | b: {b:+.4f}")
|
| 55 |
+
if errors == 0:
|
| 56 |
+
print(f"Converged at epoch {epoch} — zero training errors")
|
| 57 |
+
break
|
| 58 |
+
|
| 59 |
+
return w, b
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
# ---- Logic gates (lecture: "Neuron as Logic Gate") ----
|
| 63 |
+
def train_gate(name, inputs, targets, lr=1.0, epochs=20):
|
| 64 |
+
w = [0.0, 0.0]
|
| 65 |
+
b = 0.0
|
| 66 |
+
errors = epochs
|
| 67 |
+
for _ in range(epochs):
|
| 68 |
+
errors = 0
|
| 69 |
+
for x, y in zip(inputs, targets):
|
| 70 |
+
z = 1 if w[0] * x[0] + w[1] * x[1] + b >= 0 else 0
|
| 71 |
+
if z != y:
|
| 72 |
+
errors += 1
|
| 73 |
+
w[0] += lr * (y - z) * x[0]
|
| 74 |
+
w[1] += lr * (y - z) * x[1]
|
| 75 |
+
b += lr * (y - z)
|
| 76 |
+
if errors == 0:
|
| 77 |
+
break
|
| 78 |
+
hits = sum(
|
| 79 |
+
(1 if w[0] * x1 + w[1] * x2 + b >= 0 else 0) == y
|
| 80 |
+
for (x1, x2), y in zip(inputs, targets)
|
| 81 |
+
)
|
| 82 |
+
print(f"{name}: w = [{w[0]:+.1f}, {w[1]:+.1f}], b = {b:+.1f} → "
|
| 83 |
+
f"{hits}/{len(targets)} correct")
|
| 84 |
+
return errors
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
# ---- Main ----
|
| 88 |
+
if __name__ == "__main__":
|
| 89 |
+
print("=" * 60)
|
| 90 |
+
print("Perceptron Learning Algorithm")
|
| 91 |
+
print("=" * 60)
|
| 92 |
+
print(f"Training data: {n} examples")
|
| 93 |
+
print()
|
| 94 |
+
|
| 95 |
+
w, b = train_perceptron(lr=0.1, epochs=50)
|
| 96 |
+
|
| 97 |
+
print()
|
| 98 |
+
print("=" * 60)
|
| 99 |
+
print(f"Result: h(x) = step({w[0]:.4f}*x1 + {w[1]:.4f}*x2 + {b:.4f})")
|
| 100 |
+
print("=" * 60)
|
| 101 |
+
|
| 102 |
+
# Predictions
|
| 103 |
+
print()
|
| 104 |
+
print(f"{'Hours':>6} | {'Slept':>6} | {'Actual':>6} | {'z = w.x+b':>9} | {'Predicted':>9}")
|
| 105 |
+
print("-" * 50)
|
| 106 |
+
correct = 0
|
| 107 |
+
for xi, yi in zip(POINTS, LABELS):
|
| 108 |
+
xs = ((xi[0] - means[0]) / stds[0], (xi[1] - means[1]) / stds[1])
|
| 109 |
+
z = w[0] * xs[0] + w[1] * xs[1] + b
|
| 110 |
+
y_pred = neuron(xs, w, b)
|
| 111 |
+
correct += y_pred == yi
|
| 112 |
+
print(f"{xi[0]:>6} | {xi[1]:>6} | {yi:>6} | {z:>+9.3f} | {y_pred:>9}")
|
| 113 |
+
|
| 114 |
+
print()
|
| 115 |
+
print(f"Training accuracy: {correct}/{n} = {correct / n:.0%}")
|
| 116 |
+
|
| 117 |
+
# Predict new value
|
| 118 |
+
print()
|
| 119 |
+
new_point = (4.5, 6.0)
|
| 120 |
+
xs = ((new_point[0] - means[0]) / stds[0], (new_point[1] - means[1]) / stds[1])
|
| 121 |
+
print(f"New prediction: {new_point[0]} hours studied, {new_point[1]} hours slept "
|
| 122 |
+
f"→ {'PASS' if neuron(xs, w, b) else 'FAIL'}")
|
| 123 |
+
|
| 124 |
+
# Logic gates — can a single neuron learn them?
|
| 125 |
+
print()
|
| 126 |
+
print("Logic gates (single neuron):")
|
| 127 |
+
train_gate("AND", [(0, 0), (0, 1), (1, 0), (1, 1)], [0, 0, 0, 1])
|
| 128 |
+
train_gate("OR ", [(0, 0), (0, 1), (1, 0), (1, 1)], [0, 1, 1, 1])
|
| 129 |
+
xor_errors = train_gate("XOR", [(0, 0), (0, 1), (1, 0), (1, 1)], [0, 1, 1, 0])
|
| 130 |
+
if xor_errors:
|
| 131 |
+
print("XOR: NOT linearly separable — a single neuron can never learn it (→ need hidden layers)")
|