جواب «Step 4: Training with Mini-batches» چالش 4 از جلسهی 47
سؤال
قدم ۴: train_step (یک قدم پسانتشار) آماده است. داخل حلقهی train را کامل کن: اول order را بُر بزن، بعد برای هر دستهی ۲۰تایی train_step را صدا بزن. بعد از خروجی قدم ۳، ده خط میآید؛ اولی epoch 1 train 46%، پنجمی epoch 5 train 96% و آخری epoch 10 train 99%.
راهنما
random.shuffle(order) و بعد for start in range(0, len(order), batch): با idx = order[start:start + batch].
یک جواب درست
import random
import numpy as np
SHAPES = '''
.###. ..#.. .###. ####. #..#. ##### .###. ##### .###. .###.
#...# .##.. #...# ....# #..#. #.... #.... ....# #...# #...#
#...# ..#.. ..##. .###. ##### ####. ####. ...#. .###. .####
#...# ..#.. .#... ....# ...#. ....# #...# ..#.. #...# ....#
.###. .###. ##### ####. ...#. ####. .###. ..#.. .###. .###.
'''
def make_digits(copies, noise, seed):
random.seed(seed)
rows = SHAPES.strip().splitlines()
X, y = [], []
for c in range(copies):
for d in range(10):
pixels = [1.0 if ch == "#" else 0.0 for ch in "".join(r.split()[d] for r in rows)]
for i in random.sample(range(25), noise):
pixels[i] = 1.0 - pixels[i]
X.append(pixels)
y.append(d)
return np.array(X), np.array(y)
X, y = make_digits(20, 2, 1) # 200 training digits
Xt, yt = make_digits(10, 2, 2) # 100 test digits
def show(pixels):
for r in range(5):
row = pixels[r * 5:r * 5 + 5]
print("".join("#" if p > 0.5 else "." for p in row))
print(X.shape, Xt.shape)
print("label:", int(y[3]))
show(X[3])
random.seed(0)
def rand_matrix(rows, cols):
return np.array([[random.uniform(-0.5, 0.5) for c in range(cols)] for r in range(rows)])
net = {"W1": rand_matrix(25, 16), "b1": np.zeros(16),
"W2": rand_matrix(16, 10), "b2": np.zeros(10)}
def softmax(z):
e = np.exp(z - z.max(axis=1, keepdims=True))
return e / e.sum(axis=1, keepdims=True)
def forward(x):
h = np.maximum(0, x @ net["W1"] + net["b1"])
p = softmax(h @ net["W2"] + net["b2"])
return h, p
h, p = forward(X[:3])
print(p.shape)
print(np.round(p[0], 2))
print(round(float(p[0].sum()), 3))
def accuracy(Xs, ys):
h, p = forward(Xs)
guesses = p.argmax(axis=1)
return float((guesses == ys).mean())
print(f"before training: train {round(accuracy(X, y) * 100)}% test {round(accuracy(Xt, yt) * 100)}%")
def train_step(xb, yb, lr):
h, p = forward(xb)
dz2 = (p - np.eye(10)[yb]) / len(xb)
dh = dz2 @ net["W2"].T * (h > 0)
net["W2"] -= lr * (h.T @ dz2)
net["b2"] -= lr * dz2.sum(axis=0)
net["W1"] -= lr * (xb.T @ dh)
net["b1"] -= lr * dh.sum(axis=0)
def train(epochs, batch, lr):
order = list(range(len(X)))
for epoch in range(1, epochs + 1):
random.shuffle(order)
for start in range(0, len(order), batch):
idx = order[start:start + batch]
train_step(X[idx], y[idx], lr)
print(f"epoch {epoch} train {round(accuracy(X, y) * 100)}%")
train(10, 20, 0.5)
اول خودت امتحان کن؛ جوابهای دیگر هم قبول است. برگرد و خودت حلش کن