Patience چالش 5 از جلسهی 49، سخت
تابع train_with_patience(hidden, patience, max_epochs) را بنویس: دوره به دوره آموزش بده، کپی بهترین شبکه را (با اعتبارسنجی) نگه دار و وقتی بهترین دوره patience دوره قدیمی شد، بایست. best_net, best_epoch و دورهای که در آن ایستادی را برگردان. خروجی: patience 3: stopped at epoch 14, best epoch 11, validation 89% و patience 5: stopped at epoch 16, best epoch 11, validation 89%.
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, 4, 1) # train: 200 digits, 4 flipped pixels each
Xv, yv = make_digits(10, 4, 3) # validation: 100 digits for our choices
Xt, yt = make_digits(10, 4, 2) # test: 100 digits, used once at the very end
def softmax(z):
e = np.exp(z - z.max(axis=1, keepdims=True))
return e / e.sum(axis=1, keepdims=True)
def rand_matrix(rows, cols):
return np.array([[random.uniform(-0.5, 0.5) for c in range(cols)] for r in range(rows)])
def new_net(hidden):
random.seed(0)
return {"W1": rand_matrix(25, hidden), "b1": np.zeros(hidden),
"W2": rand_matrix(hidden, 10), "b2": np.zeros(10)}
def forward(net, x):
h = np.maximum(0, x @ net["W1"] + net["b1"])
return h, softmax(h @ net["W2"] + net["b2"])
def accuracy(net, Xs, ys):
h, p = forward(net, Xs)
return round(float((p.argmax(axis=1) == ys).mean()) * 100)
def train_epoch(net, Xs, ys, lr=0.5, batch=20):
order = list(range(len(Xs)))
random.shuffle(order)
for start in range(0, len(order), batch):
idx = order[start:start + batch]
xb, yb = Xs[idx], ys[idx]
h, p = forward(net, 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_with_patience(hidden, patience, max_epochs):
net = new_net(hidden)
# train epoch by epoch and keep a copy of the best net (by validation)
# stop when the best is `patience` epochs old
# return best_net, best_epoch, the epoch where you stopped
for patience in [3, 5]:
best_net, best_epoch, stopped = train_with_patience(32, patience, 30)
print(f"patience {patience}: stopped at epoch {stopped}, best epoch {best_epoch}, validation {accuracy(best_net, Xv, yv)}%")
راهنما و جواب این چالش · برگشت به درس جلسهی 49: بیشبرازش و تعمیم ۲ · چالش بعدی: Touch the Test Once
بیشتر بخوان: list comprehension در پایتون: از حلقه for تا یک خط کد