def auto_diff_lr():
# 注意，以下实现某些情况会有很大的数值误差，
# 所以一般真实系统实现会提供高阶算子，从而减少数值误差
N = 100
X_val, Y_val = gen_2d_data(N)
w_val = np.ones(3)
w = Tensor(w_val)
plot(N, X_val, Y_val, w_val)
test_accuracy(w_val, X_val, Y_val)
alpha = 0.01
max_iters = 300
for iteration in range(max_iters):
acc_L_val = 0
for i in range(N):
w.zero_gard()
x_val = X_val[i]
y_val = np.array(Y_val[i])
x = Tensor(x_val)
y = Tensor(y_val)
h = 1 / (1 + (-(w * x).sum()).exp())
l = y * h.log() + (1 - y) * (1 - h).log()
l.backward()
w.data += alpha * w.grad
acc_L_val += l.data
print("iter = %d, likelihood = %s, w = %s" %
(iteration, acc_L_val, w_val))
test_accuracy(w_val, X_val, Y_val)
plot(N, X_val, Y_val, w_val, True)
def test_reduce_sum_mix():
x1_val = 2 * np.ones(3)
x1 = Tensor(x1_val)
y = x1.sum().exp()
expected_y_val = np.exp(np.sum(x1_val))
assert np.array_equal(y.data, expected_y_val)
y.backward()
assert np.array_equal(x1.grad, expected_y_val * np.ones_like(x1_val))
x1.zero_gard()
y2 = x1.sum().log()
expected_y2_val = np.log(np.sum(x1_val))
assert np.array_equal(y2.data, expected_y2_val)
y2.backward()
assert np.array_equal(x1.grad, (1 / np.sum(x1_val)) * np.ones_like(x1_val))