値がおかしい？MNISTデータセットを多層パーセプトロン(MLP)で学習

###質問内容
結果の値がおかしいと思われるが、コードのどの部分がおかしいかわからない。

###条件
homework関数内のみで作成
MLPの作成にTensorflowなどのライブラリを使わない
出力yはone-of-k表現
最終層の活性化関数はソフトマックス関数, 誤差関数は多クラス交差エントロピー
###コード

lang
1def homework(train_X, train_y, test_X):
2    def sigmoid(x):
3        return 1/(1+np.exp(-x))
4    def deriv_sigmoid(x):
5        return sigmoid(x) *(1-sigmoid(x))
6    def softmax(x):
7        exp_x = np.exp(x)
8        return exp_x/np.sum(exp_x, keepdims=True)
9    def deriv_softmax(x):
10        return softmax(x)*(1 - softmax(x))
11
12    np.random.seed(0)
13    # Layer1 weights
14    W1 = np.random.uniform(low=-0.08, high=0.08, size=(train_X.shape[1],100)).astype('float32')
15    b1 = np.zeros(100).astype('float32')
16
17    # Layer2 weights
18    W2 = np.random.uniform(low=-0.08, high=0.08, size=(100, 10)).astype('float32')
19    b2 = np.zeros(10).astype('float32')
20    eps=0.1
21    epochs=10
22    for loop in range(epochs):
23        train_X, valid_X, train_y, valid_y = train_test_split(train_X, train_y,
24                                                              test_size=0.2,
25                                                              random_state=42)
26
27    
28        for X,Y in zip(train_X,train_y):
29            x=X[np.newaxis,:]
30            
31            t=np.zeros(10)
32            t[Y]=1
33            
34            u1 = np.matmul(x, W1) + b1
35            z= sigmoid(u1)
36    
37
38            u2 = np.matmul(z, W2) + b2
39            y = softmax(u2)
40    
41
42            
43            delta_2 = y - t 
44            delta_1 = deriv_sigmoid(u1) * np.matmul(delta_2, W2.T) 
45      
46        
47            W1 = W1 - eps*np.matmul(x.T, delta_1)
48            b1 = b1 - eps*np.matmul(np.ones(len(x)), delta_1)
49    
50
51            W2 = W2 - eps*np.matmul(z.T, delta_2)
52            b2 = b2 - eps*np.matmul(np.ones(len(z)), delta_2)
53
54    pred_y=np.array([])
55    for X in test_X:
56        x=X[np.newaxis,:]
57        
58        u1=np.matmul(x,W1)+b1
59        z=sigmoid(u1)
60        
61        u2=np.matmul(z,W2)+b2
62        y=softmax(u2)
63        pred_y=np.append(pred_y,np.argmax(y))
64    return pred_y
65
66from sklearn.utils import shuffle
67from sklearn.metrics import f1_score
68from sklearn.datasets import fetch_mldata
69from sklearn.model_selection import train_test_split
70
71import numpy as np
72
73def load_mnist():
74    mnist = fetch_mldata('MNIST original')
75    mnist_X, mnist_y = shuffle(mnist.data.astype('float32'),
76                               mnist.target.astype('int32'), random_state=42)
77
78    mnist_X = mnist_X / 255.0
79
80    return train_test_split(mnist_X, mnist_y,
81                test_size=0.2,
82                random_state=42)
83
84def validate_homework():
85    train_X, test_X, train_y, test_y = load_mnist()
86
87    # validate for small dataset
88    train_X_mini = train_X[:100]
89    train_y_mini = train_y[:100]
90    test_X_mini = test_X[:100]
91    test_y_mini = test_y[:100]
92
93    pred_y = homework(train_X_mini, train_y_mini, test_X_mini)
94    print(f1_score(test_y_mini, pred_y, average='macro'))
95
96
97validate_homework()
98