【Python】KerasとCNNで画像分類したいが「ValueError: Error when checking target: expected dense

問題

Python3.7とKeras2.3.1で3種類の画像の分類器を作成中です。
class0, class1, class2の3つのフォルダに、それぞれ画像を入れてCNNで分類器を作成したいのですが、以下のプログラムを実行するとエラーが出てしまします。
解決法をご教示ください。

##　エラー

python
1Model: "sequential_1"
2_________________________________________________________________
3Layer (type)                 Output Shape              Param #   
4=================================================================
5conv2d_1 (Conv2D)            (None, 32, 32, 32)        896       
6_________________________________________________________________
7conv2d_2 (Conv2D)            (None, 32, 32, 32)        9248      
8_________________________________________________________________
9max_pooling2d_1 (MaxPooling2 (None, 16, 16, 32)        0         
10_________________________________________________________________
11dropout_1 (Dropout)          (None, 16, 16, 32)        0         
12_________________________________________________________________
13conv2d_3 (Conv2D)            (None, 16, 16, 64)        18496     
14_________________________________________________________________
15conv2d_4 (Conv2D)            (None, 16, 16, 64)        36928     
16_________________________________________________________________
17max_pooling2d_2 (MaxPooling2 (None, 8, 8, 64)          0         
18_________________________________________________________________
19dropout_2 (Dropout)          (None, 8, 8, 64)          0         
20_________________________________________________________________
21flatten_1 (Flatten)          (None, 4096)              0         
22_________________________________________________________________
23dense_1 (Dense)              (None, 512)               2097664   
24_________________________________________________________________
25dropout_3 (Dropout)          (None, 512)               0         
26_________________________________________________________________
27dense_2 (Dense)              (None, 3)                 1539      
28=================================================================
29Total params: 2,164,771
30Trainable params: 2,164,771
31Non-trainable params: 0
32_________________________________________________________________
33Traceback (most recent call last):
34  File "/Users/python/test2/test1.py", line 128, in <module>
35    main()
36  File "/Users/python/test2/test1.py", line 113, in main
37    validation_split=0.1)
38  File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/keras/engine/training.py", line 1154, in fit
39    batch_size=batch_size)
40  File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/keras/engine/training.py", line 621, in _standardize_user_data
41    exception_prefix='target')
42  File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/keras/engine/training_utils.py", line 135, in standardize_input_data
43    'with shape ' + str(data_shape))
44ValueError: Error when checking target: expected dense_2 to have 2 dimensions, but got array with shape (12, 3, 3)

ソースコード

python
1from keras.layers import Dense, Dropout, Flatten
2from keras.layers import Conv2D, MaxPooling2D
3from keras.models import Sequential, model_from_json
4from keras.optimizers import RMSprop
5from keras.preprocessing.image import array_to_img, img_to_array, load_img
6from keras.utils import np_utils, to_categorical
7from sklearn.model_selection import train_test_split
8import matplotlib.pyplot as plt
9import numpy as np
10import re, os
11 
12def list_pictures(directory, ext='jpg|jpeg|bmp|png|ppm'):
13    return [os.path.join(root, f) for root, _, files in os.walk(directory) for f in files if re.match(r'([\w]+.(?:' + ext + '))', f.lower())]
14
15def main():
16    batch_size = 5 # バッチサイズ
17    num_classes = 3 # 0, 1, 2の3分類
18    epochs = 20      
19    dropout_rate = 0.2
20
21    img_width = 32 # 入力画像の幅
22    img_height = 32 # 入力画像の高さ
23    img_ch = 3 # 3ch画像で学習
24
25    # データ格納用のディレクトリパス
26    dirpath= "/Users/python/test2/data/"
27
28    data_x = []
29    data_y = []
30    num_classes = 3
31
32    # クラス0, 1, 2の画像をロード
33    for filepath in list_pictures(dirpath + "class0"):
34        img = img_to_array(load_img(filepath, target_size=(img_width,img_height, img_ch)))
35        data_x.append(img)
36        data_y.append(0) 
37
38    for filepath in list_pictures(dirpath + "class1"):
39        img = img_to_array(load_img(filepath, target_size=(img_width,img_height, img_ch)))
40        data_x.append(img)
41        data_y.append(1)
42
43    for filepath in list_pictures(dirpath + "class2"):
44        img = img_to_array(load_img(filepath, target_size=(img_width,img_height, img_ch)))
45        data_x.append(img)
46        data_y.append(2)
47
48    # NumPy配列に変換
49    data_x = np.asarray(data_x)
50    data_y = np.asarray(data_y)
51
52    # 学習データとテストデータに分割
53    x_train, x_test, y_train, y_test = train_test_split(data_x, data_y, test_size=0.2)
54
55    # float32型に変換し、正規化(0～1)
56    x_train = x_train.astype('float32')
57    x_test = x_train.astype('float32')
58    x_train = x_train / 255.0
59    x_test = x_train / 255.0
60
61    # one hotエンコーディング
62    y_train = to_categorical(y_train, num_classes)
63    y_train = to_categorical(y_train, num_classes)
64
65
66    # モデルの構築
67    model = Sequential()
68
69    model.add(Conv2D(32,(3,3), 
70                padding='same', 
71                input_shape=x_train.shape[1:],
72                activation='relu'))
73
74    model.add(Conv2D(32,(3,3),
75                padding='same',
76                activation='relu'))
77
78    model.add(MaxPooling2D(pool_size=(2, 2)))
79
80    model.add(Dropout(dropout_rate))
81
82    model.add(Conv2D(64,(3,3),
83                padding='same',
84                activation='relu'))
85
86    model.add(Conv2D(64,(3,3),
87                padding='same',
88                activation='relu'))
89
90    model.add(MaxPooling2D(pool_size=(2, 2)))
91
92    model.add(Dropout(dropout_rate))
93
94    model.add(Flatten())
95
96    model.add(Dense(512, activation='relu'))
97
98    model.add(Dropout(dropout_rate))
99    
100    model.add(Dense(num_classes, activation='softmax')) # 活性化関数：softmax
101
102    model.summary()
103
104    model.compile(loss='categorical_crossentropy', optimizer=RMSprop(), metrics=['accuracy'])
105
106    # 学習
107    history = model.fit(x_train, 
108                        y_train, 
109                        batch_size=batch_size, 
110                        epochs=epochs, 
111                        verbose=1, 
112                        validation_split=0.1)
113
114    # スコアを計算
115    score = model.evaluate(x_test, 
116                            y_test,
117                            verbose=0
118                            )
119
120    # 損失
121    print('Test loss:', score[0])
122
123    # 精度
124    print('Test accuracy:', score[1])
125    
126if __name__ == '__main__':
127    main()