keras 複数入力でinput_4 to have 3 dimensions,のエラーが出て対処方法がわからない

カラー画像192x88 と文字列:6文字　2種類の入力で　単出力のAI学習です。
"aaaa"が表示されているのでmodel.fit()で失敗しているのですが
ValueError: Error when checking input: expected input_4 to have 3 dimensions, but got array with shape (2628, 1)
のエラーで、modelでは3次元で定義されているのにデータ2次元できたと言う事のようですが
どのように直したらいいかわかりません。
分かる方がいたら教えてください。

python
1from keras.models import Model, Sequential
2from keras.layers import Input, concatenate, Activation, Dense, Dropout, Flatten, SimpleRNN, Reshape
3from keras.preprocessing import image 
4import numpy as np
5from keras.preprocessing.image import ImageDataGenerator
6from keras.layers import Conv2D, MaxPooling2D
7import keras
8
9num_classes = 6
10photo_width = 192
11photo_height = 88
12data_size = photo_height * photo_width * 3
13in_shape = (photo_height, photo_width, 3)
14infile1  = "C:/Product/pato_ai/tensorflow/learning3/val.npz"
15infile2  = "C:/Product/pato_ai/tensorflow/learning3/train.npz"
16outfile1 = "C:/Product/pato_ai/tensorflow/learning3/model/road_model.h5"
17outfile2 = "C:/Product/pato_ai/tensorflow/learning3/model/road_weights"
18outfile3 = "C:/Product/pato_ai/tensorflow/learning3/model/road_keka.csv"
19val_image = []
20val_text  = []
21
22def main():
23    global val_image
24    data = np.load(infile1)
25    val_image = data["input_image"]
26    val_text  = data["input_text"]
27    val_label = data["input_label"]
28
29    data = np.load(infile2)
30    tst_image = data["input_image"]
31    tst_text  = data["input_text"]
32    tst_label = data["input_label"]
33
34    print(val_text)
35
36    # ラベルデータをOne-Hot形式に変換
37    val_label = keras.utils.to_categorical(val_label, num_classes)
38    tst_label = keras.utils.to_categorical(tst_label, num_classes)
39
40    model = train(val_image, val_text, val_label, tst_image, tst_text, tst_label)
41    model_eval(model, x_test, y_test)
42
43
44def train(val_image, val_text, val_label, tst_image, tst_text, tst_label):
45    
46    input_image = Input(shape=(photo_height, photo_width, 3))
47    input_text  = Input(shape=(6,1))
48    
49    # image input
50    model_image = Conv2D(32, (3, 3), activation='relu')(input_image)
51    model_image = Conv2D(32, (3, 3), activation='relu')(model_image)
52    model_image = MaxPooling2D(pool_size=(2, 2))(model_image)
53    model_image = Dropout(0.25)(model_image)
54    
55    model_image = Conv2D(32, (3, 3), activation='relu')(model_image)
56    model_image = Conv2D(32, (3, 3), activation='relu')(model_image)
57    model_image = MaxPooling2D(pool_size=(2, 2))(model_image)
58
59    model_image = Flatten()(model_image)
60    model_image = Dense(255, activation='relu')(model_image)
61    model_image = Dense(10, activation='softmax')(model_image)
62    model_image = Model(inputs=input_image, outputs=model_image)
63    
64    # text input
65    model_text = SimpleRNN(1, return_sequences=True)(input_text)
66    model_text = Dense(1, activation="linear")(model_text)
67    model_text = Flatten()(model_text)
68    model_text = Model(inputs=input_text, outputs=model_text)
69    
70    # Image and text combined
71    combined = concatenate([model_image.output, model_text.output], axis=1)
72    final = Dense(32, activation="relu")(combined)
73    final = Dense(4, activation="sigmoid")(final)
74    
75    model = Model(inputs=[model_image.input, model_text.input], outputs=final)
76    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
77    model.save(outfile1)
78
79
80    i = 0
81    while i < 1:
82        print("aaaa")
83        hist = model.fit([val_image, val_text], val_label,
84                         batch_size=32, epochs=30, verbose=1, validation_data=([tst_image, tst_text], tst_label))
85        print("bbbb")
86        model.save_weights(outfile2+str(i)+".h5")
87        print("ccccc")
88        score = model.evaluate([tst_image, tst_text], tst_label)
89        print("      (loss)=", score[0])
90        print("正解率(accuracy)=", score[1])
91        f = open(outfile3,'a')
92        f.write(str(i)+" (loss)="+str(score[0])+"正解率(accuracy)=,"+str(score[1])+'\n')
93        f.close()
94        i = i + 1
95
96
97    hist = model.fit(x_train, y_train, batch_size=32, epochs=100, verbose=1, validation_data=(x_test, y_test))
98    model.save_weights(outfile2 + ".h5")
99    return model
100
101def model_eval(model, x, y):
102    score = model.evaluate(x, y)
103    print("      (loss)=", score[0])
104    print("正解率(accuracy)=", score[1])
105
106if __name__ == "__main__":
107    main()
108    
109

runfile('C:/Product/pato_ai/tensorflow/learning3/train03.py', wdir='C:/Product/pato_ai/tensorflow/learning3')
['111018' '111018' '111018' ..., '122008' '122008' '122008']
aaaa
Traceback (most recent call last):

  File "<ipython-input-2-4062c011fabe>", line 1, in <module>
    runfile('C:/Product/pato_ai/tensorflow/learning3/train03.py', wdir='C:/Product/pato_ai/tensorflow/learning3')

  File "C:\Users\Arail\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 880, in runfile
    execfile(filename, namespace)

  File "C:\Users\Arail\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 102, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "C:/Product/pato_ai/tensorflow/learning3/train03.py", line 161, in <module>
    main()

  File "C:/Product/pato_ai/tensorflow/learning3/train03.py", line 50, in main
    model = train(val_image, val_text, val_label, tst_image, tst_text, tst_label)

  File "C:/Product/pato_ai/tensorflow/learning3/train03.py", line 138, in train
    batch_size=32, epochs=30, verbose=1, validation_data=([tst_image, tst_text], tst_label))

  File "C:\Users\Arail\Anaconda3\lib\site-packages\keras\engine\training.py", line 1593, in fit
    batch_size=batch_size)

  File "C:\Users\Arail\Anaconda3\lib\site-packages\keras\engine\training.py", line 1426, in _standardize_user_data
    exception_prefix='input')

  File "C:\Users\Arail\Anaconda3\lib\site-packages\keras\engine\training.py", line 110, in _standardize_input_data
    'with shape ' + str(data_shape))

ValueError: Error when checking input: expected input_4 to have 3 dimensions, but got array with shape (2628, 1)

行動規範の内容に同意します

回答2件

自己解決

画像サイズ64x64 の2種類による複数入力ができました。
出力は6パターンです。

python
1import tensorflow as tf
2import keras
3import numpy as np
4from PIL import Image
5import os, glob, random
6from keras.preprocessing import image
7from keras.preprocessing.image import load_img, img_to_array
8
9photo_width = 64
10photo_height2 = 64
11x1 = []    # 画像
12x2 = []    # 画像
13y  = []    # ラベル
14outfile1 = "C:/Product/pato_ai/tensorflow/learning4/val.npz"
15
16def glob_images(path1, path2, label):
17    # pathフォルダ内の画像を読む
18    files = glob.glob("C:/Product/pato_ai/tensorflow/learning4/" + path1 + "/*.jpg")
19
20    for i,f in enumerate(files):
21        img = image.load_img(f, target_size=(photo_height2, photo_width))
22        data1 = img_to_array(img) 
23        data1 = data1 / 255
24        data1 = data1.reshape(-1,photo_height2, photo_width, 3)
25        x1.append(data1)     # 画像
26
27    files = glob.glob("C:/Product/pato_ai/tensorflow/learning4/" + path2 + "/*.jpg")
28    for i,f in enumerate(files):
29        img = image.load_img(f, target_size=(photo_height2, photo_width))
30        data1 = img_to_array(img) 
31        data1 = data1 / 255
32        data1 = data1.reshape(-1,photo_height2, photo_width, 3)
33        x2.append(data1)     # 画像
34        y.append(label)    # ラベル
35
36def make_dataset():
37    global x1, x2
38    global y
39    x1 = []
40    x2 = []
41    y  = []
42    
43    glob_images("DBpic1A/0"   , "DBpic1B/0"   , 0)
44    glob_images("DBpic1A/2"   , "DBpic1B/2"   , 2)
45    glob_images("DBpic1A/3"   , "DBpic1B/3"   , 3)
46    glob_images("DBpic1A/4"   , "DBpic1B/4"   , 4)
47    glob_images("DBpic1A/5"   , "DBpic1B/5"   , 5)
48    np.savez(outfile1, x1=x1, x2=x2, y=y)
49    print("saved:" + outfile1)
50

python
1import tensorflow as tf
2import keras
3import numpy as np
4from PIL import Image
5import os, glob, random
6from keras.preprocessing.image import load_img, img_to_array
7
8photo_width = 64
9photo_height2 = 64
10x1 = []    # 画像
11x2 = []    # 画像
12y  = []    # ラベル
13outfile1 = "C:/Product/pato_ai/tensorflow/learning4/train.npz"
14
15def glob_images(path1, path2, label):
16
17    # pathフォルダ内の画像を読む
18    files = glob.glob("C:/Product/pato_ai/tensorflow/learning4/" + path1 + "/*.jpg")
19
20    for i,f in enumerate(files):
21        img = image.load_img(f, target_size=(photo_height2, photo_width))
22        data1 = img_to_array(img) 
23        data1 = data1 / 255
24        data1 = data1.reshape(-1,photo_height2, photo_width, 3)
25        x1.append(data1)     # 画像
26
27    files = glob.glob("C:/Product/pato_ai/tensorflow/learning4/" + path2 + "/*.jpg")
28    for i,f in enumerate(files):
29        img = image.load_img(f, target_size=(photo_height2, photo_width))
30        data1 = img_to_array(img) 
31        data1 = data1 / 255
32        data1 = data1.reshape(-1,photo_height2, photo_width, 3)
33        x2.append(data1)     # 画像
34        y.append(label)    # ラベル
35
36def make_dataset():
37    global x1, x2
38    global y
39    x1 = []
40    x2 = []
41    y  = []
42    
43    glob_images("DBpic2A/0", "DBpic2B/0", 0)
44    glob_images("DBpic2A/2", "DBpic2B/2", 2)
45    glob_images("DBpic2A/3", "DBpic2B/3", 3)
46    glob_images("DBpic2A/4", "DBpic2B/4", 4)
47    glob_images("DBpic2A/5", "DBpic2B/5", 5)
48    np.savez(outfile1, x1=x1, x2=x2, y=y)
49    print("saved:" + outfile1)
50    
51# データセットの作成
52make_dataset()

python
1from keras.models import Model, Sequential
2from keras.layers import Input, concatenate, Activation, Dense, Dropout, Flatten, SimpleRNN, Reshape
3from keras.preprocessing import image 
4import numpy as np
5from keras.preprocessing.image import ImageDataGenerator
6from keras.layers import Conv2D, MaxPooling2D
7import keras
8from keras.layers import Add
9from keras.layers import Merge
10
11num_classes = 6
12photo_width = 64
13photo_height2 = 64
14infile1  = "C:/Product/pato_ai/tensorflow/learning4/val.npz"
15infile2  = "C:/Product/pato_ai/tensorflow/learning4/train.npz"
16outfile1 = "C:/Product/pato_ai/tensorflow/learning4/model/road_model.h5"
17outfile2 = "C:/Product/pato_ai/tensorflow/learning4/model/road_weights"
18outfile3 = "C:/Product/pato_ai/tensorflow/learning4/model/road_keka.csv"
19val_x1 = []
20val_x2 = []
21
22def main():
23    global val_image
24    data = np.load(infile1)
25    val_x1 = data["x1"]
26    val_x2 = data["x2"]
27    val_y  = data["y"]
28
29    data = np.load(infile2)
30    tst_x1 = data["x1"]
31    tst_x2 = data["x2"]
32    tst_y  = data["y"]
33
34
35    # ラベルデータをOne-Hot形式に変換
36    val_y = keras.utils.to_categorical(val_y, num_classes)
37    tst_y = keras.utils.to_categorical(tst_y, num_classes)
38    val_x1 = val_x1.reshape(-1, photo_height2, photo_width,3)
39    val_x2 = val_x2.reshape(-1, photo_height2, photo_width,3)
40    tst_x1 = tst_x1.reshape(-1, photo_height2, photo_width,3)
41    tst_x2 = tst_x2.reshape(-1, photo_height2, photo_width,3)
42    model = train(val_x1, val_x2, val_y, tst_x1, tst_x2, tst_y)
43
44
45def train(val_x1, val_x2, val_y, tst_x1, tst_x2, tst_y):
46    
47    left_branch = Sequential()
48    left_branch.add(Dense(units = 64, input_shape =(photo_height2, photo_width, 3)))
49    
50    right_branch = Sequential()
51    right_branch.add(Dense(units = 64, input_shape =(photo_height2, photo_width, 3)))
52   
53    merged = Merge([left_branch, right_branch], mode='concat')
54    final_model = Sequential()
55    final_model.add(merged)
56    final_model.add(Flatten())
57    final_model.add(Dense(6, activation='softmax'))
58    
59    final_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
60    final_model.summary()
61
62    i = 0
63    while i < 3:
64        final_model.fit([val_x1, val_x2], val_y, batch_size=32, epochs=30, validation_split=0.1,)  
65        final_model.save_weights(outfile2+str(i)+".h5")
66        score = final_model.evaluate([tst_x1, tst_x2], tst_y)
67        print("      (loss)=", score[0])
68        print("正解率(accuracy)=", score[1])
69        f = open(outfile3,'a')
70        f.write(str(i)+" (loss)="+str(score[0])+"正解率(accuracy)=,"+str(score[1])+'\n')
71        f.close()
72        i = i + 1
73    return final_model
74
75if __name__ == "__main__":
76    main()
77

投稿2021/04/21 01:05

yk3125

総合スコア91

合ってるか分かりませんが,質問者様が仰っている通り,期待される入力の次元が3次元であるのに対し,実際に入力された値は2次元となっていると思われます.
入力が3次元の可能性として考えられるのが画像入力時だと思われるので,入力する値が2次元になっているのではないでしょうか.

python
1print(val_image.shape)
2print(tst_image.shape)

投稿2021/04/09 03:37

HelloQ

総合スコア81

yk3125

2021/04/09 13:16

model = の前にprint()を追加してみた print(val_image.shape) print(tst_image.shape) print(val_text.shape) print(tst_text.shape) model = train(val_image, val_text, val_label, tst_image, tst_text, tst_label) 結果 (2628, 88, 192, 3) (2160, 88, 192, 3) (2628,) (2160,) と表示されました

行動規範の内容に同意します

あなたの回答