Input 0 of layer sequential is incompatible with the layerを解決したい

model.fitした際に、Input 0 of layer sequential is incompatible with the layer: expected axis -1 of input shape to have value 64 but received input with shape [None, 64, 64, 3]と表示され、正常にmodel.fit出来ない。

エラー文

ValueError: Input 0 of layer sequential is incompatible with the layer: expected axis -1 of input shape to have value 64 but received input with shape [None, 64, 64, 3]

Python
1import os, cv2, random
2import numpy as np
3import pandas as pd
4
5import matplotlib.pyplot as plt
6from matplotlib import ticker
7import seaborn as sns
8%matplotlib inline 
9
10import tensorflow
11from tensorflow.keras.models import Sequential
12from tensorflow.keras.layers import Input, Dropout, Flatten, Convolution2D, MaxPooling2D, Dense, Activation
13from tensorflow.keras.optimizers import RMSprop
14from tensorflow.keras.callbacks import ModelCheckpoint, Callback, EarlyStopping
15import np_utils
16
17TRAIN_DIR = './archive/casting_data/casting_data/train/'
18TEST_DIR = './archive/casting_data/casting_data/test/'
19
20
21ROWS = 64
22COLS = 64
23CHANNELS = 3
24
25train_images = [TRAIN_DIR+i for i in os.listdir(TRAIN_DIR)] # use this for full dataset
26train_ok =   [TRAIN_DIR+i for i in os.listdir(TRAIN_DIR) if 'cast_ok' in i]
27train_def =   [TRAIN_DIR+i for i in os.listdir(TRAIN_DIR) if 'cast_def' in i]
28
29test_images =  [TEST_DIR+i for i in os.listdir(TEST_DIR)]
30
31# slice datasets for memory efficiency on Kaggle Kernels, delete if using full dataset
32train_images = train_ok + train_def
33random.shuffle(train_images)
34test_images =  test_images
35
36def read_image(file_path):
37    img = cv2.imread(file_path, cv2.IMREAD_COLOR) #cv2.IMREAD_GRAYSCALE
38    return cv2.resize(img, (ROWS, COLS), interpolation=cv2.INTER_CUBIC)
39
40
41def prep_data(images):
42    count = len(images)
43    data = np.ndarray((count, CHANNELS, ROWS, COLS), dtype=np.uint8)
44
45    for i, image_file in enumerate(images):
46        image = read_image(image_file)
47        data[i] = image.T #transpose mxn > nxm
48        if i%250 == 0: print('Processed {} of {}'.format(i, count))
49    
50    return data
51
52train = prep_data(train_images)
53test = prep_data(test_images)
54
55print("Train shape: {}".format(train.shape))
56print("Test shape: {}".format(test.shape))
57
58labels = []
59for i in train_images:
60    if 'cast_ok_' in i:
61        labels.append(1)
62    else:
63        labels.append(0)
64
65sns.countplot(labels)
66plt.title('OK and def')
67
68labels = np.array(labels)
69
70def show_OK_and_def(idx):
71    OK = read_image(train_ok[idx])
72    DEF = read_image(train_def[idx])
73    pair = np.concatenate((DEF, OK), axis=1)
74    plt.figure(figsize=(10,5))
75    plt.imshow(pair)
76    plt.show()
77
78DEF_avg = np.array([DEF[0].T for i, DEF in enumerate(train) if labels[i]==1]).mean(axis=0)
79plt.imshow(DEF_avg)
80plt.title('Average DEF')
81
82OK_avg = np.array([OK[0].T for i, OK in enumerate(train) if labels[i]==0]).mean(axis=0)
83plt.imshow(OK_avg)
84plt.title('Average OK')
85
86optimizer = RMSprop(lr=1e-4)
87objective = 'binary_crossentropy'
88
89
90def OKDEF():
91    
92    model = Sequential()
93
94    model.add(Convolution2D(32, (3, 3), input_shape=(3, ROWS, COLS), activation='relu', padding="same"))
95    model.add(Convolution2D(32, (3, 3), activation='relu', padding="same"))
96    model.add(MaxPooling2D(data_format="channels_first", pool_size=(2, 2)))
97    
98    model.add(Convolution2D(64, (3, 3), activation='relu', padding="same"))
99    model.add(Convolution2D(64, (3, 3), activation='relu', padding="same"))
100    model.add(MaxPooling2D(data_format="channels_first", pool_size=(2, 2)))
101    
102    model.add(Convolution2D(128, (3, 3), activation='relu', padding="same"))
103    model.add(Convolution2D(128, (3, 3), activation='relu', padding="same"))
104    model.add(MaxPooling2D(data_format="channels_first", pool_size=(2, 2)))
105    
106    model.add(Convolution2D(256, (3, 3), activation='relu', padding="same"))
107    model.add(Convolution2D(256, (3, 3), activation='relu', padding="same"))
108    model.add(MaxPooling2D(data_format="channels_first", pool_size=(2, 2)))
109    
110    model.add(Flatten())
111    model.add(Dense(256, activation='relu'))
112    model.add(Dropout(0.5))
113    
114    model.add(Dense(256, activation='relu'))
115    model.add(Dropout(0.5))
116
117    model.add(Dense(1))
118    model.add(Activation('sigmoid'))
119
120    model.compile(loss=objective, optimizer=optimizer, metrics=['accuracy'])
121    return model
122
123
124model = OKDEF()
125
126train = tensorflow.transpose(train, [0, 2, 3, 1])
127test = tensorflow.transpose(test,[0, 2, 3, 1])
128
129epochs = 10
130batch_size = 16
131
132class LossHistory(Callback):
133    def on_train_begin(self, logs={}):
134        self.losses = []
135        self.val_losses = []
136        
137    def on_epoch_end(self, batch, logs={}):
138        self.losses.append(logs.get('loss'))
139        self.val_losses.append(logs.get('val_loss'))
140
141early_stopping = EarlyStopping(monitor='val_loss', patience=3, verbose=1, mode='auto')        
142        
143def run_OKDEF():
144    
145    history = LossHistory()
146    model.fit(train, labels, batch_size=batch_size, epochs=epochs,
147              validation_split=0.25, verbose=0, shuffle=True, callbacks=[history, early_stopping])
148    
149    predictions = model.predict(test, verbose=0)
150    return predictions, history
151
152predictions, history = run_OKDEF()