Python3 4次元のデータセットが3次元として認識されてしまう

前提・実現したいこと

kerasのImageDataGeneratorを使って各サンプルごとの平均を0にする処理を行いたいのですが、
以下の通り、データセットの次元が4次元であるべきところ、3次元であるというエラーが発生してしまいます。
しかし、ndimを使ってデータセットの次元を確認すると、ちゃんと4次元になっています。
1次元（サンプルの数）の部分をうまく認識してくれていないようなのですが、このような場合、どうしたら4次元として読み込んでもらえるか、アドバイスをいただけないでしょうか。

発生している問題・エラーメッセージ

ValueError                                Traceback (most recent call last)
<ipython-input-19-2f1bdd86e710> in <module>()
    186 #各サンプルの平均を0にする
    187 train_datagen = ImageDataGenerator(samplewise_center=True)
--> 188 train_generator = train_datagen.flow(x_train, None, batch_size=32)
    189 
    190 validation_datagen = ImageDataGenerator(samplewise_center=True)

ValueError: ('Input data in `NumpyArrayIterator` should have rank 4. You passed an array with shape', (512, 496, 1))

該当のソースコード

Python
1from __future__ import absolute_import
2from __future__ import division
3from __future__ import print_function
4
5from keras.layers import Lambda, Input, Dense
6from keras.models import Model
7from keras.models import Sequential, model_from_json
8from keras.losses import mse, binary_crossentropy
9from keras.layers import Conv2D, Flatten
10from keras.layers import Reshape, Conv2DTranspose
11from keras.utils import plot_model, np_utils 
12from keras.utils import plot_model
13from keras.callbacks import Callback, EarlyStopping, TensorBoard, ModelCheckpoint, LearningRateScheduler, CSVLogger
14from keras import optimizers
15from keras import backend as K
16from keras.preprocessing.image import array_to_img, img_to_array,load_img
17from keras.preprocessing.image import ImageDataGenerator
18from keras.layers import Activation, BatchNormalization
19
20import numpy as np
21import matplotlib.pyplot as plt
22import argparse
23import os
24import re
25import glob
26import random as rn
27import tensorflow as tf
28import cv2
29from PIL import Image
30
31import warnings
32warnings.filterwarnings('ignore')
33
34%matplotlib inline
35get_ipython().run_line_magic('matplotlib', 'inline')
36
37
38#reparameterization trick
39# instead of sampling from Q(z|X), sample eps = N(0,I)
40# z = z_mean + sqrt(var)*eps
41def sampling(args):
42    """Reparameterization trick by sampling fr an isotropic unit Gaussian.
43
44    # Arguments
45        args (tensor): mean and log of variance of Q(z|X)
46
47    # Returns
48        z (tensor): sampled latent vector
49    """
50    z_mean, z_log_var = args
51    batch = K.shape(z_mean)[0]
52    dim = K.int_shape(z_mean)[1]
53    # by default, random_normal has mean=0 and std=1.0
54    epsilon = K.random_normal(shape=(batch, dim))
55    return z_mean + K.exp(0.5 * z_log_var) * epsilon
56
57
58def plot_results(models,
59                 data,
60                batch_size=25,
61                 model_name="vae_OCT"):
62    """Plots labels and MNIST digits as function of 2-dim latent vector
63
64    # Arguments
65        models (tuple): encoder and decoder models
66        data (tuple): test data and label
67        batch_size (int): prediction batch size
68        model_name (string): which model is using this function
69    """
70
71    encoder, decoder = models
72    x_test = data
73    os.makedirs(model_name, exist_ok=True)
74
75    filename = os.path.join(model_name, "vae_mean.png")
76    # display a 2D plot of the digit classes in the latent space
77    z_mean, _, _ = encoder.predict(x_test,
78                                   batch_size=batch_size)
79
80#original dataset
81#train
82filenames = glob.glob("./NORMAL_resize_pixel_100_1_0506/*.jpeg")
83x_train= []
84
85for filename in filenames:
86    img = img_to_array(load_img(
87    filename, color_mode = "grayscale"
88    , target_size=(512,496)))
89    x_train.append(img)
90
91x_train = np.asarray(x_train)
92print(x_train.shape)# (100, 512, 496, 1)
93
94#test
95filenames = glob.glob("./NORMAL_resize_pixel_0506/*.jpeg")
96x_test= []
97
98for filename in filenames:
99    img = img_to_array(load_img(
100    filename, color_mode = "grayscale"
101    , target_size=(512,496)))
102    x_test.append(img)
103
104x_test = np.asarray(x_test)
105
106
107image_size_width = x_train.shape[1]
108image_size_height = x_train.shape[2]
109original_dim = 512 * 496  #3削除
110x_train = np.reshape(x_train, [-1, image_size_width,image_size_height,1])
111x_test = np.reshape(x_test,[ -1, image_size_width,image_size_height,1])
112x_train = x_train.astype('float32') / 255
113x_test = x_test.astype('float32') / 255
114
115#numpy配列をタプルに変換
116x_train = tuple(x_train)
117x_test = tuple(x_test)
118
119#各サンプルの平均を0にする
120train_datagen = ImageDataGenerator(samplewise_center=True)
121train_generator = train_datagen.flow(x_train, batch_size=32)
122
123validation_datagen = ImageDataGenerator(samplewise_center=True)
124validation_generator = validation_datagen.flow(x_test, batch_size=32)
125
126print(x_train.shape)# 表示されず（タプルに変換したため？）
127print(x_test.shape) # 表示されず（タプルに変換したため？）
128
129# network parameters
130input_shape = (image_size_width,image_size_height,1)
131batch_size = 25#50
132kernel_size = 3
133filters = 16
134latent_dim = 2
135epochs = 100
136
137# VAE model = encoder + decoder
138# build encoder model
139inputs = Input(shape=input_shape, name='encoder_input')
140x = inputs
141for i in range(4):
142    filters *= 2
143    x = Conv2D(filters=filters,
144               kernel_size=kernel_size,
145               activation = 'relu',
146               strides=2,padding='same')(x)
147    #x = BatchNormalization()(x)
148    #x = Activation('relu')(x)
149
150# shape info needed to build decoder model
151shape = K.int_shape(x)
152
153# generate latent vector Q(z|X)
154x = Flatten()(x)
155x = Dense(64, activation='relu')(x)
156z_mean = Dense(latent_dim, name='z_mean')(x)
157z_log_var = Dense(latent_dim, name='z_log_var')(x)
158
159# use reparameterization trick to push the sampling out as input
160# note that "output_shape" isn't necessary with the TensorFlow backend
161z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])
162
163# instantiate encoder model
164encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')
165encoder.summary()
166plot_model(encoder, to_file='vae_cnn_encoder.png', show_shapes=True)
167
168# build decoder model
169latent_inputs = Input(shape=(latent_dim,), name='z_sampling')
170x = Dense(shape[1] * shape[2] * shape[3], activation='relu')(latent_inputs)
171x = Reshape((shape[1], shape[2], shape[3]))(x)
172
173for i in range(4):
174    x = Conv2DTranspose(filters=filters,
175               kernel_size=kernel_size,
176               activation = 'relu',         
177               strides=2,padding='same')(x)
178    #x = BatchNormalization()(x)
179    #x = Activation('relu')(x)
180    filters //= 2
181
182outputs = Conv2DTranspose(filters=1,
183                          kernel_size=kernel_size,
184                          activation='sigmoid',
185                          padding='same',
186                          name='decoder_output')(x)
187#outputs = BatchNormalization()(outputs)#(x)
188#outputs = Activation('sigmoid')(outputs)#(x)
189
190# instantiate decoder model
191decoder = Model(latent_inputs, outputs, name='decoder')
192decoder.summary()
193plot_model(decoder, to_file='vae_cnn_decoder.png', show_shapes=True)
194
195# instantiate VAE model
196outputs = decoder(encoder(inputs)[2])
197vae = Model(inputs, outputs, name='vae')
198
199
200
201def plot_history(history):
202
203    # ??????????
204    plt.plot(history.history['loss'])
205    plt.plot(history.history['val_loss'])
206    plt.title('model loss')
207    plt.xlabel('epoch')
208    plt.ylabel('loss')
209    plt.legend(['loss', 'val_loss'], loc='lower right')
210    plt.savefig('loss.png')  # -----(2)
211    plt.show()
212
213
214
215if __name__ == '__main__':
216    parser = argparse.ArgumentParser()
217    help_ = "Load h5 model trained weights"
218    parser.add_argument("-w", "--weights", help=help_)
219    help_ = "Use mse loss instead of binary cross entropy (default)"
220    parser.add_argument("-m", "--mse", help=help_, action='store_true')
221    args = parser.parse_args([])
222    models = (encoder, decoder)
223    data = (x_test)
224
225
226
227
228    # VAE loss = mse_loss or xent_loss + kl_loss
229    if args.mse:
230        reconstruction_loss = mse(K.flatten(inputs), K.flatten(outputs))
231    else:
232        reconstruction_loss = binary_crossentropy(K.flatten(inputs),
233                                                  K.flatten(outputs))
234
235    reconstruction_loss *= image_size_width * image_size_height
236    kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
237    kl_loss = K.sum(kl_loss, axis=-1)
238    kl_loss *= -0.5
239    vae_loss = K.mean(reconstruction_loss + kl_loss)
240    vae.add_loss(vae_loss)
241    Adam = optimizers.Adam(lr=0.0005)
242    vae.compile(optimizer=Adam)
243    vae.summary()
244    plot_model(vae, to_file='vae_cnn.png', show_shapes=True)
245
246    callbacks = []
247    callbacks.append(ModelCheckpoint(filepath="model.ep{epoch:02d}.h5", save_best_only = True, period=5))
248    callbacks.append(EarlyStopping(monitor='val_loss', patience=7, verbose=1))
249    callbacks.append(CSVLogger("history.csv"))
250
251    if args.weights:
252        vae.load_weights(args.weights)
253    else:
254        # train the autoencoder
255        history = vae.fit_generator(train_generator,
256                          steps_per_epoch=x_train.shape[0] // batch_size,
257                          epochs=epochs,#20,
258                          validation_data=(validation_generator,None),
259                          validation_steps=x_test.shape[0] // batch_size,
260                          verbose=1,
261                          callbacks=callbacks)
262
263    plot_results(models, data, batch_size=batch_size, model_name="vae_OCT")
264    plot_history(history)

試したこと

以下を参考に試しに次元を増やしてみると、今度は「tuple is out of index」のエラーが出てしまいます。

https://deepage.net/features/numpy-newaxis.html

http://www.kamishima.net/mlmpyja/nbayes2/shape.html

補足情報（FW/ツールのバージョンなど）

google colabを使っています。

mather

2019/06/06 09:34

print(x_train.shape) などの出力値を書いてください。また、あちこちで x_train 自体が更新されているようですが、エラーになった箇所の直前ではどのような状態になっているのですか？加えて、公式ドキュメントは参照しましたか？ https://keras.io/preprocessing/image/#flow