Python UnboundLocalError: local variable 'batch_x' referenced before assignment

前提・実現したいこと

以下のサイトを参考に、自作のgeneratorを作成しました。
https://qiita.com/koshian2/items/909360f50e3dd5922f32
しかし、コードを実行してみると、以下のようなエラーが発生してしまいますが、どう修正したらいいかわからず困っております。
お手数をお掛けするのですが、修正方法についてお分かりの方がいらっしゃいましたら教えていただけないでしょうか。
どうぞよろしくお願いいたします。

while True:
40 batch_x_2, batch_y_2 = next(batches)
---> 41 m1, m2 = batch_x.shape[0], batch_x_2.shape[0]
42 if m1 < m2:
43 batch_x_2 = batch_x_2[:m1]

UnboundLocalError: local variable 'batch_x' referenced before assignment

該当のソースコード

Python
1# model構築の準備
2import keras
3from keras.models import Model
4from keras.layers import Dense, GlobalAveragePooling2D,Input,Dropout,Activation
5from keras.preprocessing.image import ImageDataGenerator
6from keras import applications
7from keras.preprocessing.image import ImageDataGenerator
8from keras.optimizers import Adam
9from keras.callbacks import CSVLogger,EarlyStopping
10import numpy as np
11from keras import backend as K
12from keras.engine.topology import Layer
13import numpy as np
14import tensorflow as tf
15from keras.preprocessing.image import load_img, img_to_array, array_to_img
16import os
17import matplotlib.pyplot as plt
18%matplotlib inline
19
20
21class MyImageDataGenerator(ImageDataGenerator):
22    def __init__(self, featurewise_center = False, samplewise_center = False, 
23                 featurewise_std_normalization = False, samplewise_std_normalization = False, 
24                 zca_whitening = False, zca_epsilon = 1e-06, rotation_range = 0.0, width_shift_range = 0.0, 
25                 height_shift_range = 0.0, brightness_range = None, shear_range = 0.0, zoom_range = 0.0, 
26                 channel_shift_range = 0.0, fill_mode = 'nearest', cval = 0.0, horizontal_flip = False, 
27                 vertical_flip = False, rescale = None, preprocessing_function = None, data_format = None, validation_split = 0.0, 
28                 random_crop = None, mix_up_alpha = 0.0):
29        # 親クラスのコンストラクタ
30        super().__init__(featurewise_center, samplewise_center, featurewise_std_normalization, samplewise_std_normalization, zca_whitening, zca_epsilon, rotation_range, width_shift_range, height_shift_range, brightness_range, shear_range, zoom_range, channel_shift_range, fill_mode, cval, horizontal_flip, vertical_flip, rescale, preprocessing_function, data_format, validation_split)
31        # 拡張処理のパラメーター
32        # Mix-up
33        assert mix_up_alpha >= 0.0
34        self.mix_up_alpha = mix_up_alpha
35     # Mix-up
36    # 参考 https://qiita.com/yu4u/items/70aa007346ec73b7ff05
37    def mix_up(self, X1, y1, X2, y2):
38        assert X1.shape[0] == y1.shape[0] == X2.shape[0] == y2.shape[0]
39        batch_size = X1.shape[0]
40        l = np.random.beta(self.mix_up_alpha, self.mix_up_alpha, batch_size)
41        X_l = l.reshape(batch_size, 1, 1, 1)
42        y_l = l.reshape(batch_size, 1)
43        X = X1 * X_l + X2 * (1-X_l)
44        y = y1 * y_l + y2 * (1-y_l)
45        return X, y
46
47    def flow_from_directory(self, directory, target_size = (299,299), color_mode = 'rgb', 
48                            classes = None, class_mode = 'categorical', batch_size = 12, shuffle = True, 
49                            seed = None, save_to_dir = None, save_prefix = '', save_format = 'png', 
50                            follow_links = False, subset = None, interpolation = 'nearest'):
51        # 親クラスのflow_from_directory
52        batches = super().flow_from_directory(directory, target_size, color_mode, classes, class_mode, batch_size, shuffle, seed, save_to_dir, save_prefix, save_format, follow_links, subset, interpolation)
53        # 拡張処理
54        while True:
55            if self.mix_up_alpha > 0:
56                while True:
57                    batch_x_2, batch_y_2 = next(batches)
58                    m1, m2 = batch_x.shape[0], batch_x_2.shape[0]
59                    if m1 < m2:
60                        batch_x_2 = batch_x_2[:m1]
61                        batch_y_2 = batch_y_2[:m1]
62                        break
63                    elif m1 == m2:
64                        break
65                batch_x, batch_y = self.mix_up(batch_x, batch_y, batch_x_2, batch_y_2)
66             # 返り値
67            yield (batch_x, batch_y)
68
69train_dir = './train'
70validation_dir = './validation'
71
72train_datagen = MyImageDataGenerator(
73    rescale=1/255.0,
74    mix_up_alpha=0.2,
75    #rotation_range=8,
76    horizontal_flip=True)
77
78train_generator=train_datagen.flow_from_directory(
79          train_dir,
80          target_size=(299,299),
81          batch_size=12,#25,
82          class_mode='categorical',
83          shuffle=True)
84
85
86validation_datagen=ImageDataGenerator(rescale=1.0/255.)
87      
88validation_generator=validation_datagen.flow_from_directory(
89            validation_dir,
90            target_size=(299,299),
91            batch_size=12,#25,
92            class_mode='categorical',
93            shuffle=True)
94
95
96batch_size = 12
97
98base_model=keras.applications.inception_resnet_v2.InceptionResNetV2(input_shape=(299,299,3),
99                       weights='imagenet',
100                       include_top=False)
101    
102x = base_model.output
103x = GlobalAveragePooling2D()(x)
104x = Dense(1024, activation='relu')(x)
105predictions = Dense(4, activation='softmax')(x)
106model = Model(inputs=base_model.input,outputs=predictions)
107    
108model.summary()
109
110 
111model.compile(optimizer=Adam(lr=0.001),
112                  loss='categorical_crossentropy',
113                  metrics=['accuracy'])
114
115callbacks_list = [
116        callbacks.ModelCheckpoint(
117        filepath="model.ep{epoch:02d}.h5",#delsavepath,
118        save_best_only=True),
119        
120        #バリデーションlossが改善しなくなったら学習率を変更する
121callbacks.ReduceLROnPlateau(
122            monitor="val_loss",
123            factor=0.8,
124            patience=5,
125            verbose=1)]#,
126
127#callbacks.EarlyStopping(monitor='val_loss', patience=7, verbose=1)]
128    
129model.fit_generator(train_generator, steps_per_epoch=7208, epochs=10, validation_steps=1805, validation_data=validation_generator, callbacks=callbacks_list)
130