顔画像から年齢を予測でのエラー

顔画像から年齢を予測（https://qiita.com/ha9kberry/items/314afb56ee7484c53e6f）というサイトを参考にしつつ、進めていました。しかし、下のコードの一番最後の部分で一つエラーが出てしまいました。その改善方法が見当もつかないので、どうぞよろしくお願いいたします。

Python
1import numpy as np
2import pandas as pd
3import matplotlib.pyplot as plt
4%config InlineBackend.figure_formats = {'png', 'retina'}
5import os, zipfile, io, re
6from PIL import Image
7from sklearn.model_selection import train_test_split
8from keras.applications.xception import Xception
9from keras.models import Model, load_model
10from keras.layers.core import Dense
11from keras.layers.pooling import GlobalAveragePooling2D
12from keras.optimizers import Adam, RMSprop, SGD
13from keras.utils.np_utils import to_categorical
14from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard, ReduceLROnPlateau
15from keras.preprocessing.image import ImageDataGenerator
16
17# 画像入力サイズ
18image_size=100
19# ZIP読み込み
20z = zipfile.ZipFile('./UTKFace_train.zip')
21# 画像ファイルパスのみ取得S
22imgfiles = [ x for x in z.namelist() if re.search(r"^UTKFace_train.*jpg$", x) ]
23
24X=[]
25Y=[]
26for imgfile in imgfiles:
27    # ZIPから画像読み込み
28    image = Image.open(io.BytesIO(z.read(imgfile)))
29    # RGB変換
30    image = image.convert('RGB')
31    # リサイズ
32    image = image.resize((image_size, image_size))
33    # 画像から配列に変換
34    data = np.asarray(image)
35    file = os.path.basename(imgfile)
36    file_split = [i for i in file.split('_')]
37    X.append(data)
38    Y.append(int(file_split[0]))
39z.close()
40X = np.array(X)
41Y = np.array(Y)
42del z, imgfiles
43
44print(X.shape, Y.shape)
45>>>>(24107, 100, 100, 3) (24107,)
46
47X_train, X_test, y_train, y_test = train_test_split(
48    X,
49    Y,
50    random_state = 0,
51    test_size = 0.2
52)
53del X,Y
54print(X_train.shape, y_train.shape, X_test.shape, y_test.shape) 
55>>>>(19285, 100, 100, 3) (19285,) (4822, 100, 100, 3) (4822,)
56
57# データ型の変換＆正規化
58X_train = X_train.astype('float32') / 255
59X_test = X_test.astype('float32') / 255
60# trainデータからvalidデータを分割
61X_train, X_valid, y_train, y_valid = train_test_split(
62    X_train,
63    y_train,
64    random_state = 0,
65    test_size = 0.2
66)
67print(X_train.shape, y_train.shape, X_valid.shape, y_valid.shape)
68>>>>(15428, 100, 100, 3) (15428,) (3857, 100, 100, 3) (3857,)
69
70# model building
71base_model = Xception(
72    include_top = False,
73    weights = "imagenet",
74    input_shape = None
75)
76x = base_model.output
77x = GlobalAveragePooling2D()(x)
78x = Dense(1024, activation='relu')(x)
79predictions = Dense(1)(x)
80datagen = ImageDataGenerator(
81    featurewise_center = False,
82    samplewise_center = False,
83    featurewise_std_normalization = False,
84    samplewise_std_normalization = False,
85    zca_whitening = False,
86    rotation_range = 0,
87    width_shift_range = 0.1,
88    height_shift_range = 0.1,
89    horizontal_flip = True,
90    vertical_flip = False
91)
92datagen = ImageDataGenerator(
93    featurewise_center = False,
94    samplewise_center = False,
95    featurewise_std_normalization = False,
96    samplewise_std_normalization = False,
97    zca_whitening = False,
98    rotation_range = 0,
99    width_shift_range = 0.1,
100    height_shift_range = 0.1,
101    horizontal_flip = True,
102    vertical_flip = False
103)
104# RMSE
105from keras import backend as K
106def root_mean_squared_error(y_true, y_pred):
107        return K.sqrt(K.mean(K.square(y_pred - y_true), axis = -1)) 
108# ネットワーク定義
109model = Model(inputs = base_model.input, outputs = predictions)
110
111#108層までfreeze
112for layer in model.layers[:108]:
113    layer.trainable = False
114
115    # Batch Normalizationのfreeze解除
116    if layer.name.startswith('batch_normalization'):
117        layer.trainable = True
118    if layer.name.endswith('bn'):
119        layer.trainable = True
120
121#109層以降、学習させる
122for layer in model.layers[108:]:
123    layer.trainable = True
124
125# layer.trainableの設定後にcompile
126model.compile(
127    optimizer = Adam(),
128    loss = root_mean_squared_error,
129)
130
131%%time
132hist = model.fit_generator(
133    datagen.flow(X_train, y_train, batch_size = 32),
134    steps_per_epoch = X_train.shape[0] // 32,
135    epochs = 50,
136    validation_data = (X_valid, y_valid),
137    callbacks = [early_stopping, reduce_lr],
138    shuffle = True,
139    verbose = 1
140)
141>>>> #Epochが 1からずら～っと並びます
142>>>>Epoch 00037: ReduceLROnPlateau reducing learning rate to 1.000000082740371e-09.
143>>>>Epoch 38/50
144>>>>482/482 [==============================] - 3221s 7s/step - loss: nan - val_loss: >>>>nan
145>>>>Epoch 00038: early stopping
146>>>>Wall time: 1d 5h 59min 46s
147
148score = model.evaluate(X_test, y_test, verbose=1)
149print("evaluate loss: {}".format(score))
150>>>>4822/4822 [==============================] - 605s 125ms/step
151>>>>evaluate loss: nan
152