KerasでUNetとLSTMを組み合わせたモデルを構築したい

UNetとLSTMの組み合わせでのエラー（Graph disconnected: cannot obtain value for tensor Tensor）

Kerasを用いて、UNetとLSTMを組み合わせたモデルを作成しようとしております
（LSTMへの入力は時系列画像を畳み込んだベクトルとし、LSTMから出力される時系列の
隠れ状態ベクトルに転置畳み込みを実行して再び画像に戻す）。
その際に以下のようなエラーが出てしまいます。
原因や解決方法をご教示いただきたいです。

発生している問題・エラーメッセージ

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("first:0", shape=(?, 256, 256, 3), dtype=float32) at layer "first". The following previous layers were accessed without issue: []

該当のソースコード

【追記】後者のソースコードがmodel.pyで、
前者のソースコードを実行しています。

python
1mport numpy as np
2from models import *
3from keras.optimizers import Adam
4
5frames = 20
6height = 256
7width = 256
8in_channels = 3
9out_channels = 3
10dropout_rate = 0.0
11optimizer = Adam(lr=0.0002, beta_1=0.5, beta_2=0.999)
12
13UNet_LSTM = UNet_LSTM(frames, height, width, in_channels, out_channels, dropout_rate)
14
15all = UNet_LSTM.build_all()
16all.compile(loss='binary_crossentropy', optimizer=optimizer)
17all.summary()

python
1from keras.layers.advanced_activations import LeakyReLU
2from keras.layers.convolutional import Conv2D, UpSampling2D
3from keras_contrib.layers.normalization.instancenormalization import InstanceNormalization
4from keras.layers import Input, Dropout, ReLU, Concatenate, Reshape, CuDNNLSTM
5from keras.layers.wrappers import TimeDistributed
6from keras.initializers import glorot_normal
7from keras.models import Model
8
9class UNet_LSTM():
10    def __init__(self, frames, height, width, in_channels, out_channels, dropout_rate):
11        self.frames = frames
12        self.height = height
13        self.width = width
14        self.in_channels = in_channels
15        self.out_channels = out_channels
16        self.input_imgs_shape = (self.frames, self.height, self.width, self.in_channels)
17        self.output_imgs_shape = (self.frames, self.height, self.width, self.out_channels)
18        self.input_img_shape = (self.height, self.width, self.in_channels)
19        self.output_img_shape = (self.height, self.width, self.out_channels)
20        self.dropout_rate = dropout_rate
21
22    def build_all(self):
23        """U-Net+LSTM"""
24        input_imgs = Input(shape=self.input_imgs_shape, name='all')
25        first = self.build_first()
26        encoded_frames = TimeDistributed(first)(input_imgs)
27
28        length_of_sequences = encoded_frames.shape[1].value
29        latent_dim = encoded_frames.shape[-1].value
30        encoded_frames = Reshape((length_of_sequences, latent_dim))(encoded_frames)
31
32        latent_vector = self.build_LSTM(length_of_sequences, latent_dim)(encoded_frames)
33        latent_vector = Reshape((length_of_sequences, 1, 1, latent_dim))(latent_vector)
34#        return Model(input_imgs, latent_vector)
35
36        latter = self.build_latter(latent_dim)
37        output_imgs = TimeDistributed(latter)(latent_vector)
38        return Model(input_imgs, output_imgs)
39
40    def build_first(self):
41        #Image input
42        d0 = Input(shape=self.input_img_shape, name='first')
43
44        # Downsampling
45        self.d1 = self.conv2d(layer_input=d0, filters=64, f_size=4,
46                              normalization=False, alpha=0.2, dropout_rate=0)
47        self.d2 = self.conv2d(layer_input=self.d1, filters=128, f_size=4,
48                              normalization=True, alpha=0.2, dropout_rate=0)
49        self.d3 = self.conv2d(layer_input=self.d2, filters=256, f_size=4,
50                              normalization=True, alpha=0.2, dropout_rate=0)
51        self.d4 = self.conv2d(layer_input=self.d3, filters=512, f_size=4,
52                              normalization=True, alpha=0.2, dropout_rate=self.dropout_rate)
53        self.d5 = self.conv2d(layer_input=self.d4, filters=512, f_size=4,
54                              normalization=True, alpha=0.2, dropout_rate=self.dropout_rate)
55        self.d6 = self.conv2d(layer_input=self.d5, filters=512, f_size=4,
56                              normalization=True, alpha=0.2, dropout_rate=self.dropout_rate)
57        self.d7 = self.conv2d(layer_input=self.d6, filters=512, f_size=4,
58                              normalization=True, alpha=0.2, dropout_rate=self.dropout_rate)
59        d8 = self.conv2d(layer_input=self.d7, filters=512, f_size=4,
60                         normalization=False, alpha=0.2, dropout_rate=self.dropout_rate)
61
62        return Model(d0, d8)
63
64    def build_LSTM(self, length_of_sequences, latent_dim):
65        x = Input(shape=(length_of_sequences, latent_dim, ), name='LSTM')
66        # LSTM
67        y = CuDNNLSTM(latent_dim, batch_input_shape=(None, length_of_sequences, latent_dim),
68                      kernel_initializer=glorot_normal(seed=1),
69                      return_sequences=True, stateful=False)(x)
70        return Model(x, y)
71
72    def build_latter(self, latent_dim):
73        # Latent vector input
74        z = Input(shape=(1, 1, latent_dim), name='latter')
75
76        # Upsampling
77        u1 = self.deconv2d(layer_input=z, skip_input=self.d7, filters=512,
78                           f_size=4, dropout_rate=self.dropout_rate)
79        u2 = self.deconv2d(layer_input=u1, skip_input=self.d6, filters=512,
80                           f_size=4, dropout_rate=self.dropout_rate)
81        u3 = self.deconv2d(layer_input=u2, skip_input=self.d5, filters=512,
82                           f_size=4, dropout_rate=self.dropout_rate)
83        u4 = self.deconv2d(layer_input=u3, skip_input=self.d4, filters=512,
84                           f_size=4, dropout_rate=self.dropout_rate)
85        u5 = self.deconv2d(layer_input=u4, skip_input=self.d3, filters=256,
86                           f_size=4, dropout_rate=0)
87        u6 = self.deconv2d(layer_input=u5, skip_input=self.d2, filters=128,
88                           f_size=4, dropout_rate=0)
89        u7 = self.deconv2d(layer_input=u6, skip_input=self.d1, filters=64,
90                           f_size=4, dropout_rate=0)
91
92        u8 = UpSampling2D(size=2)(u7)
93        output_img = Conv2D(self.out_channels, kernel_size=4, strides=1,
94                            padding='same', activation='tanh')(u8)
95        print('output_img.shape= ', output_img.shape)
96        return Model(z, output_img)
97
98    @staticmethod
99    def conv2d(layer_input, filters, f_size=4,
100               normalization=True, alpha=0.2, dropout_rate=0):
101               # layer_input:   input to this layer (figure or feather map)
102               # filters:       Number of channels
103               # f_size:        kernel size
104               # normalization: impliment the Instance Normalization if True
105               # alpha:         parameter of LeakyReLU (gradient in x < 0)
106               # dropout_rate:  dropout ratio (0-1)
107        # convolution===>(normalization===>)activation===>(dropout===>)
108        d = Conv2D(filters, kernel_size=f_size, strides=2, padding='same')(layer_input)
109        if normalization:
110            d = InstanceNormalization()(d)
111        d = LeakyReLU(alpha=alpha)(d)
112        if dropout_rate:
113            d = Dropout(dropout_rate)(d)
114
115        return d
116
117    @staticmethod
118    def deconv2d(layer_input, skip_input, filters,
119                 f_size=4, dropout_rate=0):
120                 # layer_input:   input to this layer (figure or feather map)
121                 # skip_input:    input from skip-connection
122                 # filters:       Number of channels
123                 # f_size:        kernel size
124                 # dropout_rate:  dropout ratio (0-1)
125        # transpose convolution===>normalization===>activation===>(dropout===>)
126        u = UpSampling2D(size=2)(layer_input)
127        u = Conv2D(filters, kernel_size=f_size, strides=1, padding='same')(u)
128        u = InstanceNormalization()(u)
129        u = ReLU()(u)
130        if dropout_rate:
131            u = Dropout(dropout_rate)(u)
132        print('u.shape= ',u.shape)
133        print('skip_input.shape= ', skip_input.shape)
134        u = Concatenate()([u, skip_input])
135        return u
136

試したこと

input層の入力データの形に関するエラーのようなので、
print文にて様々な場所でデータの形を出力してみました。
結果としては、自分が意図した通りの形になっていました。

input_imgs.shape=  (?, 20, 256, 256, 3)
self.d7.shape=  (?, 2, 2, 512)
encoded_frames.shape=  (?, 20, 1, 1, 512)
encoded_frames.shape=  (?, 20, 512)
latent_vector.shape=  (?, 20, 512)
latent_vector.shape=  (?, 20, 1, 1, 512)
u.shape=  (?, 2, 2, 512)
skip_input.shape=  (?, 2, 2, 512)
u.shape=  (?, 4, 4, 512)
skip_input.shape=  (?, 4, 4, 512)
u.shape=  (?, 8, 8, 512)
skip_input.shape=  (?, 8, 8, 512)
u.shape=  (?, 16, 16, 512)
skip_input.shape=  (?, 16, 16, 512)
u.shape=  (?, 32, 32, 256)
skip_input.shape=  (?, 32, 32, 256)
u.shape=  (?, 64, 64, 128)
skip_input.shape=  (?, 64, 64, 128)
u.shape=  (?, 128, 128, 64)
skip_input.shape=  (?, 128, 128, 64)
output_img.shape=  (?, 256, 256, 3)

LSTMまで、つまりコメントアウトしているreturn Model(input_imgs, latent_vector)まで
であればエラーは出ず、モデルのsummary情報が出ます。

Model: "model_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
all (InputLayer)             (None, 20, 256, 256, 3)   0         
_________________________________________________________________
time_distributed_1 (TimeDist (None, 20, 1, 1, 512)     19535820  
_________________________________________________________________
reshape_1 (Reshape)          (None, 20, 512)           0         
_________________________________________________________________
model_2 (Model)              (None, 20, 512)           2101248   
_________________________________________________________________
reshape_2 (Reshape)          (None, 20, 1, 1, 512)     0         
=================================================================

UNetの後半部分を足すとエラーが出てしまいます。
また、エラーが出るのが
Tensor("first:0", shape=(?, 256, 256, 3), dtype=float32) at layer "first"
にあるようにUNetの前半部分の入力層（"first"）であるのが理解できず、
試行錯誤したのですが解決できませんでした。

是非ともお力をお借りしたく、どうぞよろしくお願いします。