如何将前置模型数据适配到最后一个模型？代码fit报错求助

我正在构建包含约40个模型的复杂链式结构，当前最后一个模型的fit方法无法运行。需求是仅训练最后一个模型cModel，它接收前置a、b模型的输出及额外输入s_in，生成3维输出匹配c_out，后续要构建数十个此类实例链式串联。

原代码 #

#Input
a_in=np.array([[3,4,5],[2,3,4],[1,2,3],[4,6,8]])
b_in=np.array([[9,5,3],[1,0,2],[7,3,1],[6,6,6]])
s_in = np.array([[5],[7],[9],[0]])

#Output
c_out = np.array([[6,2,3],[0,2,7],[9,6,1],[8,6,5]])

a = keras.Input(shape=(3,), dtype=tf.float64)
aModel = keras.layers.Flatten()(a)

b = keras.Input(shape=(3,), dtype=tf.float64)
bModel = keras.layers.Flatten()(b)

s = keras.Input(shape=(1,), dtype=tf.float64)

aModel = keras.layers.Dense(10, activation='sigmoid')(a)
aModel = keras.layers.Dense(1, activation='sigmoid')(aModel)
aModel = keras.Model(inputs=a, outputs=aModel, name="a")

bModel = keras.layers.Dense(10, activation='sigmoid')(b)
bModel = keras.layers.Dense(1, activation='sigmoid')(bModel)
bModel = keras.Model(inputs=b, outputs=bModel, name="b")

combine = keras.layers.concatenate([aModel.output, bModel.output, s], dtype=tf.float64) #inject s in the middle stage

cModel = keras.layers.Dense(10, activation='sigmoid')(combine)
cModel = keras.layers.Dense(3, activation='sigmoid')(cModel)
cModel = keras.Model(inputs=combine, outputs=cModel, name="c")

cModel.compile(optimizer='adam', loss='mean_absolute_error', metrics='accuracy')

aModel.summary()
bModel.summary()
cModel.summary()

keras.utils.plot_model(cModel, "baby-architecture.png", show_shapes=True)


cModel.fit([a_in, b_in, s_in],[c_out],epochs=1, shuffle=False, verbose=1)

模型摘要 #

Model: "a"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, 3)]               0         
                                                                 
 dense (Dense)               (None, 10)                40        
                                                                 
 dense_1 (Dense)             (None, 1)                 11        
                                                                 
=================================================================
Total params: 51 (204.00 Byte)
Trainable params: 51 (204.00 Byte)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
Model: "b"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 3)]               0         
                                                                 
 dense_2 (Dense)             (None, 10)                40        
                                                                 
 dense_3 (Dense)             (None, 1)                 11        
                                                                 
=================================================================
Total params: 51 (204.00 Byte)
Trainable params: 51 (204.00 Byte)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
Model: "c"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_4 (InputLayer)        [(None, 3)]               0         
                                                                 
 dense_4 (Dense)             (None, 10)                40        
                                                                 
 dense_5 (Dense)             (None, 3)                 33        
                                                                 
=================================================================
Total params: 73 (292.00 Byte)
Trainable params: 73 (292.00 Byte)
Non-trainable params: 0 (0.00 Byte)

错误信息 #

ValueError: Layer "c" expects 1 input(s), but it received 3 input tensors. 
    Inputs received: [<tf.Tensor 'IteratorGetNext:0' shape=(None, 3) dtype=int64>, 
    <tf.Tensor 'IteratorGetNext:1' shape=(None, 3) dtype=int64>, 
<tf.Tensor 'IteratorGetNext:2' shape=(None, 1) dtype=int64>]

错误原因 #

1. cModel输入定义错误：原代码中cModel = keras.Model(inputs=combine, outputs=cModel, name="c")将拼接后的张量combine作为输入，导致模型只期望接收1个3维张量，但fit时传入了3个不同形状的输入。
2. 未固定前置模型权重：若要仅训练cModel，需将aModel和bModel设置为不可训练，避免其权重被更新。

修改后的代码 #

import numpy as np
import tensorflow as tf
from tensorflow import keras

#Input
a_in=np.array([[3,4,5],[2,3,4],[1,2,3],[4,6,8]], dtype=np.float64)
b_in=np.array([[9,5,3],[1,0,2],[7,3,1],[6,6,6]], dtype=np.float64)
s_in = np.array([[5],[7],[9],[0]], dtype=np.float64)

#Output
c_out = np.array([[6,2,3],[0,2,7],[9,6,1],[8,6,5]], dtype=np.float64)

# 定义并构建a模型
a = keras.Input(shape=(3,), dtype=tf.float64)
a_x = keras.layers.Dense(10, activation='sigmoid')(a)
a_x = keras.layers.Dense(1, activation='sigmoid')(a_x)
aModel = keras.Model(inputs=a, outputs=a_x, name="a")
# 固定a模型权重，不可训练
aModel.trainable = False

# 定义并构建b模型
b = keras.Input(shape=(3,), dtype=tf.float64)
b_x = keras.layers.Dense(10, activation='sigmoid')(b)
b_x = keras.layers.Dense(1, activation='sigmoid')(b_x)
bModel = keras.Model(inputs=b, outputs=b_x, name="b")
# 固定b模型权重，不可训练
bModel.trainable = False

# 定义额外输入s
s = keras.Input(shape=(1,), dtype=tf.float64)

# 获取a、b模型的输出，拼接s输入
a_out = aModel(a)
b_out = bModel(b)
combine = keras.layers.concatenate([a_out, b_out, s], dtype=tf.float64)

# 构建c模型，输入为a、b、s的原始输入，输出为最终预测
c_x = keras.layers.Dense(10, activation='sigmoid')(combine)
c_x = keras.layers.Dense(3, activation='sigmoid')(c_x)
cModel = keras.Model(inputs=[a, b, s], outputs=c_x, name="c")

# 编译c模型
cModel.compile(optimizer='adam', loss='mean_absolute_error', metrics=['accuracy'])

# 查看模型结构
aModel.summary()
bModel.summary()
cModel.summary()

# 训练c模型，传入三个输入
cModel.fit([a_in, b_in, s_in], c_out, epochs=1, shuffle=False, verbose=1)

关键修改点说明 #

1. 固定前置模型权重：通过aModel.trainable = False和bModel.trainable = False锁定a、b模型的参数，训练时仅更新cModel的权重。
2. 正确定义cModel输入：将cModel的输入设置为[a, b, s]（三个原始输入张量），而非拼接后的combine，这样fit时传入[a_in, b_in, s_in]就能匹配模型输入要求。
3. 统一数据类型：确保输入数据的 dtype 与模型输入定义的tf.float64一致，避免类型不匹配问题。

内容的提问来源于stack exchange，提问作者Majoris