concatenate() 缺少 1 个必需的位置参数：“输入”

Question

我想使用三个导入的网络创建一个多输入 CNN。但是我不完全理解下面的错误。 keras 函数 api 不应该识别导入网络中的输入层吗？

import tensorflow.keras
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import layers, models, Sequential
from tensorflow.keras.applications import ResNet152, InceptionV3, InceptionResNetV2

导入 1 个模型：

resnet152 = ResNet152(
weights='imagenet',
include_top=False,
input_shape=((224, 224, 3))
)

for layer in resnet152.layers:
    layer.trainable = True

合并 3 个模型的输入和输出。

from tensorflow.keras import Model

resnet152_copy = Model(inputs=resnet152.input, outputs=resnet152.output)
inceptionV3_copy = Model(inputs=inceptionV3.input, outputs=inceptionV3.output)
inception_resnet152_copy = Model(inputs=inceptionV3_resnet152.input, outputs=inceptionV3_resnet152.output)


concat_feature_layer = layers.concatenate(axis=4)([resnet152_copy, inceptionV3_copy, inception_resnet152_copy])
fully_connected_dense_big = layers.Dense(1024, activation='relu')(concat_feature_layer)
dropout_one = layers.Dropout(0.5)(fully_connected_dense_big)
flatten_layer = layers.Flatten()(dropout_one)
fully_connected_dense_small = layers.Dense(512, activation='relu')(flatten_layer)
dropout_two = layers.Dropout(0.5)(fully_connected_dense_small)
fully_connected_dense_class = layers.Dense(4, activation='softmax')

model = Model(
    inputs=[resnet152.input, inceptionV3.input, inceptionV3_resnet152.input],
    outputs=fully_connected_dense_class
)

我在 concat 层中不断收到以下错误：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-23-7cef36b1c56e> in <module>
      6 
      7 
----> 8 concat_feature_layer = layers.concatenate(axis=4)([resnet152_copy, inceptionV3_copy, inception_resnet152_copy])
      9 fully_connected_dense_big = layers.Dense(1024, activation='relu')(concat_feature_layer)
     10 dropout_one = layers.Dropout(0.5)(fully_connected_dense_big)

TypeError: concatenate() missing 1 required positional argument: 'inputs'

Answer 1

tf.keras中有两种形式的连接层：

• 你有functional form tf.keras.layers.concatenate，它应该被称为

concat_feature_layer = layers.concatenate([resnet152_copy, inceptionV3_copy, inception_resnet152_copy], axis=4)

• 然后你有 layer object tf.keras.layers.Concatenate 应该首先调用它来实例化对象，然后再对输入进行操作：

concat_feature_layer = layers.Concatenate(axis=4)([resnet152_copy, inceptionV3_copy, inception_resnet152_copy])