如何从同一个 Keras 层获得不同的输出，然后将它们组合起来？

debugcn 发表于 Dev

菲利波·卡普尔索

所以基本上，我正在创建一个带有 Keras 和 Tensorflow 后端的 CNN。我现在想插入两个具有相同输入层的层，然后将它们连接起来，如下所示：

model = Sequential()

model.add(Convolution1D(128, (4), activation='relu',input_shape=(599,128))
model.add(MaxPooling1D(pool_size=(4)))
model.add(Convolution1D(256, (4), activation='relu')
model.add(MaxPooling1D(pool_size=(2)))
model.add(Convolution1D(256, (4), activation='relu')
model.add(MaxPooling1D(pool_size=(2)))
model.add(Convolution1D(512, (4), activation='relu')

# output 1 = GlobalMaxPooling1D() # from last conv layer
# output 2 = GlobalAveragePooling1D() # from last conv layer
# model.add(Concatenate((output 1, output 2))
# at this point output should have a shape of 1024,1 (from 512 * 2)

model.add(Dense(1024))
model.add(Dense(512))

要以简单的方式以图形方式显示：

    ...
    cv4
    / \
   /   \
gMaxP|gAvrgP   (each 512,)
   \   /
    \ /
   dense(1024,)

我有一种感觉，我错过了一些非常明显的东西。有人能叫醒我吗？

达霍夫

使用Model 类 API，那么它应该是这样的：

inputs = Input(shape=(599,128), name='image_input')

x = Convolution1D(128, (4), activation='relu')(inputs)
x = MaxPooling1D(pool_size=(4))(x)
x = Convolution1D(256, (4), activation='relu')(x)
x = MaxPooling1D(pool_size=(2))(x)
x = Convolution1D(256, (4), activation='relu')(x)
x = MaxPooling1D(pool_size=(2))(x)
x = Convolution1D(512, (4), activation='relu')(x)


output_1 = GlobalMaxPooling1D()(x) # from last conv layer
output_2 = GlobalAveragePooling1D()(x) # from last conv layer
x = concatenate([output_1, output_2])
# at this point output should have a shape of 1024,1 (from 512 * 2)

x = Dense(1024)(x)
x = Dense(512)(x)

本文收集自互联网，转载请注明来源。

如有侵权，请联系[email protected] 删除。