通过for循环使用功能api创建keras输入层？

Question

假设我需要动态生成特定于用户的 keras 模型。 每个用户都可以拥有可变数量的分类输入，但是一旦您知道分类输入的数量，手动构建模型就变得很简单了。

我想编写一个函数，给出每个分类变量的基数列表将返回一个适当的模型。 我对这个问题的第一次尝试产生了以下解决方案，但执行这样的字符串似乎不正确。

from keras.layers import Dense,Embedding,Input,Flatten,Add
from keras.models import Model

def build_model(input_cardinalities,num_outputs):
    layers = []
    inputs = []
    for i,cardinality in enumerate(input_cardinalities):
        exec("input{0} = Input(shape=[1], name='input{0}')".format(i))
        exec("embedding{0} =  Embedding({1}, 20, name='embedding{0}')(input{0})".format(i,cardinality))
        exec("vec{0} = Flatten(name='flatten{0}')(embedding{0})".format(i))
        exec("layers.append(vec{0})".format(i))
        exec("inputs.append(input{0})".format(i))
    context_layer = Add(layers)
    dense1 = Dense(50, name='Dense1',activation='relu')(context_layer)
    dense2 = Dense(num_outputs, name='Output', activation='softmax')(dense1)
    model = Model(inputs,dense2)
    model.compile('sgd','categorical_crossentropy')
    return model

我只是觉得像这样执行字符串不太舒服，但这是我能想到的做我想做的事情的唯一方法。 我只是觉得应该有更好的方法来做到这一点。

Answer 1

实际上根本不需要使用exec ，您一次构建一个输入/嵌入，然后将它们存储到列表中。 这是正确的方法，它不需要exec 。

def build_model(input_cardinalities,num_outputs):
    layers = []
    inputs = []
    for i,cardinality in enumerate(input_cardinalities):
        input = Input(shape=[1], name='input{0}'.format(i))
        embedding =  Embedding(cardinality, 20, name='embedding{0}'.format(i))
        vec = Flatten(name='flatten{0}'.format(i))(embedding)
        layers.append(vec)
        inputs.append(input)
    context_layer = Add()(layers)
    dense1 = Dense(50, name='Dense1',activation='relu')(context_layer)
    dense2 = Dense(num_outputs, name='Output', activation='softmax')(dense1)
    model = Model(inputs,dense2)
    model.compile('sgd','categorical_crossentropy')
    return model

另请注意，我更正了Add()(layers)调用。

Answer 2

我遇到了类似的问题，我想使用 for 循环实例化多个输入层。 首先，我尝试重现上述内容，并仅完成上述答案的全面性（和打印模型摘要）。

from keras.layers import Dense,Embedding,Input,Flatten,Add
from keras.models import Model
from tensorflow.keras.utils import plot_model

def build_model(input_cardinalities,num_outputs):
    layers = []
    inputs = []
    for i,cardinality in enumerate(input_cardinalities):
        input = Input(shape=[1], name='input{0}'.format(i))
        embedding =  Embedding(cardinality, 20, name='embedding{0}'.format(i))(input)
        vec = Flatten(name='flatten{0}'.format(i))(embedding)
        layers.append(vec)
        inputs.append(input)
    context_layer = Add(name='context')(layers)
    dense1 = Dense(50, name='Dense1',activation='relu')(context_layer)
    dense2 = Dense(num_outputs, name='Output', activation='softmax')(dense1)
    model = Model(inputs,dense2)
    model.compile('sgd','categorical_crossentropy')
    return model

然后构建模型

input_cardinalities = [1,2,3]
num_outputs = 6
model = build_model(input_cardinalities, num_outputs)
model.summary()
plot_model(model, 'model.png', show_shapes=True)

输出看起来像这样

Model: "model_14"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input0 (InputLayer)             [(None, 1)]          0                                            
__________________________________________________________________________________________________
input1 (InputLayer)             [(None, 1)]          0                                            
__________________________________________________________________________________________________
input2 (InputLayer)             [(None, 1)]          0                                            
__________________________________________________________________________________________________
embedding0 (Embedding)          (None, 1, 20)        20          input0[0][0]                     
__________________________________________________________________________________________________
embedding1 (Embedding)          (None, 1, 20)        40          input1[0][0]                     
__________________________________________________________________________________________________
embedding2 (Embedding)          (None, 1, 20)        60          input2[0][0]                     
__________________________________________________________________________________________________
flatten0 (Flatten)              (None, 20)           0           embedding0[0][0]                 
__________________________________________________________________________________________________
flatten1 (Flatten)              (None, 20)           0           embedding1[0][0]                 
__________________________________________________________________________________________________
flatten2 (Flatten)              (None, 20)           0           embedding2[0][0]                 
__________________________________________________________________________________________________
context (Add)                   (None, 20)           0           flatten0[0][0]                   
                                                                 flatten1[0][0]                   
                                                                 flatten2[0][0]                   
__________________________________________________________________________________________________
Dense1 (Dense)                  (None, 50)           1050        context[0][0]                    
__________________________________________________________________________________________________
Output (Dense)                  (None, 6)            306         Dense1[0][0]                     
==================================================================================================
Total params: 1,476
Trainable params: 1,476
Non-trainable params: 0
__________________________________________________________________________________________________

为了更好的可视化，请在此处查看此模型的 plot_model() 返回

Answer 3

末尾缺少一个(input) ：

embedding =  Embedding(cardinality, 20, name='embedding{0}'.format(i))(input)