将keras中两层的输出相乘

Question

我正试图用这种设计在keras中实现一个神经（ish）网络：http://nlp.cs.rpi.edu/paper/AAAI15.pdf

该算法基本上有三个输入。输入2和输入3乘以相同的权重矩阵W1以产生O2和O3。输入1乘以W2，得到O1。然后，我们需要取O1*O2和O1*O3的点积。

我正试图在keras中实现这一点。

我的第一个想法是使用keras Graph类，并使W1成为具有两个输入和两个输出的共享节点层。到目前为止还好。

然后出现了如何取这两个输出与O1的点积的问题。

我试图定义一个自定义函数：

   def layer_mult(X, Y):
       return K.dot(X * K.transpose(Y))

然后：

ntm.add_node(Lambda(layer_mult, output_shape = (1,1)), name = "ls_pos", inputs = ["O1", "O2"])
ntm.add_node(Lambda(layer_mult, output_shape = (1,1)), name = "ls_neg", inputs = ["O1", "O3"])

编译时出现的问题是，keras只想给Lambda层一个输入：

   1045         func = types.FunctionType(func, globals())
   1046         if hasattr(self, 'previous'):
-> 1047             return func(self.previous.get_output(train))
   1048         else:
   1049             return func(self.input)
TypeError: layer_mult() takes exactly 2 arguments (1 given)

我认为另一种选择可能是使用Merge类，该类将dot作为允许的合并类型。但是，Merge类的输入层必须传递给构造函数。因此，似乎没有办法将共享节点的输出放入Merge，以将Merge添加到Graph。

如果我使用的是Sequential容器，我可以将它们放入Merge中。但是，没有办法实现两个Sequential层需要共享相同的权重矩阵。

我想尝试将O1、O2和O3连接到一个单独的向量中作为输出层，然后在目标函数中进行乘法运算。但是，这需要目标函数分割其输入，这在keras中似乎是不可能的（相关Theano函数不能传递给keras API）。

有人知道解决方案吗？

编辑：

我认为我已经取得了一些进展，因为我发现shared_node正在实现dot（即使文档中没有）。

所以我去了：

ntm = Graph()
ntm.add_input(name='g', input_shape=(300,))  #  Vector of 300 units, normally distributed around zero
ntm.add_node([pretrained bit], name = "lt", input = "g") # 300 * 128, output = (,128)
n_docs = 1000
ntm.add_input("d_pos", input_shape = (n_docs,)) # (,n_docs)
ntm.add_input("d_neg", input_shape = (n_docs,)) # (,n_docs)
ntm.add_shared_node(Dense(128, activation = "softmax", 
#                      weights = pretrained_W1, 
                      W_constraint = unitnorm(), 
                      W_regularizer = l2(0.001)
                      ), name = "ld", 
                    inputs = ["d_pos", "d_neg"],  
                    outputs = ["ld_pos", "ld_neg"], 
                    merge_mode=None) # n_docs * 128, output = (,128) * 2
ntm.add_shared_node(ActivityRegularization(0,0),   #ActivityRegularization is being used as a passthrough - the function of the node is to dot* its inputs
                    name = "ls_pos", 
                    inputs = ["lt", "d_pos"], 
                    merge_mode = 'dot')  # output = (,1)
ntm.add_shared_node(ActivityRegularization(0,0), 
                    name = "ls_neg", 
                    inputs = ["lt", "d_neg"], 
                    merge_mode = 'dot')  # output = (,1)
ntm.add_shared_node(ActivityRegularization(0,0), 
                    name = "summed", 
                    inputs = ["ls_pos", "ls_neg"], 
                    merge_mode = 'sum') # output = (,1)
ntm.add_node(ThresholdedReLU(0.5), 
             input = "summed", name = "loss") # output = (,1)
ntm.add_output(name = "loss_out", 
               input= "loss")
def obj(X, Y):
    return K.sum(Y)
ntm.compile(loss = {'loss_out' : obj},  optimizer = "sgd")

现在的错误是：

>>> ntm.compile(loss = {'loss_out' : obj},  optimizer = "sgd")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "build/bdist.macosx-10.5-x86_64/egg/keras/models.py", line 602, in compile
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/advanced_activations.py", line 149, in get_output
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 117, in get_input
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 1334, in get_output
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 1282, in get_output_sum
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 1266, in get_output_at
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 730, in get_output
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 117, in get_input
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 1340, in get_output
  File "build/bdist.macosx-10.5-x86_64/egg/keras/layers/core.py", line 1312, in get_output_dot
  File "/Volumes/home500/anaconda/envs/[-]/lib/python2.7/site-packages/theano/tensor/var.py", line 360, in dimshuffle
    pattern)
  File "/Volumes/home500/anaconda/envs/[-]/lib/python2.7/site-packages/theano/tensor/elemwise.py", line 164, in __init__
    (input_broadcastable, new_order))
ValueError: ('You cannot drop a non-broadcastable dimension.', ((False, False, False, False), (0, 'x')))

Answer 1

您可以使用此

main_branch.add（Merge（[branch_1，branch_2]，mode='dot'））

Answer 2

我也面临类似的问题。我想了一个解决方案，但还没有尝试。

1. 将卷积层用于将Input2和Input3都作为输入的序列模型A。以这种方式，相同的卷积核将被应用于Input2和Input3，也就是相同的权重W1。

2. 将Input1作为另一个Sequential model B.的输入

3. 使用合并层合并a和B的输出。点也可以通过合并层的自定义功能来完成。