Question

我正在使用Keras，我想制作一个以[a0, a1]，[b0, b1, b2]作为输入并以[a0*b0, a0*b1, a0*b2, a1*b0, a1*b1, a1*b2]作为输出的图层。我尝试使用Lambda，但无法成功。这是我的代码：

import numpy as np
from keras.models import Input
from keras.layers import Lambda

def mix(A):
    reshaped = [np.reshape(A[m], (1,np.size(A[m]))) for m in range(len(A))]
    mixed = reshaped[-1]

    for i in range(len(A)-1):
        mixed = np.matmul(np.transpose(reshaped[-i-2]), mixed)
        mixed = np.reshape(mixed, (1,np.size(mixed)))

    return np.reshape(mixed, np.size(mixed))

a = Input(shape=(2,))
b = Input(shape=(3,))
c = Lambda(mix)([a, b])

这是我得到的错误：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-32-07bbf930b48b> in <module>()
      1 a = Input(shape=(2,))
      2 b = Input(shape=(3,))
----> 3 c = Lambda(mix)([a, b])

~\Anaconda3\envs\mind\lib\site-packages\keras\engine\base_layer.py in __call__(self, inputs, **kwargs)
    455             # Actually call the layer,
    456             # collecting output(s), mask(s), and shape(s).
--> 457             output = self.call(inputs, **kwargs)
    458             output_mask = self.compute_mask(inputs, previous_mask)
    459 

~\Anaconda3\envs\mind\lib\site-packages\keras\layers\core.py in call(self, inputs, mask)
    685         if has_arg(self.function, 'mask'):
    686             arguments['mask'] = mask
--> 687         return self.function(inputs, **arguments)
    688 
    689     def compute_mask(self, inputs, mask=None):

<ipython-input-31-bbc21320d8af> in mix(A)
      4 
      5     for i in range(len(A)-1):
----> 6         mixed = np.matmul(np.transpose(reshaped[-i-2]), mixed)
      7         mixed = np.reshape(mixed, (1,np.size(mixed)))
      8 

TypeError: Object arrays are not currently supported

但是，如果我放：

a = np.array([1,2])
b = np.array([3,4,5])
print(mix([a,b]))

然后我得到：

[ 3  4  5  6  8 10]

这正是我的意图。但是我不知道如何正确地将其放入Lambda。

有人可以告诉我如何处理吗？我是Keras的新手，所以我不知道Lambda，Input或其他内容的内部结构。

在Abhijit的评论之后，我更改了代码，如下所示：

import numpy as np
import tensorflow as tf
from keras.models import Input
from keras.layers import Lambda

def mix(A):
    reshaped = [tf.reshape(A[m], (1,tf.size(A[m]))) for m in range(len(A))]
    mixed = reshaped[-1]

    for i in range(len(A)-1):
        mixed = tf.matmul(tf.transpose(reshaped[-i-2]), mixed)
        mixed = tf.reshape(mixed, (1,tf.size(mixed)))

    return tf.reshape(mixed, [tf.size(mixed)])

a = Input(shape=(2,))
b = Input(shape=(3,))
c = Lambda(mix)([a, b])

现在我没有任何错误，但是我认为我没有正确的神经网络。因为执行：

model = Model(inputs=[a,b], outputs=c)
print(model.summary())

我得到：

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_22 (InputLayer)           (None, 2)            0                                            
__________________________________________________________________________________________________
input_23 (InputLayer)           (None, 3)            0                                            
__________________________________________________________________________________________________
lambda_3 (Lambda)               (None,)              0           input_22[0][0]                   
                                                                 input_23[0][0]                   
==================================================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
__________________________________________________________________________________________________

但是请参见图层lambda_3。输出形状不应该是(None, 6)吗？

Answer 1

除了需要使用Keras后端函数（即keras.backend.*）或直接使用后端函数（即tf.*或th.*）外，我认为您正在定义mix的内容不必要地复杂。可以这样简单得多：

from keras import backend as K

def mix(ts):
    t0 = K.expand_dims(ts[0], axis=-1)
    t1 = K.expand_dims(ts[1], axis=1)
    return K.batch_flatten(t0 * t1)

a = Input(shape=(2,))
b = Input(shape=(3,))
c = Lambda(mix)([a, b])

model = Model(inputs=[a,b], outputs=c)

这是测试：

# the reshapes are necessary to make them a batch
a = np.array([1,2]).reshape(1,2)
b = np.array([3,4,5]).reshape(1,3)
print(model.predict([a, b]))

# output
[[ 3.  4.  5.  6.  8. 10.]]

此外，有时Lambda层可以自动推断输出形状。但是，如果您愿意，可以显式设置其输出形状：

c = Lambda(mix, output_shape=(6,))([a, b])

模型摘要：

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_9 (InputLayer)            (None, 2)            0                                            
__________________________________________________________________________________________________
input_10 (InputLayer)           (None, 3)            0                                            
__________________________________________________________________________________________________
lambda_5 (Lambda)               (None, 6)            0           input_9[0][0]                    
                                                                 input_10[0][0]                   
==================================================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
__________________________________________________________________________________________________

使用keras Lambda时，如何处理“ TypeError：当前不支持对象数组”？

1 个答案: