Keras模型与自定义图层的拟合失败

Question

我在keras（1.1）中创建了自定义图层：

from keras import backend as K
from keras.engine.topology import Layer
import numpy as np

class MyLayer(Layer):

def __init__(self,input_shape,**kwargs):
    self.W_init = np.random.rand(input_shape[0], input_shape[1], input_shape[2])
    self.input_len = input_shape[0]
    self.output_dim = 1
    super(MyLayer, self).__init__(**kwargs)

def build(self, input_shape):
    # Create a trainable weight variable for this layer.
    self.W = K.variable(self.W_init, name="W")
    self.trainable_weights = [ self.W ]
    super(MyLayer, self).build(input_shape)  # Be sure to call this somewhere!

def call(self, x, mask=None):
    res= K.sum(x*self.W,axis=(1,2))
    res= K.expand_dims(res, -1)
    res = K.expand_dims(res, -1)
    return res

def get_output_shape_for(self, input_shape):
    return (input_shape[0], self.input_len, self.output_dim, self.output_dim)

模型编译成功：

但是当我试图适应它时，我得到了错误：

ValueError: cannot reshape array of size 64 into shape (1,4)
Apply node that caused the error: Reshape{2}(HostFromGpu.0, MakeVector{dtype='int64'}.0)
Toposort index: 895
Inputs types: [TensorType(float32, vector), TensorType(int64, vector)]
Inputs shapes: [(64,), (2,)]
Inputs strides: [(4,), (8,)]
Inputs values: ['not shown', array([1, 4])]
Inputs type_num: [11, 7]
Outputs clients: [[InplaceDimShuffle{0,1,x,x}(Reshape{2}.0)]]

Backtrace when the node is created(use Theano flag traceback.limit=N to make it longer):
  File "<ipython-input-155-09ee1207017c>", line 22, in get_my_model_2
    Dense(10, activation='softmax')
  File "/home/universal/anaconda3/envs/practicecourse2/lib/python2.7/site-packages/keras/models.py", line 255, in __init__
    self.add(layer)

自定义图层中的可训练权重有问题吗？

Answer 1

由于您使用的是Theano，因此宽度和高度轴为(2,3)，而不是(1,2)。

也就是说，你应该改变这一行：

    res= K.sum(x*self.W,axis=(1,2))

到

    res= K.sum(x*self.W,axis=(2,3))

引发错误的原因是call()函数的输出形状为(None, 4, 1, 1)而不是(None, 64, 1, 1)，如get_output_shape_for()中所述。