我试图关注这篇博客文章，他们在其中构建了一个编码器和解码器共享权重的自动编码器。
https://medium.com/@lmayrandprovencher/building-an-autoencoder-with-tid-weights-in-keras-c4a559c529a2
class DenseTranspose(keras.layers.Layer):
def __init__(self,dense,activation=None,**kwargs):
self.dense = dense
self.activation = keras.activations.get(activation)
super().__init__(**kwargs)
def build(self, batch_input_shape):
self.biases = self.add_weight(name='bias',shape=[self.dense.input_shape[-1]],initializer='zero')
return super().build(batch_input_shape)
def call(self,inputs):
z = tf.matmul(inputs,self.dense.weights[0],transpose_b=True)
return self.activation(z+self.biases)
inputs = Input(shape = (28,28,1))
dense_1 = Dense(784,activation='relu')
dense_2 = Dense(392,activation='relu')
dense_3 = Dense(196,activation='relu')
#Flatting input data to feed into Dense layer.
x = Flatten()(inputs)
x = dense_1(x)
x = dense_2(x)
x = dense_3(x)
x = DenseTranspose(dense_3,activation='relu')(x)
x = DenseTranspose(dense_2,activation='relu')(x)
x = DenseTranspose(dense_1,activation='sigmoid')(x)
outputs = Reshape([28,28])(x)
但当试图编译代码时，它会显示
Traceback (most recent call last):
File "<string>", line 1, in <module>
File ".ProgramsPythonPython37libsite-packageskerasenginebase_layer.py", line 474, in __call__
output_shape = self.compute_output_shape(input_shape)
File ".ProgramsPythonPython37libsite-packageskeraslayerscore.py", line 398, in
compute_output_shape
input_shape[1:], self.target_shape)
File ".ProgramsPythonPython37libsite-packageskeraslayerscore.py", line 386, in _fix_unknown_dimension
raise ValueError(msg)
ValueError: total size of new array must be unchanged
当在行outputs = Rehsape([28,28])(x)之前打印x的值时，如下所示。
<tf.Tensor 'dense_transpose_3/Sigmoid:0' shape=(?, 784) dtype=float32>
既然784=28*28，为什么我不能像他们在博客中那样将其重塑为(28,28(？
我不完全理解这个问题，但我建议用Lambda层切换Reshape
outputs = Lambda(lambda x: K.reshape(x, (28, 28)))(x)
参见此处