我试验了一种基于https://github.com/keras-team/keras/blob/master/examples/variational_autoencoder.py在Keras中实现的变分自动编码器。
在从Keras中的非顺序计算图中获取层的输出方面存在问题，例如在Keras中实现的变分自动编码器。
编码器如下：
inputs = Input(shape=(input_shape,), name='encoder_input')
x = Dense(intermediate_dim, activation='relu')(inputs)
z_mean = Dense(latent_dim, name='z_mean')(x)
z_log_var = Dense(latent_dim, name='z_log_var')(x)
z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])
encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')
解码器如下：
latent_inputs = Input(shape=(latent_dim,), name='z_sampling')
x = Dense(intermediate_dim, activation='relu')(latent_inputs)
outputs = Dense(original_dim, activation='sigmoid', name='decoder_outpu')(x)
# instantiate decoder model
decoder = Model(latent_inputs, outputs, name='decoder')
变分自动编码器（VAE）模型如下：
# instantiate VAE model
outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs, name='vae_mlp')
为简单起见，我刚刚做了以下事情：
vae.compile(optimizer='adam', loss='binary_crossentropy')
vae.fit(x=x_train, y=x_train,epochs=epochs, batch_size=batch_size, validation_data=(x_test, x_test))
我可以得到解码器的输出，例如，如下：
decoder.summary()
intermediate_layer_outputs = [layer.output for layer in decoder.layers]
intermediate_layer_model = Mod el(inputs=decoder.get_input_at(0),outputs=intermediate_layer_outputs)
decoder_input = np.random.normal(0,10,size=(3,2))
intermediate_output = intermediate_layer_model.predict(x=decoder_input)
intermediate_output
图层的输出是：
[array([[ 2.74899244, 3.58746958],
[ -3.37609863, 3.22382379],
[ 6.70482016, 13.96735668]], dtype=float32),
array([[ 0. , 0.64544445, 0.45420599, ..., 0. ,
0.72464317, 1.18030334],
[ 1.79546463, 0. , 1.65797734, ..., 1.93947244,
0. , 0.15902017],
[ 0. , 1.51342845, 1.34246469, ..., 0. ,
2.08303833, 3.98163915]], dtype=float32),
array([[ 4.30661340e-11, 4.72995476e-10, 1.68229250e-10, ...,
4.98356023e-10, 6.42546918e-11, 2.90629031e-11],
[ 9.11776947e-13, 4.95661184e-13, 4.81620494e-13, ...,
2.19281014e-13, 1.79197141e-12, 6.77271390e-13],
[ 1.75418052e-27, 1.18550066e-24, 1.44973726e-25, ...,
2.22284772e-24, 5.12405949e-27, 7.81602653e-28]], dtype=float32)]
VAE摘要如下：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
encoder_input (InputLayer) (None, 784) 0
_________________________________________________________________
encoder (Model) [(None, 2), (None, 2), (N 403972
_________________________________________________________________
decoder (Model) (None, 784) 403728
=================================================================
从上面的摘要中我可以得到中间层的输出，即编码器（模型）层，如下所示：
intermediate_layer_outputs = [layer.get_output_at(0) for layer in vae.layers]
intermediate_layer_model = Model(inputs=vae.get_input_at(0), outputs=intermediate_layer_outputs[1])
intermediate_output_1 = intermediate_layer_model.predict(x=x_train[:1])
print('Layer 1 contents:')
display(intermediate_output_1)
，输出结果为：
Layer 1 contents:
[array([[ 0.22815476, -0.59258038]], dtype=float32),
array([[-4.97396278, -4.52285957]], dtype=float32),
array([[ 0.13466805, -0.67670715]], dtype=float32)]
以下是我的问题： 我想得到最后一层的输出，其中输入来自编码器的第三个输出。但是，我收到了错误。这是我尝试过的一个例子：
intermediate_layer_outputs = [layer.get_output_at(0) for layer in vae.layers]
intermediate_layer_model = Model(inputs=encoder(Input(shape=(input_shape,)))[2], outputs=intermediate_layer_outputs[2])
intermediate_layer_outputs
错误是
/opt/conda/lib/python3.5/site-packages/keras/engine/topology.py in __init__(self, inputs, outputs, name)
1608 # It's supposed to be an input layer, so only one node
1609 # and one tensor output.
-> 1610 assert node_index == 0
1611 assert tensor_index == 0
1612 self.input_layers.append(layer)
AssertionError:
如果我用
替换中间线intermediate_layer_model = Model(inputs=vae.layers[1].get_output_at(0)[2], outputs=intermediate_layer_outputs[2])
我收到以下错误
TypeError: Input layers to a `Model` must be `InputLayer` objects. Received inputs: Tensor("z/add:0", shape=(?, 2), dtype=float32). Input 0 (0-based) originates from layer type `Lambda`.
如果我尝试以下
intermediate_layer_model = Model(inputs=Input(shape=(latent_dim,)), outputs=intermediate_layer_outputs[2])
我收到以下错误
RuntimeError: Graph disconnected: cannot obtain value for tensor Tensor("z_sampling:0", shape=(?, 2), dtype=float32) at layer "z_sampling". The following previous layers were accessed without issue: []
我的感觉是它与计算图的非顺序性有关。我真的很感谢这方面的帮助！
非常感谢！