我正在尝试实现一个简化的模块，例如具有扩展卷积的wavenet。下面是一个简单的示例：
import tensorflow as tf
tfkl = tf.keras.layers
output_dim = 3
def waveres(inpt, n_filters, kernel_size, i):
tanh_out = tfkl.Conv1D(n_filters,
kernel_size,
dilation_rate = kernel_size ** i,
padding='causal',
name=f'dilated_conv_{kernel_size**i}_tanh',
activation='tanh'
)(inpt)
sigm_out = tfkl.Conv1D(n_filters,
kernel_size,
dilation_rate = kernel_size**i,
padding='causal',
name=f'dilated_conv_{kernel_size**i}_sigm',
activation='sigmoid'
)(inpt)
z = tfkl.Multiply(name=f'gated_activation_{i}')([tanh_out, sigm_out])
skip = tfkl.Conv1D(n_filters, 1, name=f'skip_{i}')(z)
res = tfkl.Add(name=f'residual_block_{i}')([skip, inpt])
return res, skip
def wavenet(inpt,depth,n_filters=32,kernel_size=2):
skip_connections = []
out = tfkl.Conv1D(n_filters, kernel_size, dilation_rate=1, activation='linear',padding='causal', name='wavenet_conv_1',input_shape=inpt.get_shape())(inpt)
for i in range(1, depth + 1):
out, skip = waveres(out, n_filters, kernel_size, i)
skip_connections.append(skip)
out = tfkl.Add(name='skip_connections')(skip_connections)
out = tfkl.Activation('relu')(out)
out = tfkl.Conv1D(n_filters, 32, strides=1, padding='causal', name='wavenet_final_conv', activation='relu')(out)
out = tfkl.AveragePooling1D(7, 1, padding='same', name='wavenet_avgpool')(out)
return out
def _model(inputs,wave_depth=4):
x = tfkl.Dense(256)(inputs)
kyma = wavenet(inputs,wave_depth)
junc = tfkl.Concatenate()([x,kyma])
fc = tfkl.Dense(32)(junc)
out = tfkl.Dense(output_dim)(fc)
return out
model = tf.keras.Model(inpt_,_model(inpt_))
我面临的问题是，在扩大的卷积之后，第二维变为无，这将阻止与平行层的连接。
我有两个问题：
这是预期的行为吗？
如何处理膨胀卷积层的输出？
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-29-3bcd4eb984b7> in <module>
49
50
---> 51 model = tf.keras.Model(inpt_,_model(inpt_))
<ipython-input-29-3bcd4eb984b7> in _model(inputs, wave_depth)
43 kyma = wavenet(inputs,wave_depth)
44 print(kyma.get_shape())
---> 45 junc = tfkl.Concatenate()([x,kyma])
46 fc = tfkl.Dense(32)(junc)
47 out = tfkl.Dense(output_dim)(fc)
~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\base_layer.py in __call__(self, inputs, *args, **kwargs)
614 # Build layer if applicable (if the `build` method has been
615 # overridden).
--> 616 self._maybe_build(inputs)
617
618 # Wrapping `call` function in autograph to allow for dynamic control
~\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\base_layer.py in _maybe_build(self, inputs)
1964 # operations.
1965 with tf_utils.maybe_init_scope(self):
-> 1966 self.build(input_shapes)
1967 # We must set self.built since user defined build functions are not
1968 # constrained to set self.built.
~\Anaconda3\lib\site-packages\tensorflow\python\keras\utils\tf_utils.py in wrapper(instance, input_shape)
294 if input_shape is not None:
295 input_shape = convert_shapes(input_shape, to_tuples=True)
--> 296 output_shape = fn(instance, input_shape)
297 # Return shapes from `fn` as TensorShapes.
298 if output_shape is not None:
~\Anaconda3\lib\site-packages\tensorflow\python\keras\layers\merge.py in build(self, input_shape)
389 'inputs with matching shapes '
390 'except for the concat axis. '
--> 391 'Got inputs shapes: %s' % (input_shape))
392
393 def _merge_function(self, inputs):
ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 10, 256), (None, None, 32)]
NB：
在Tensorflow 2.0.0beta1上运行此