请注意：此问题在github上也以issue形式存在。
更新
我试图在Keras中实现一个自定义图层：一个通过卷积计算其门的GRU图层。 该代码有效，但仅限于Theano后端。
在github上，我得到了提示切换＆＃34;图像布局＆＃34;到＆＃34; tf＆＃34;在keras.json。 事实上，我已经硬编了“＃34; th＆＃34;图像布局，所以我不得不修复我的代码的一些问题。 现在它支持＆＃34; th＆＃34;和＆＃34; tf＆＃34;布局。 该模型也进行了调整：现在我有两个模型，一个用于＆＃34; th＆＃34;和＃34; tf＆＃34;。
这是对问题的完全重写。问题是一样的，但代码是新的。
以theano作为后端，两个图像布局都运行良好。网络能够从数据中学习，新层似乎有效。我非常肯定我实施了第二次和第二次操作。和＆＃34; tf＆＃34;布局是正确的。 但是当我切换到tensorflow时，它总是会崩溃。布局似乎没有效果。
代码中有一个技巧：GRU使用卷积来计算其门。但该卷积不包含在图层中。我只是假设每次使用这个图层时，都会有一个2D卷积，它产生3个输出特征（每个门一个）。 这是代码：
import numpy as np
from keras import backend as K
from keras import initializations, activations
from keras.engine import InputSpec
from keras.layers.recurrent import Recurrent
# copied from the keras github source. removed lots of unnecessary (for me) code
# assuming a 2D Convolution was run by hand before this layer.
# please note that this has no variables of its own.
# TODO: incorporate the 2D Convolution into this layer
class CGRU(Recurrent):
def __init__(self,
init='glorot_uniform', inner_init='orthogonal',
activation='tanh', inner_activation='hard_sigmoid', **kwargs):
self.init = initializations.get(init)
self.inner_init = initializations.get(inner_init)
self.activation = activations.get(activation)
self.inner_activation = activations.get(inner_activation)
self.dim_ordering = K.image_dim_ordering()
assert self.dim_ordering in {'tf', 'th'}, '`dim_ordering` must be in {"tf", "th"}.'
# removing the regularizers and the dropout
super(CGRU, self).__init__(**kwargs)
# this seems necessary in order to accept 5 input dimensions
# (samples, timesteps, features, x, y)
self.input_spec = [InputSpec(ndim=5)]
def build(self, input_shape):
self.input_spec = [InputSpec(shape=input_shape)]
if self.dim_ordering == "th":
nb_features = input_shape[2]
x_dim = input_shape[3]
y_dim = input_shape[4]
# the input_dim was broken and didn't cause any problems. I'm commenting it out to see
# whether it is really necessary
# self.input_dim = (3,x_dim, y_dim)
# moved here from the constructor. Curiously it seems like batch_size has been removed from here.
# I also don't understand why I don't need to specify the time_sequence if there is the flag
# output_sequence is set to True
# TODO: debug where this is used
self.output_dim = [1, x_dim, y_dim]
elif self.dim_ordering == "tf":
nb_features = input_shape[4]
x_dim = input_shape[2]
y_dim = input_shape[3]
# see above
# self.input_dim = (x_dim, y_dim, 3)
# see above
self.output_dim = [x_dim, y_dim, 1]
assert (nb_features == 3, "the CGRU expects 3 input features")
if self.stateful:
self.reset_states()
else:
# initial states: all-zero tensor of shape (output_dim)
self.states = [None]
def reset_states(self):
assert self.stateful, 'Layer must be stateful.'
input_shape = self.input_spec[0].shape
if not input_shape[0]:
raise Exception('If a RNN is stateful, a complete ' +
'input_shape must be provided (including batch size).')
if self.dim_ordering == "th":
nb_features = input_shape[2]
x_dim = input_shape[3]
y_dim = input_shape[4]
elif self.dim_ordering == "tf":
nb_features = input_shape[4]
x_dim = input_shape[2]
y_dim = input_shape[3]
if hasattr(self, 'states'):
K.set_value(self.states[0],
np.zeros((input_shape[0], x_dim, y_dim)))
else:
self.states = [K.zeros((input_shape[0], x_dim, y_dim))]
def preprocess_input(self, x):
# TODO: if at some point, this layer does calculations, revisit and optimise this code
return x
def step(self, x, states):
h_tm1 = states[0] # previous memory
# usually here would be matrix multiplications for the different gates
# I assume that a convolution before this layer has prepared 3 features per coordinate
# I'll simply use those 3 features
# in fact: this layer doesn't have any trainable weights
# TODO: is an if else problematic wrt performance.
# TODO: since I'm currently trying to get the layer to work with tensorflow, I'll now hardcode for tf
if self.dim_ordering == "th":
z = self.inner_activation(x[:, 0, :, :])
r = self.inner_activation(x[:, 1, :, :])
hh = self.activation(x[:, 2, :, :])
else: # self.dim_ordering == "th":
z = self.inner_activation(x[:, :, :, 0])
r = self.inner_activation(x[:, :, :, 1])
hh = self.activation(x[:, :, :, 2])
h = z * h_tm1 + (1 - z) * hh
return h, [h]
def get_initial_states(self, x):
initial_state = K.zeros_like(x) # (samples, timesteps, input_dim)
# input_dim = (3, x_dim, y_dim) for th
# input_dim = (x_dim, y_dim, 3) for tf
# please note: the number of features is expected to be 3, one feature for each gate
if self.dim_ordering == "th":
initial_state = K.sum(initial_state, axis=(1, 2)) # (samples, x_dim, y_dim)
else: # "tf"
initial_state = K.sum(initial_state, axis=(1, -1)) # (samples, x_dim, y_dim)
return [initial_state]
def get_output_shape_for(self, input_shape):
if self.return_sequences:
return tuple([input_shape[0], input_shape[1]] + self.output_dim)
else:
return tuple([input_shape[0]] + self.output_dim)
def get_config(self):
config = {'output_dim': self.output_dim,
'init': self.init.__name__,
'inner_init': self.inner_init.__name__,
'activation': self.activation.__name__,
'inner_activation': self.inner_activation.__name__}
# removed the various regularizers and dropouts.
# surely this isn't needed if not present ?
base_config = super(CGRU, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
很抱歉为您提供了如此多的代码，我已尽力对其进行评论和格式化。
这里是设置，这就是我使用代码的方式：
if K.image_dim_ordering()=="tf":
# this is the actual input, fed to the network
inputs = Input((40, 40, 40, 1))
# now reshape to a sequence
#reshaped = Reshape((40, 1, 40, 40))(inputs)
# now create a model like uNet which is fed the time-distributed data
# this model will perform a segmentation on a 2D slice of data.
conv_inputs = Input((40, 40, 1))
base_filters_size = 16
conv1 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(conv_inputs)
#conv1 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv1)
conv1 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(conv1)
#conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv1)
conv1 = Convolution2D(3, 3, 3, activation='relu', border_mode='same')(conv1)
convmodel = Model(input=conv_inputs, output=conv1)
convmodel.summary()
#apply the segmentation to each layer
time_dist=TimeDistributed(convmodel)(inputs)
# now a model with the two GRUs which move in different directions
time_dist=Reshape((40,40,40, 3))(time_dist)
# The next line will crash
up=CGRU(go_backwards=False, return_sequences=True, name="up", input_shape=[40,40,40,3])(time_dist)
down=CGRU(go_backwards=True, return_sequences=True, name="down", input_shape=[40,40,40,3])(time_dist)
# merge the two outputs
merged_updown=merge([up,down], mode="concat", concat_axis=2)
reshaped=Reshape((40,40,40,2))(merged_updown)
# and apply a final convolution to map from 2 features back to 1
mergeconv_inputs=Input((40,40,2))
mergeconv=Convolution2D(1,3,3, activation="sigmoid", border_mode="same")(mergeconv_inputs)
merge_model=Model(input=mergeconv_inputs, output=mergeconv)
final_time_dist=TimeDistributed(merge_model)(reshaped)
output=Reshape((40,40,40,1))(final_time_dist)
model=Model(input=inputs, output=output)
print(model.summary())
model.compile(optimizer=Adam(lr=1e-4), loss=dice_coef_loss, metrics=[dice_coef])
完全相同的代码，使用＆＃34; tf＆＃34;图像布局，在theano下运行正常，并且有张量流崩溃。
问题是：ValueError: Shapes (?, ?, 40, 40) and (40, ?, 40) are not compatible
以下是完整的错误消息：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-8-295ff1bf414f> in <module>()
25 # now a model with the two GRUs which move in different directions
26 time_dist=Reshape((40,40,40, 3))(time_dist)
---> 27 up=CGRU(go_backwards=False, return_sequences=True, name="up", input_shape=[40,40,40,3])(time_dist)
28 down=CGRU(go_backwards=True, return_sequences=True, name="down", input_shape=[40,40,40,3])(time_dist)
29
/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in __call__(self, x, mask)
513 if inbound_layers:
514 # this will call layer.build() if necessary
--> 515 self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
516 input_added = True
517
/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in add_inbound_node(self, inbound_layers, node_indices, tensor_indices)
571 # creating the node automatically updates self.inbound_nodes
572 # as well as outbound_nodes on inbound layers.
--> 573 Node.create_node(self, inbound_layers, node_indices, tensor_indices)
574
575 def get_output_shape_for(self, input_shape):
/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in create_node(cls, outbound_layer, inbound_layers, node_indices, tensor_indices)
148
149 if len(input_tensors) == 1:
--> 150 output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0]))
151 output_masks = to_list(outbound_layer.compute_mask(input_tensors[0], input_masks[0]))
152 # TODO: try to auto-infer shape if exception is raised by get_output_shape_for
/usr/local/lib/python2.7/dist-packages/keras/layers/recurrent.pyc in call(self, x, mask)
211 constants=constants,
212 unroll=self.unroll,
--> 213 input_length=input_shape[1])
214 if self.stateful:
215 self.updates = []
/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.pyc in rnn(step_function, inputs, initial_states, go_backwards, mask, constants, unroll, input_length)
1193 parallel_iterations=32,
1194 swap_memory=True,
-> 1195 sequence_length=None)
1196
1197 if nb_states > 1:
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.pyc in _dynamic_rnn_loop(cell, inputs, initial_state, parallel_iterations, swap_memory, sequence_length, dtype)
1023 shape = _state_size_with_prefix(
1024 output_size, prefix=[const_time_steps, const_batch_size])
-> 1025 output.set_shape(shape)
1026
1027 final_outputs = nest.pack_sequence_as(
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc in set_shape(self, shape)
406 this tensor.
407 """
--> 408 self._shape = self._shape.merge_with(shape)
409
410 @property
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_shape.pyc in merge_with(self, other)
568 except ValueError:
569 raise ValueError("Shapes %s and %s are not compatible" %
--> 570 (self, other))
571
572 def concatenate(self, other):
ValueError: Shapes (?, ?, 40, 40) and (40, ?, 40) are not compatible