我正在尝试使用keras实现标准偏差合并层。这个想法类似于用功能类似于AveragePooling1D的层来实现，但是要计算标准偏差。
我的第一步是尝试将其实现为Lambda层。它应该使用3d张量，例如（batch_size，time，features）和一个跨度整数（指示窗口的大小）。它应该返回一个形状为（batch_size，time，features）的张量。
我的实现如下：
import tensorflow
import keras
from keras.layers import Dense, TimeDistributed, Lambda, Input
import numpy as np
import keras.backend as K
def stdev_pooling(inputs):
data, stride = inputs
stride = K.cast(stride, dtype='int32')
print K.dtype(stride), K.dtype(data), '---'
num_windows = K.shape(data)[1] / stride
idxs = K.arange(num_windows) * stride
windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())
windows = K.permute_dimensions(windows, (1,0,2,3))
stds = K.map_fn(lambda w: K.std(w, axis=1), windows)
return stds
ipt = Input(shape=(None,10))
d = TimeDistributed(Dense(10))(ipt)
out = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])
m = keras.Model(inputs=ipt, outputs=out)
x = np.arange(1000).reshape(1,-1,10)
m.predict(x).shape
但是，我的输出（按顺序显示了步幅和数据张量的数据类型）是
int32 float32 ---
float32 float32 ---
堆栈跟踪是这样的：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
in ()
1 ipt = Input(shape=(None,10))
2 d = TimeDistributed(Dense(10))(ipt)
----> 3 out = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/base_layer.pyc in __call__(self, inputs, **kwargs)
472 if all([s is not None
473 for s in to_list(input_shape)]):
--> 474 output_shape = self.compute_output_shape(input_shape)
475 else:
476 if isinstance(input_shape, list):
/home/juliano/.local/lib/python2.7/site-packages/keras/layers/core.pyc in compute_output_shape(self, input_shape)
643 if isinstance(input_shape, list):
644 xs = [K.placeholder(shape=shape) for shape in input_shape]
--> 645 x = self.call(xs)
646 else:
647 x = K.placeholder(shape=input_shape)
/home/juliano/.local/lib/python2.7/site-packages/keras/layers/core.pyc in call(self, inputs, mask)
680 if has_arg(self.function, 'mask'):
681 arguments['mask'] = mask
--> 682 return self.function(inputs, **arguments)
683
684 def compute_mask(self, inputs, mask=None):
in stdev_pooling(inputs)
5 print K.dtype(stride), K.dtype(data), '---'
6
----> 7 num_windows = K.shape(data)[1] / stride
8
9 idxs = K.arange(num_windows-1) * stride
/home/juliano/.local/lib/python2.7/site-packages/tensorflow/python/ops/math_ops.pyc in binary_op_wrapper(x, y)
848 with ops.name_scope(None, op_name, [x, y]) as name:
849 if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):
--> 850 return func(x, y, name=name)
851 elif not isinstance(y, sparse_tensor.SparseTensor):
852 try:
/home/juliano/.local/lib/python2.7/site-packages/tensorflow/python/ops/math_ops.pyc in _div_python2(x, y, name)
972 with ops.name_scope(name, "div", [x, y]) as name:
973 x = ops.convert_to_tensor(x, name="x")
--> 974 y = ops.convert_to_tensor(y, name="y", dtype=x.dtype.base_dtype)
975 x_dtype = x.dtype.base_dtype
976 y_dtype = y.dtype.base_dtype
/home/juliano/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in convert_to_tensor(value, dtype, name, preferred_dtype)
996 name=name,
997 preferred_dtype=preferred_dtype,
--> 998 as_ref=False)
999
1000
/home/juliano/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx)
1092
1093 if ret is None:
-> 1094 ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
1095
1096 if ret is NotImplemented:
/home/juliano/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in _TensorTensorConversionFunction(t, dtype, name, as_ref)
929 raise ValueError(
930 "Tensor conversion requested dtype %s for Tensor with dtype %s: %r" %
--> 931 (dtype.name, t.dtype.name, str(t)))
932 return t
933
ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor("lambda_9/Placeholder_1:0", shape=(), dtype=float32)'
有趣的是，据我了解，它表示stride变量是float32，尽管它被声明为int32，但应将其转换为int32。变量K.variable(20, dtype='int32', name='stride_var')。
这是怎么了？任何帮助将非常感激！谢谢！
编辑：
正如@BlackBear所建议的，我添加了一个显式的强制转换，看来已经解决了部分问题：
def stdev_pooling(inputs):
data, stride = inputs
stride = K.cast(stride, dtype='int32')
print K.dtype(stride), K.dtype(data), '---'
num_windows = K.shape(data)[1] / stride
idxs = K.arange(num_windows) * stride
windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())
windows = K.permute_dimensions(windows, (1,0,2,3))
stds = K.map_fn(lambda w: K.std(w, axis=1), windows)
return stds
输出：
int32 float32 ---
int32 float32 ---
但是，现在我有一个新错误，我不知道它来自哪里！
这是堆栈跟踪：
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
in ()
3 x = np.arange(2000).reshape(2,-1,10)
4
----> 5 m = keras.Model(inputs=ipt, outputs=out)
6
7 m.predict(x).shape
/home/juliano/.local/lib/python2.7/site-packages/keras/legacy/interfaces.pyc in wrapper(*args, **kwargs)
89 warnings.warn('Update your `' + object_name +
90 '` call to the Keras 2 API: ' + signature, stacklevel=2)
---> 91 return func(*args, **kwargs)
92 wrapper._original_function = func
93 return wrapper
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/network.pyc in __init__(self, *args, **kwargs)
91 'inputs' in kwargs and 'outputs' in kwargs):
92 # Graph network
---> 93 self._init_graph_network(*args, **kwargs)
94 else:
95 # Subclassed network
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/network.pyc in _init_graph_network(self, inputs, outputs, name)
235 # Keep track of the network's nodes and layers.
236 nodes, nodes_by_depth, layers, layers_by_depth = _map_graph_network(
--> 237 self.inputs, self.outputs)
238 self._network_nodes = nodes
239 self._nodes_by_depth = nodes_by_depth
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/network.pyc in _map_graph_network(inputs, outputs)
1351 layer=layer,
1352 node_index=node_index,
-> 1353 tensor_index=tensor_index)
1354
1355 for node in reversed(nodes_in_decreasing_depth):
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/network.pyc in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index)
1338 tensor_index = node.tensor_indices[i]
1339 build_map(x, finished_nodes, nodes_in_progress, layer,
-> 1340 node_index, tensor_index)
1341
1342 finished_nodes.add(node)
/home/juliano/.local/lib/python2.7/site-packages/keras/engine/network.pyc in build_map(tensor, finished_nodes, nodes_in_progress, layer, node_index, tensor_index)
1310 ValueError: if a cycle is detected.
1311 """
-> 1312 node = layer._inbound_nodes[node_index]
1313
1314 # Prevent cycles.
AttributeError: 'NoneType' object has no attribute '_inbound_nodes'
编辑：我已经更新了stdev_pooling函数，它现在返回正确的输出。但是，我仍然遇到AttributeError: 'NoneType' object has no attribute '_inbound_nodes'错误...
答案 0 :(得分：1)
在对代码进行了更多摆弄并阅读了keras如何与tensorflow交互之后（在许多不同的地方，包括tensorflow和keras的源代码），我发现了哪里出了问题。
首先，这是我想要做的一个最小的工作示例：
import tensorflow
import keras
from keras.layers import Dense, TimeDistributed, Lambda, Input
import numpy as np
import keras.backend as K
def stdev_pooling(inputs, stride):
data = inputs
padding = K.shape(data)[1] % stride
data = K.switch(padding > 0, K.temporal_padding(data, padding=(0,stride-padding)), data )
num_windows = K.shape(data)[1] / stride
idxs = K.arange(num_windows) * stride
windows = K.map_fn(lambda w: data[:, w: (w + stride), :], idxs, dtype=K.floatx())
windows = K.permute_dimensions(windows, (1,0,2,3))
stds = K.map_fn(lambda w: K.std(w, axis=1), windows)
return stds
ipt = Input(shape=(None,10))
d = TimeDistributed(Dense(10))(ipt)
#stride is an argument to stdev_pooling, not a signal coming from
#a previous layer. Thus it must be passed in the `arguments`
#dictionary of the `Lambda` layer.
out = Lambda(stdev_pooling, arguments={'stride':15})(d)
x = np.arange(2000).reshape(2,-1,10)
m = keras.Model(inputs=ipt, outputs=out)
y = m.predict(x)
print y
print y.shape
问题出在先前代码中的out = Lambda(stdev_pooling)([d,K.variable(20, dtype='int32', name='stride_var')])行。
当提供信号（例如K.variable（...））作为Lambda层的输入时，keras希望将其连接到Input层。因此，错误AttributeError: 'NoneType' object has no attribute '_inbound_nodes'。
解决方案只是通过Lambda层构造函数的arguments字典提供stride参数：
out = Lambda(stdev_pooling, arguments={'stride':15})(d)
我希望这段代码可以帮助所有人尝试在keras中构建某种池化层。有空的时候，我会将其写为适当的池化层。目前，此Lambda版本应该可以。