我正在将此工具（ann4brains）从Caffe转换为Keras。
我已经实现了两种自定义类型的2D卷积（E2E和E2N）。 我是根据_Conv的源代码（来自Keras的源代码）实现的。1
模型已编译，但在拟合过程中失败，并显示以下消息错误：
ValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.
我在Google上搜索了以下主题：Custom Keras Layer Troubles
解决方案（删除self.kernel = self.add_weight(...)适用于我的情况。但是我对此解决方案并不安全。如果它在_Conv类中，为什么还要删除它？发表评论吗？还有其他推荐的解决方案吗？
谢谢！
以下是有关此案的更多信息：
Keras版本： 2.2.4
输入形状： （21、21、1）
模型摘要：
Layer (type) Output Shape Param #
=================================================================
conv_e2e_20 (ConvE2E) (None, 21, 21, 1, 32) 1376
_________________________________________________________________
conv_e2n_16 (ConvE2N) (None, 21, 1, 1, 64) 43072
_________________________________________________________________
flatten_28 (Flatten) (None, 1344) 0
_________________________________________________________________
dense_91 (Dense) (None, 128) 172160
_________________________________________________________________
dropout_84 (Dropout) (None, 128) 0
_________________________________________________________________
dense_92 (Dense) (None, 30) 3870
_________________________________________________________________
dropout_85 (Dropout) (None, 30) 0
_________________________________________________________________
dense_93 (Dense) (None, 30) 930
_________________________________________________________________
dropout_86 (Dropout) (None, 30) 0
_________________________________________________________________
dense_94 (Dense) (None, 2) 62
=================================================================
Total params: 221,470
Trainable params: 221,470
Non-trainable params: 0
_________________________________________________________________
端到端层
class ConvE2E(Layer):
def __init__(self,
filters,
kernel_size,
strides=1,
padding='valid',
data_format=None,
dilation_rate=1,
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super(ConvE2E, self).__init__(**kwargs)
self.rank = 2
self.filters = filters
self.kernel_size = kernel_size
self.strides = conv_utils.normalize_tuple(strides, self.rank, 'strides')
self.padding = conv_utils.normalize_padding(padding)
self.data_format = K.common.normalize_data_format(data_format)
self.dilation_rate = conv_utils.normalize_tuple(dilation_rate, self.rank,
'dilation_rate')
self.activation = activations.get(activation)
self.use_bias = use_bias
self.kernel_initializer = initializers.get(kernel_initializer)
self.bias_initializer = initializers.get(bias_initializer)
self.kernel_regularizer = regularizers.get(kernel_regularizer)
self.bias_regularizer = regularizers.get(bias_regularizer)
self.activity_regularizer = regularizers.get(activity_regularizer)
self.kernel_constraint = constraints.get(kernel_constraint)
self.bias_constraint = constraints.get(bias_constraint)
self.input_spec = InputSpec(ndim=self.rank + 2)
def build(self, input_shape):
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape[channel_axis] is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
self.input_dim = input_shape[channel_axis]
kernel_shape = self.kernel_size + (self.input_dim, self.filters)
# self.kernel = self.add_weight(shape=kernel_shape,
# initializer=self.kernel_initializer,
# name='kernel',
# regularizer=self.kernel_regularizer,
# constraint=self.kernel_constraint)
if self.use_bias:
self.bias = self.add_weight(shape=(self.filters,),
initializer=self.bias_initializer,
name='bias',
regularizer=self.bias_regularizer,
constraint=self.bias_constraint)
else:
self.bias = None
# Set input spec.
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: self.input_dim})
self.built = True
def call(self, inputs):
kernel_h = self.kernel_size[0]
kernel_w = self.kernel_size[1]
kernel_size_h = conv_utils.normalize_tuple((kernel_h, 1), 2, 'kernel_size')
kernel_shape_h = kernel_size_h + (self.input_dim, self.filters)
kernel_dx1 = self.add_weight(shape=kernel_shape_h,
initializer=self.kernel_initializer,
name='kernel',
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
kernel_size_w = conv_utils.normalize_tuple((1, kernel_w), 2, 'kernel_size')
kernel_shape_w = kernel_size_w + (self.input_dim, self.filters)
kernel_1xd = self.add_weight(shape=kernel_shape_w,
initializer=self.kernel_initializer,
name='kernel',
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
outputs_dx1 = K.conv2d(inputs, kernel_dx1,strides=self.strides,
padding=self.padding, data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs_dx1_dxd = K.repeat_elements(outputs_dx1, kernel_w, 1)
outputs_1xd = K.conv2d(inputs, kernel_1xd, strides=self.strides,
padding=self.padding, data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs_1xd_dxd = K.repeat_elements(outputs_1xd, kernel_h, 2)
outputs = Add()([outputs_dx1_dxd, outputs_1xd_dxd])
if self.use_bias:
outputs = K.bias_add(
outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def compute_output_shape(self, input_shape):
if self.data_format == 'channels_last':
output_shape = (input_shape) + (self.filters,)
return output_shape
if self.data_format == 'channels_first':
output_shape = (input_shape[0], self.filters) + (input_shape[1:])
return output_shape
def get_config(self):
config = {
'rank': self.rank,
'filters': self.filters,
'kernel_size': self.kernel_size,
'strides': self.strides,
'padding': self.padding,
'data_format': self.data_format,
'dilation_rate': self.dilation_rate,
'activation': activations.serialize(self.activation),
'use_bias': self.use_bias,
'kernel_initializer': initializers.serialize(self.kernel_initializer),
'bias_initializer': initializers.serialize(self.bias_initializer),
'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
'bias_regularizer': regularizers.serialize(self.bias_regularizer),
'activity_regularizer':
regularizers.serialize(self.activity_regularizer),
'kernel_constraint': constraints.serialize(self.kernel_constraint),
'bias_constraint': constraints.serialize(self.bias_constraint)
}
base_config = super(ConvE2E, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
E2N层
class ConvE2N(Layer):
def __init__(self,
filters,
kernel_size,
strides=1,
padding='valid',
data_format=None,
dilation_rate=1,
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super(ConvE2N, self).__init__(**kwargs)
self.rank = 2
self.filters = filters
self.kernel_size = kernel_size
self.strides = conv_utils.normalize_tuple(strides, self.rank, 'strides')
self.padding = conv_utils.normalize_padding(padding)
self.data_format = K.common.normalize_data_format(data_format)
self.dilation_rate = conv_utils.normalize_tuple(dilation_rate, self.rank,
'dilation_rate')
self.activation = activations.get(activation)
self.use_bias = use_bias
self.kernel_initializer = initializers.get(kernel_initializer)
self.bias_initializer = initializers.get(bias_initializer)
self.kernel_regularizer = regularizers.get(kernel_regularizer)
self.bias_regularizer = regularizers.get(bias_regularizer)
self.activity_regularizer = regularizers.get(activity_regularizer)
self.kernel_constraint = constraints.get(kernel_constraint)
self.bias_constraint = constraints.get(bias_constraint)
self.input_spec = InputSpec(ndim=self.rank + 2)
def build(self, input_shape):
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape[channel_axis] is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
self.input_dim = input_shape[channel_axis]
kernel_shape = self.kernel_size + (self.input_dim, self.filters)
# self.kernel = self.add_weight(shape=kernel_shape,
# initializer=self.kernel_initializer,
# name='kernel',
# regularizer=self.kernel_regularizer,
# constraint=self.kernel_constraint)
if self.use_bias:
self.bias = self.add_weight(shape=(self.filters,),
initializer=self.bias_initializer,
name='bias',
regularizer=self.bias_regularizer,
constraint=self.bias_constraint)
else:
self.bias = None
# Set input spec.
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: self.input_dim})
self.built = True
def call(self, inputs):
kernel_h = self.kernel_size[0]
kernel_w = self.kernel_size[1]
kernel_size_w = conv_utils.normalize_tuple((1, kernel_w), 2, 'kernel_size')
kernel_shape_w = kernel_size_w + (self.input_dim, self.filters)
kernel_1xd = self.add_weight(shape=kernel_shape_w,
initializer=self.kernel_initializer,
name='kernel',
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
outputs_1xd = K.conv2d(inputs, kernel_1xd, strides=self.strides,
padding=self.padding, data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs = outputs_1xd
if self.use_bias:
outputs = K.bias_add(
outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def compute_output_shape(self, input_shape):
if self.data_format == 'channels_last':
output_shape = (input_shape[0], self.kernel_size[0], 1, input_shape[-2], self.filters)
return output_shape
if self.data_format == 'channels_first':
output_shape = input_shape[0:2] + (self.kernel_size[0], 1, self.filters)
return output_shape
def get_config(self):
config = {
'rank': self.rank,
'filters': self.filters,
'kernel_size': self.kernel_size,
'strides': self.strides,
'padding': self.padding,
'data_format': self.data_format,
'dilation_rate': self.dilation_rate,
'activation': activations.serialize(self.activation),
'use_bias': self.use_bias,
'kernel_initializer': initializers.serialize(self.kernel_initializer),
'bias_initializer': initializers.serialize(self.bias_initializer),
'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
'bias_regularizer': regularizers.serialize(self.bias_regularizer),
'activity_regularizer':
regularizers.serialize(self.activity_regularizer),
'kernel_constraint': constraints.serialize(self.kernel_constraint),
'bias_constraint': constraints.serialize(self.bias_constraint)
}
base_config = super(ConvE2N, self).get_config()
return dict(list(base_config.items()) + list(config.items()))