以下是一段代码，用于基于一些简单的块来设计Keras深Model()：
def conv_bn_activation(x, filters, kernel_size, strides, data_format, is_training, activation=relu,
name=None):
"""
Parameters
----------
x
filters
kernel_size
strides
data_format
activation
is_training
name
Returns
-------
"""
conv = tks.Conv2D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
data_format=data_format,
activation=None,
use_bias=False,
name=name)(x)
bn = batch_norm(x=conv, is_training=is_training, data_format=data_format)
if activation is not None:
return activation(bn)
else:
return bn
def basic_block(x, filters, is_training, data_format):
"""
Basic block function
Parameters
----------
x
filters
data_format
is_training
Returns
-------
"""
conv = conv_bn_activation(x=x, filters=filters, kernel_size=3, strides=1,
is_training=is_training, data_format=data_format)
conv = conv_bn_activation(x=conv, filters=filters, kernel_size=3, strides=1,
is_training=is_training, data_format=data_format)
axis = 3 if data_format == 'channels_last' else 1
input_channels = x.get_shape().as_list()[axis]
if input_channels == filters:
short_cut = x
else:
short_cut = conv_bn_activation(x=x, filters=filters, kernel_size=1, strides=1,
is_training=is_training, data_format=data_format)
return conv + short_cut
img_input_shape = [None, None, 3]
inputs = []
img_input = Input(shape=img_input_shape)
inputs.append(img_input)
conv = conv_bn_activation(x=img_input, filters=16,
kernel_size=7, strides=1,
data_format="channels_last",
is_training=True,
name="base_layer_conv_1")
conv = conv_bn_activation(x=conv, filters=16, kernel_size=3, strides=1,
data_format="channels_last",
is_training=True, name="base_layer_conv_2")
out = basic_block(conv, 64, True, data_format="channels_last")
# outputs = [conv] this work fine
outputs = [out]
model = Model(inputs=inputs, outputs=outputs)
我收到此错误 ValueError：模型的输出张量必须是TensorFlow Layer的输出（因此保留了过去的层元数据）。找到：Tensor（“ add：0”，shape =（?,？，？，64），dtype = float32）
问题出在basic_block（）函数中，但我找不到它！
答案 0 :(得分：0)
问题出在另外的conv + short_cut
def conv_bn_activation(x, filters, kernel_size, strides, data_format, is_training, activation=relu,
name=None):
"""
Parameters
----------
x
filters
kernel_size
strides
data_format
activation
is_training
name
Returns
-------
"""
conv = tks.Conv2D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
data_format=data_format,
activation=None,
use_bias=False,
name=name)(x)
bn = batch_norm(x=conv, is_training=is_training, data_format=data_format)
if activation is not None:
return activation(bn)
else:
return bn
def basic_block(x, filters, is_training, data_format):
"""
Basic block function
Parameters
----------
x
filters
data_format
is_training
Returns
-------
"""
conv = conv_bn_activation(x=x, filters=filters, kernel_size=3, strides=1,
is_training=is_training, data_format=data_format)
conv = conv_bn_activation(x=conv, filters=filters, kernel_size=3, strides=1,
is_training=is_training, data_format=data_format)
axis = 3 if data_format == 'channels_last' else 1
input_channels = x.get_shape().as_list()[axis]
if input_channels == filters:
short_cut = x
else:
short_cut = conv_bn_activation(x=x, filters=filters, kernel_size=1, strides=1,
is_training=is_training, data_format=data_format)
return tks.Add()([conv, short_cut])
img_input_shape = [None, None, 3]
inputs = []
img_input = Input(shape=img_input_shape)
inputs.append(img_input)
conv = conv_bn_activation(x=img_input, filters=16,
kernel_size=7, strides=1,
data_format="channels_last",
is_training=True,
name="base_layer_conv_1")
conv = conv_bn_activation(x=conv, filters=16, kernel_size=3, strides=1,
data_format="channels_last",
is_training=True, name="base_layer_conv_2")
out = basic_block(conv, 64, True, data_format="channels_last")
# outputs = [conv] this work fine
outputs = [out]
model = Model(inputs=inputs, outputs=outputs)