如何解决带有BatchNormalizationaxis = CHANNEL_AXIS的“ TypeError:不可哈希类型:'Dimension'”输入
我有以下从github获得的resnet 3D架构。这是R3D的Keras实现。该体系结构旨在训练视频分类模型
## resnet 3D architecture
# Taken from https://github.com/JihongJu/keras-resnet3d/blob/master/resnet3d/resnet3d.py
def _bn_relu(input):
"""Helper to build a BN -> relu block (by @raghakot)."""
norm = Batchnormalization(axis=CHANNEL_AXIS)(input)
return Activation("relu")(norm)
def _conv_bn_relu3D(**conv_params):
filters = conv_params["filters"]
kernel_size = conv_params["kernel_size"]
strides = conv_params.setdefault("strides",(1,1,1))
kernel_initializer = conv_params.setdefault(
"kernel_initializer","he_normal")
padding = conv_params.setdefault("padding","same")
kernel_regularizer = conv_params.setdefault("kernel_regularizer",l2(1e-4))
def f(input):
conv = Conv3D(filters=filters,kernel_size=kernel_size,strides=strides,kernel_initializer=kernel_initializer,padding=padding,kernel_regularizer=kernel_regularizer)(input)
return _bn_relu(conv)
return f
def _bn_relu_conv3d(**conv_params):
"""Helper to build a BN -> relu -> conv3d block."""
filters = conv_params["filters"]
kernel_size = conv_params["kernel_size"]
strides = conv_params.setdefault("strides",1))
kernel_initializer = conv_params.setdefault("kernel_initializer",l2(1e-4))
def f(input):
activation = _bn_relu(input)
return Conv3D(filters=filters,kernel_regularizer=kernel_regularizer)(activation)
return f
def _shortcut3d(input,residual):
"""3D shortcut to match input and residual and merges them with "sum"."""
stride_dim1 = math.ceil(int(input.shape[DIM1_AXIS]) \
/ int(residual.shape[DIM1_AXIS]))
stride_dim2 = math.ceil(int(input.shape[DIM2_AXIS]) \
/ int(residual.shape[DIM2_AXIS]))
stride_dim3 = math.ceil(int(input.shape[DIM3_AXIS]) \
/ int(residual.shape[DIM3_AXIS]))
equal_channels = int(residual.shape[CHANNEL_AXIS]) \
== int(input.shape[CHANNEL_AXIS])
shortcut = input
if stride_dim1 > 1 or stride_dim2 > 1 or stride_dim3 > 1 \
or not equal_channels:
shortcut = Conv3D(
filters=residual.shape[CHANNEL_AXIS],kernel_size=(1,1),strides=(stride_dim1,stride_dim2,stride_dim3),kernel_initializer="he_normal",padding="valid",kernel_regularizer=l2(1e-4)
)(input)
return add([shortcut,residual])
def _residual_block3d(block_function,filters,kernel_regularizer,repetitions,is_first_layer=False):
def f(input):
for i in range(repetitions):
strides = (1,1)
if i == 0 and not is_first_layer:
strides = (2,2,2)
input = block_function(filters=filters,kernel_regularizer=kernel_regularizer,is_first_block_of_first_layer=(
is_first_layer and i == 0)
)(input)
return input
return f
def basic_block(filters,strides=(1,kernel_regularizer=l2(1e-4),is_first_block_of_first_layer=False):
"""Basic 3 X 3 X 3 convolution blocks. Extended from raghakot's 2D impl."""
def f(input):
if is_first_block_of_first_layer:
# don't repeat bn->relu since we just did bn->relu->maxpool
conv1 = Conv3D(filters=filters,kernel_size=(3,3,3),padding="same",kernel_regularizer=kernel_regularizer
)(input)
else:
conv1 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
residual = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(conv1)
return _shortcut3d(input,residual)
return f
def bottleneck(filters,is_first_block_of_first_layer=False):
"""Basic 3 X 3 X 3 convolution blocks. Extended from raghakot's 2D impl."""
def f(input):
if is_first_block_of_first_layer:
# don't repeat bn->relu since we just did bn->relu->maxpool
conv_1_1 = Conv3D(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
else:
conv_1_1 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
conv_3_3 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(conv_1_1)
residual = _bn_relu_conv3d(filters=filters * 4,kernel_regularizer=kernel_regularizer
)(conv_3_3)
return _shortcut3d(input,residual)
return f
def _handle_data_format():
global DIM1_AXIS
global DIM2_AXIS
global DIM3_AXIS
global CHANNEL_AXIS
if K.image_data_format() == 'channels_last':
DIM1_AXIS = 1
DIM2_AXIS = 2
DIM3_AXIS = 3
CHANNEL_AXIS = 4
else:
CHANNEL_AXIS = 1
DIM1_AXIS = 2
DIM2_AXIS = 3
DIM3_AXIS = 4
def _get_block(identifier):
if isinstance(identifier,six.string_types):
res = globals().get(identifier)
if not res:
raise ValueError('Invalid {}'.format(identifier))
return res
return identifier
class resnet3DBuilder(object):
"""resnet3D."""
@staticmethod
def build(input_shape,num_outputs,block_fn,reg_factor):
"""Instantiate a vanilla resnet3D keras model.
# Arguments
input_shape: Tuple of input shape in the format
(conv_dim1,conv_dim2,conv_dim3,channels) if dim_ordering='tf'
(filter,conv_dim1,conv_dim3) if dim_ordering='th'
num_outputs: The number of outputs at the final softmax layer
block_fn: Unit block to use {'basic_block','bottlenack_block'}
repetitions: Repetitions of unit blocks
# Returns
model: a 3D resnet model that takes a 5D tensor (volumetric images
in batch) as input and returns a 1D vector (prediction) as output.
"""
_handle_data_format()
if len(input_shape) != 4:
raise ValueError("Input shape should be a tuple "
"(conv_dim1,channels) "
"for tensorflow as backend or "
"(channels,conv_dim3) "
"for theano as backend")
block_fn = _get_block(block_fn)
input = Input(shape=input_shape)
# first conv
conv1 = _conv_bn_relu3D(filters=64,kernel_size=(7,7,7),strides=(2,2),kernel_regularizer=l2(reg_factor)
)(input)
pool1 = MaxPooling3D(pool_size=(3,padding="same")(conv1)
# repeat blocks
block = pool1
filters = 64
for i,r in enumerate(repetitions):
block = _residual_block3d(block_fn,filters=filters,kernel_regularizer=l2(reg_factor),repetitions=r,is_first_layer=(i == 0)
)(block)
filters *= 2
# last activation
block_output = _bn_relu(block)
# average poll and classification
pool2 = AveragePooling3D(pool_size=(block.shape[DIM1_AXIS],block.shape[DIM2_AXIS],block.shape[DIM3_AXIS]),1))(block_output)
flatten1 = Flatten()(pool2)
if num_outputs > 1:
dense = Dense(units=num_outputs,activation="softmax",kernel_regularizer=l2(reg_factor))(flatten1)
else:
dense = Dense(units=num_outputs,activation="sigmoid",kernel_regularizer=l2(reg_factor))(flatten1)
model = Model(inputs=input,outputs=dense)
return model
@staticmethod
def build_resnet_18(input_shape,reg_factor=1e-4):
"""Build resnet 18."""
return resnet3DBuilder.build(input_shape,basic_block,[2,2],reg_factor=reg_factor)
@staticmethod
def build_resnet_34(input_shape,reg_factor=1e-4):
"""Build resnet 34."""
return resnet3DBuilder.build(input_shape,[3,4,6,3],reg_factor=reg_factor)
@staticmethod
def build_resnet_50(input_shape,reg_factor=1e-4):
"""Build resnet 50."""
return resnet3DBuilder.build(input_shape,bottleneck,reg_factor=reg_factor)
@staticmethod
def build_resnet_101(input_shape,reg_factor=1e-4):
"""Build resnet 101."""
return resnet3DBuilder.build(input_shape,23,reg_factor=reg_factor)
@staticmethod
def build_resnet_152(input_shape,reg_factor=1e-4):
"""Build resnet 152."""
return resnet3DBuilder.build(input_shape,8,36,reg_factor=reg_factor)
在视频上训练网络时,出现以下错误:
Error: unhashable type: 'Dimension'
File "<ipython-input-29-788d091a6763>",line 1961,in main
trained_model_name)
File "<ipython-input-29-788d091a6763>",line 1805,in train
model = train_load_model(model_type,training_condition,sample_input.shape,nb_classes)
File "<ipython-input-29-788d091a6763>",line 1684,in train_load_model
model = resnet3DBuilder.build_resnet_50((96,96,20)
File "<ipython-input-29-788d091a6763>",line 1543,in build_resnet_50
[3,reg_factor=reg_factor)
File "<ipython-input-29-788d091a6763>",line 1501,in build
)(block)
File "<ipython-input-29-788d091a6763>",line 1372,in f
)(input)
File "<ipython-input-29-788d091a6763>",line 1419,line 1334,in f
activation = _bn_relu(input)
File "<ipython-input-29-788d091a6763>",line 1300,in _bn_relu
norm = Batchnormalization(axis=CHANNEL_AXIS)(input)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py",line 440,in __call__
self.assert_input_compatibility(inputs)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py",line 345,in assert_input_compatibility
x_shape[int(axis)] not in {value,None}):
我有以下Tensorflow,Keras和Python版本:
- Tensorflow:1.15.0
- Keras:2.2.4
- Python:3.6
您能告诉我如何解决此错误吗?我看到将Dimension转换为int可以在其他地方解决问题,但是我不知道该在这里强制转换。
解决方法
要解决此问题,我们需要将每个形状的访问权限都转换为int。
示例:
residual.shape[CHANNEL_AXIS]
需要重写int(residual.shape[CHANNEL_AXIS])
该代码的新版本如下:
## Resnet 3D architecture
# Taken from https://github.com/JihongJu/keras-resnet3d/blob/master/resnet3d/resnet3d.py
def _bn_relu(input):
"""Helper to build a BN -> relu block (by @raghakot)."""
norm = BatchNormalization(axis=CHANNEL_AXIS)(input)
return Activation("relu")(norm)
def _conv_bn_relu3D(**conv_params):
filters = conv_params["filters"]
kernel_size = conv_params["kernel_size"]
strides = conv_params.setdefault("strides",(1,1,1))
kernel_initializer = conv_params.setdefault(
"kernel_initializer","he_normal")
padding = conv_params.setdefault("padding","same")
kernel_regularizer = conv_params.setdefault("kernel_regularizer",l2(1e-4))
def f(input):
conv = Conv3D(filters=filters,kernel_size=kernel_size,strides=strides,kernel_initializer=kernel_initializer,padding=padding,kernel_regularizer=kernel_regularizer)(input)
return _bn_relu(conv)
return f
def _bn_relu_conv3d(**conv_params):
"""Helper to build a BN -> relu -> conv3d block."""
filters = conv_params["filters"]
kernel_size = conv_params["kernel_size"]
strides = conv_params.setdefault("strides",1))
kernel_initializer = conv_params.setdefault("kernel_initializer",l2(1e-4))
def f(input):
activation = _bn_relu(input)
return Conv3D(filters=filters,kernel_regularizer=kernel_regularizer)(activation)
return f
def _shortcut3d(input,residual):
"""3D shortcut to match input and residual and merges them with "sum"."""
stride_dim1 = math.ceil(int(input.shape[DIM1_AXIS]) \
/ int(residual.shape[DIM1_AXIS]))
stride_dim2 = math.ceil(int(input.shape[DIM2_AXIS]) \
/ int(residual.shape[DIM2_AXIS]))
stride_dim3 = math.ceil(int(input.shape[DIM3_AXIS]) \
/ int(residual.shape[DIM3_AXIS]))
equal_channels = int(residual.shape[CHANNEL_AXIS]) \
== int(input.shape[CHANNEL_AXIS])
shortcut = input
if stride_dim1 > 1 or stride_dim2 > 1 or stride_dim3 > 1 \
or not equal_channels:
shortcut = Conv3D(
filters=int(residual.shape[CHANNEL_AXIS]),kernel_size=(1,1),strides=(stride_dim1,stride_dim2,stride_dim3),kernel_initializer="he_normal",padding="valid",kernel_regularizer=l2(1e-4)
)(input)
return add([shortcut,residual])
def _residual_block3d(block_function,filters,kernel_regularizer,repetitions,is_first_layer=False):
def f(input):
for i in range(repetitions):
strides = (1,1)
if i == 0 and not is_first_layer:
strides = (2,2,2)
input = block_function(filters=filters,kernel_regularizer=kernel_regularizer,is_first_block_of_first_layer=(
is_first_layer and i == 0)
)(input)
return input
return f
def basic_block(filters,strides=(1,kernel_regularizer=l2(1e-4),is_first_block_of_first_layer=False):
"""Basic 3 X 3 X 3 convolution blocks. Extended from raghakot's 2D impl."""
def f(input):
if is_first_block_of_first_layer:
# don't repeat bn->relu since we just did bn->relu->maxpool
conv1 = Conv3D(filters=filters,kernel_size=(3,3,3),padding="same",kernel_regularizer=kernel_regularizer
)(input)
else:
conv1 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
residual = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(conv1)
return _shortcut3d(input,residual)
return f
def bottleneck(filters,is_first_block_of_first_layer=False):
"""Basic 3 X 3 X 3 convolution blocks. Extended from raghakot's 2D impl."""
def f(input):
if is_first_block_of_first_layer:
# don't repeat bn->relu since we just did bn->relu->maxpool
conv_1_1 = Conv3D(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
else:
conv_1_1 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(input)
conv_3_3 = _bn_relu_conv3d(filters=filters,kernel_regularizer=kernel_regularizer
)(conv_1_1)
residual = _bn_relu_conv3d(filters=filters * 4,kernel_regularizer=kernel_regularizer
)(conv_3_3)
return _shortcut3d(input,residual)
return f
def _handle_data_format():
global DIM1_AXIS
global DIM2_AXIS
global DIM3_AXIS
global CHANNEL_AXIS
if K.image_data_format() == 'channels_last':
print("here CHANNELS last")
DIM1_AXIS = 1
DIM2_AXIS = 2
DIM3_AXIS = 3
CHANNEL_AXIS = 4
else:
CHANNEL_AXIS = 1
DIM1_AXIS = 2
DIM2_AXIS = 3
DIM3_AXIS = 4
def _get_block(identifier):
if isinstance(identifier,six.string_types):
res = globals().get(identifier)
if not res:
raise ValueError('Invalid {}'.format(identifier))
return res
return identifier
class Resnet3DBuilder(object):
"""ResNet3D."""
@staticmethod
def build(input_shape,num_outputs,block_fn,reg_factor):
"""Instantiate a vanilla ResNet3D keras model.
# Arguments
input_shape: Tuple of input shape in the format
(conv_dim1,conv_dim2,conv_dim3,channels) if dim_ordering='tf'
(filter,conv_dim1,conv_dim3) if dim_ordering='th'
num_outputs: The number of outputs at the final softmax layer
block_fn: Unit block to use {'basic_block','bottlenack_block'}
repetitions: Repetitions of unit blocks
# Returns
model: a 3D ResNet model that takes a 5D tensor (volumetric images
in batch) as input and returns a 1D vector (prediction) as output.
"""
_handle_data_format()
if len(input_shape) != 4:
raise ValueError("Input shape should be a tuple "
"(conv_dim1,channels) "
"for tensorflow as backend or "
"(channels,conv_dim3) "
"for theano as backend")
block_fn = _get_block(block_fn)
input = Input(shape=input_shape)
# first conv
conv1 = _conv_bn_relu3D(filters=64,kernel_size=(7,7,7),strides=(2,2),kernel_regularizer=l2(reg_factor)
)(input)
pool1 = MaxPooling3D(pool_size=(3,padding="same")(conv1)
# repeat blocks
block = pool1
filters = 64
for i,r in enumerate(repetitions):
block = _residual_block3d(block_fn,filters=filters,kernel_regularizer=l2(reg_factor),repetitions=r,is_first_layer=(i == 0)
)(block)
filters *= 2
# last activation
block_output = _bn_relu(block)
# average poll and classification
pool2 = AveragePooling3D(pool_size=(int(block.shape[DIM1_AXIS]),int(block.shape[DIM2_AXIS]),int(block.shape[DIM3_AXIS])),1))(block_output)
flatten1 = Flatten()(pool2)
if num_outputs > 1:
dense = Dense(units=num_outputs,activation="softmax",kernel_regularizer=l2(reg_factor))(flatten1)
else:
dense = Dense(units=num_outputs,activation="sigmoid",kernel_regularizer=l2(reg_factor))(flatten1)
model = Model(inputs=input,outputs=dense)
return model
@staticmethod
def build_resnet_18(input_shape,reg_factor=1e-4):
"""Build resnet 18."""
return Resnet3DBuilder.build(input_shape,basic_block,[2,2],reg_factor=reg_factor)
@staticmethod
def build_resnet_34(input_shape,reg_factor=1e-4):
"""Build resnet 34."""
return Resnet3DBuilder.build(input_shape,[3,4,6,3],reg_factor=reg_factor)
@staticmethod
def build_resnet_50(input_shape,reg_factor=1e-4):
"""Build resnet 50."""
return Resnet3DBuilder.build(input_shape,bottleneck,reg_factor=reg_factor)
@staticmethod
def build_resnet_101(input_shape,reg_factor=1e-4):
"""Build resnet 101."""
return Resnet3DBuilder.build(input_shape,23,reg_factor=reg_factor)
@staticmethod
def build_resnet_152(input_shape,reg_factor=1e-4):
"""Build resnet 152."""
return Resnet3DBuilder.build(input_shape,8,36,reg_factor=reg_factor)
最新版本的Keras和Tensorflow不会出现此问题,但是我需要保留这两个库的旧版本,因为我的其他脚本无法在Tensorflow / Keras的最新版本上运行。
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。