def identity_block_td(input_tensor, kernel_size, filters, stage, block, trainable=True):
# identity block time distributed
nb_filter1, nb_filter2, nb_filter3 = filters
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
# 我们可以使用包装器TimeDistributed包装Dense，以产生针对各个时间步（默认输入tensor的第一维，也就是batch之后的那个维度是时间维度）信号的独立全连接。
x = TimeDistributed(Convolution2D(nb_filter1, (1, 1), trainable=trainable, kernel_initializer='normal'),
name=conv_name_base + '2a')(input_tensor)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = TimeDistributed(
Convolution2D(nb_filter2, (kernel_size, kernel_size), trainable=trainable, kernel_initializer='normal',
padding='same'), name=conv_name_base + '2b')(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = TimeDistributed(Convolution2D(nb_filter3, (1, 1), trainable=trainable, kernel_initializer='normal'),
name=conv_name_base + '2c')(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2c')(x)
x = Add()([x, input_tensor])
x = Activation('relu')(x)
return x
def conv_block_td(input_tensor, kernel_size, filters, stage, block, input_shape, strides=(2, 2), trainable=True):
# conv block time distributed
nb_filter1, nb_filter2, nb_filter3 = filters
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = TimeDistributed(
Convolution2D(nb_filter1, (1, 1), strides=strides, trainable=trainable, kernel_initializer='normal'),
input_shape=input_shape, name=conv_name_base + '2a')(input_tensor)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = TimeDistributed(Convolution2D(nb_filter2, (kernel_size, kernel_size), padding='same', trainable=trainable,
kernel_initializer='normal'), name=conv_name_base + '2b')(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = TimeDistributed(Convolution2D(nb_filter3, (1, 1), kernel_initializer='normal'), name=conv_name_base + '2c',
trainable=trainable)(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '2c')(x)
shortcut = TimeDistributed(
Convolution2D(nb_filter3, (1, 1), strides=strides, trainable=trainable, kernel_initializer='normal'),
name=conv_name_base + '1')(input_tensor)
shortcut = TimeDistributed(FixedBatchNormalization(axis=bn_axis), name=bn_name_base + '1')(shortcut)
x = Add()([x, shortcut])
x = Activation('relu')(x)
return x
def identity_block(input_tensor, kernel_size, filters, stage, block, trainable=True):
nb_filter1, nb_filter2, nb_filter3 = filters
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = Convolution2D(nb_filter1, (1, 1), name=conv_name_base + '2a', trainable=trainable)(input_tensor)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Convolution2D(nb_filter2, (kernel_size, kernel_size), padding='same', name=conv_name_base + '2b',
trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = Convolution2D(nb_filter3, (1, 1), name=conv_name_base + '2c', trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
x = Add()([x, input_tensor])
x = Activation('relu')(x)
return x
def conv_block(input_tensor,
kernel_size,
filters,
stage,
block,
strides=(2, 2),
trainable=True):
nb_filter1, nb_filter2, nb_filter3 = filters
# -if K.image_dim_ordering() == 'tf':
if K.image_data_format() == 'channels_last':
# -bn_axis = 3
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = Convolution2D(nb_filter1, (1, 1),
strides=strides,
name=conv_name_base + '2a',
trainable=trainable)(input_tensor)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Convolution2D(nb_filter2, (kernel_size, kernel_size),
padding='same',
name=conv_name_base + '2b',
trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = Convolution2D(nb_filter3, (1, 1),
name=conv_name_base + '2c',
trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
shortcut = Convolution2D(nb_filter3, (1, 1),
strides=strides,
name=conv_name_base + '1',
trainable=trainable)(input_tensor)
shortcut = FixedBatchNormalization(axis=bn_axis,
name=bn_name_base + '1')(shortcut)
x = Add()([x, shortcut])
x = Activation('relu')(x)
return x
def nn_base(input_tensor=None, trainable=False, channels = 3):
# Determine proper input shape
if K.image_dim_ordering() == 'th':
input_shape = (channels, None, None)
else:
input_shape = (None, None, channels)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
x = ZeroPadding2D((3, 3), name='conv1_pad')(img_input)
if channels == 3:
x = Convolution2D(64, (7, 7), strides=(2, 2), name='conv1', trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name='bn_conv1')(x)
else:
x = Convolution2D(64, (7, 7), strides=(2, 2), name='conv1_channel_'+str(channels), trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name='bn_conv1_channel_'+str(channels))(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1), trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='b', trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='c', trainable=trainable)
x = conv_block(x, 3, [128, 128, 512], stage=3, block='a', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='b', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='c', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='d', trainable=trainable)
x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f', trainable=trainable)
return x
def identity_block_td(input_tensor,
kernel_size,
filters,
stage,
block,
trainable=True):
# Identity block time distributed
nb_filter1, nb_filter2, nb_filter3 = filters
# -if K.image_dim_ordering() == 'tf':
if K.image_data_format() == 'channels_last':
# -bn_axis = 3
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = TimeDistributed(Convolution2D(nb_filter1, (1, 1),
trainable=trainable,
kernel_initializer='normal'),
name=conv_name_base + '2a')(input_tensor)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis),
name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = TimeDistributed(Convolution2D(nb_filter2, (kernel_size, kernel_size),
trainable=trainable,
kernel_initializer='normal',
padding='same'),
name=conv_name_base + '2b')(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis),
name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = TimeDistributed(Convolution2D(nb_filter3, (1, 1),
trainable=trainable,
kernel_initializer='normal'),
name=conv_name_base + '2c')(x)
x = TimeDistributed(FixedBatchNormalization(axis=bn_axis),
name=bn_name_base + '2c')(x)
x = Add()([x, input_tensor])
x = Activation('relu')(x)
return x
def nn_base(input_tensor=None, trainable=False):
# Determine proper input shape
if K.image_dim_ordering() == 'th':
input_shape = (3, None, None)
else:
input_shape = (None, None, 3)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
x = ZeroPadding2D((3, 3))(img_input)
# print('++ zero padding: ', x)
x = Convolution2D(64, (7, 7), strides=(2, 2), name='conv1', trainable=trainable)(x)
# print('+ conv 2d: ', x)
print('NOTE: this code only support to keras 2.0.3, newest version this line will got errors. see trace back.')
x = FixedBatchNormalization(axis=bn_axis, name='bn_conv1')(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1), trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='b', trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='c', trainable=trainable)
x = conv_block(x, 3, [128, 128, 512], stage=3, block='a', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='b', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='c', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='d', trainable=trainable)
x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f', trainable=trainable)
return x
def nn_base(input_tensor=None, trainable=False):
# Determine proper input shape
if K.image_dim_ordering() == 'th':
input_shape = (3, None, None)
else:
input_shape = (None, None, 3)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
# ZeroPadding2D层的功能：对2D输入（如图片）的边界填充0，以控制卷积以后特征图的大小，（3,3）表示在图片边缘填充3个0
x = ZeroPadding2D((3, 3))(img_input)
x = Convolution2D(64, (7, 7), strides=(2, 2), name='conv1', trainable=trainable)(x)
x = FixedBatchNormalization(axis=bn_axis, name='bn_conv1')(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1), trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='b', trainable=trainable)
x = identity_block(x, 3, [64, 64, 256], stage=2, block='c', trainable=trainable)
x = conv_block(x, 3, [128, 128, 512], stage=3, block='a', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='b', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='c', trainable=trainable)
x = identity_block(x, 3, [128, 128, 512], stage=3, block='d', trainable=trainable)
x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e', trainable=trainable)
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f', trainable=trainable)
return x
def nn_base(input_tensor=None, trainable=False):
# Determine proper input shape
if K.image_dim_ordering() == 'th':
input_shape = (3, None, None)
else:
input_shape = (None, None, 3)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
# 自定义输入，为了显示
# img_numpy = np.full((1,38,166,3),2)
# x = tf.convert_to_tensor(img_numpy,dtype=tf.float32)
x = ZeroPadding2D((3, 3))(img_input)
# stage1
# conv1 7*7,64,stride 2
conv1 = Convolution2D(64, (6, 6),
strides=(2, 2),
name='conv1',
trainable=trainable)(x)
bn_conv1 = FixedBatchNormalization(axis=bn_axis, name='bn_conv1')(conv1)
conv1_relu = Activation('relu')(bn_conv1)
conv1_pool = MaxPooling2D((3, 3), strides=(2, 2))(conv1_relu)
print("stage1:{}".format(conv1_pool.get_shape())) # (1, 8, 40, 64)
# print("stage1:{}".format(x.shape))
# stage2
# conv2_x [ [1*1,64], [3*3,64], [1*1,256] ]*3
conv2_a = conv_block(conv1_pool,
3, [64, 64, 256],
stage=2,
block='a',
strides=(1, 1),
trainable=trainable)
conv2_b = identity_block(conv2_a,
3, [64, 64, 256],
stage=2,
block='b',
trainable=trainable)
conv2_c = identity_block(conv2_b,
3, [64, 64, 256],
stage=2,
block='c',
trainable=trainable)
# 最后相加的
x2 = UpSampling2D((2, 2))(conv2_c)
# conv2_output (1, 16, 80, 1024)
x2_output = Conv2D(filters=1024,
kernel_size=3,
strides=(1, 1),
padding='same')(x2)
# x2_output = conv_block(x2, 3, [64, 64, 1024], stage=2, block='diy', strides=(1, 1), trainable=trainable)
# tmp_conv2_1024 = identity_block(conv2_b, 3, [128, 128, 1024], stage=2, block='c', trainable=trainable)
# ValueError: Operands could not be broadcast together with shapes (8, 40, 1024) (8, 40, 256) 是因为shortcut要同一尺度相加
# x2 = UpSampling2D((2, 2))(conv2_c)
print("stage2:{}".format(x2.get_shape()))
print("x2_output:{}".format(x2_output.shape))
# stage3
# conv3_x [ [1*1,128], [3*3,128], [1*1,512] ]*4
conv3_a = conv_block(x2,
3, [128, 128, 512],
stage=3,
block='a',
trainable=trainable)
conv3_b = identity_block(conv3_a,
3, [128, 128, 512],
stage=3,
block='b',
trainable=trainable)
conv3_c = identity_block(conv3_b,
3, [128, 128, 512],
stage=3,
block='c',
trainable=trainable)
conv3_d = identity_block(conv3_c,
3, [128, 128, 512],
stage=3,
block='d',
trainable=trainable)
x3 = UpSampling2D((2, 2))(conv3_d)
print("stage3:{}".format(x3.get_shape()))
# stage4
# conv4_x [ [1*1,256], [3*3,256], [1*1,1024] ]*6
conv4_a = conv_block(x3,
3, [256, 256, 1024],
stage=4,
block='a',
trainable=trainable)
conv4_b = identity_block(conv4_a,
3, [256, 256, 1024],
stage=4,
block='b',
trainable=trainable)
conv4_c = identity_block(conv4_b,
3, [256, 256, 1024],
stage=4,
block='c',
trainable=trainable)
conv4_d = identity_block(conv4_c,
3, [256, 256, 1024],
stage=4,
block='d',
trainable=trainable)
conv4_e = identity_block(conv4_d,
3, [256, 256, 1024],
stage=4,
block='e',
trainable=trainable)
conv4_f = identity_block(conv4_e,
3, [256, 256, 1024],
stage=4,
block='f',
trainable=trainable)
x4 = UpSampling2D((2, 2))(conv4_f)
x = x4 + x2_output
print("stage4:{}".format(x4.get_shape()))
print("return x_shape:{}".format(x.shape))
return x
def nn_base(input_tensor=None, trainable=False):
# Determine proper input shape
if K.image_dim_ordering() == 'th':
input_shape = (3, None, None) # RGB通道在前
else:
input_shape = (None, None, 3) # RGB通道在尾
if input_tensor is None:
img_input = Input(shape=input_shape) # 输入的shape大小不做限定
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
if K.image_dim_ordering() == 'tf':
bn_axis = 3
else:
bn_axis = 1
# ZeroPadding2D层的功能：对2D输入（如图片）的边界填充0，以控制卷积以后特征图的大小，（3,3）表示在图片边缘填充3个0
x = ZeroPadding2D((3, 3))(img_input)
x = Convolution2D(64, (7, 7),
strides=(2, 2),
name='conv1',
trainable=trainable)(x)
print("Convolution2D shape:", x.shape)
x = FixedBatchNormalization(axis=bn_axis, name='bn_conv1')(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
print("MaxPooling2D shape:", x.shape)
x = conv_block(x,
3, [64, 64, 256],
stage=2,
block='a',
strides=(1, 1),
trainable=trainable) #conv_block的默认stride是2
print("x stage2 block a shape:", x.shape)
x = identity_block(x,
3, [64, 64, 256],
stage=2,
block='b',
trainable=trainable)
print("x stage2 block b shape:", x.shape)
x = identity_block(x,
3, [64, 64, 256],
stage=2,
block='c',
trainable=trainable)
print("x stage2 block c shape:", x.shape)
x = conv_block(x,
3, [128, 128, 512],
stage=3,
block='a',
trainable=trainable) #conv_block的默认stride是2
print("x stage3 block a shape:", x.shape)
x = identity_block(x,
3, [128, 128, 512],
stage=3,
block='b',
trainable=trainable) # identity_block不改变输入输出的shape
print("x stage3 block b shape:", x.shape)
x = identity_block(x,
3, [128, 128, 512],
stage=3,
block='c',
trainable=trainable)
print("x stage3 block c shape:", x.shape)
xstage3 = identity_block(x,
3, [128, 128, 512],
stage=3,
block='d',
trainable=trainable)
print("x stage3 block d shape:", xstage3.shape)
x = conv_block(x,
3, [256, 256, 1024],
stage=4,
block='a',
trainable=trainable)
print("x stage4 block a shape:", x.shape)
x = identity_block(x,
3, [256, 256, 1024],
stage=4,
block='b',
trainable=trainable)
print("x stage4 block b shape:", x.shape)
x = identity_block(x,
3, [256, 256, 1024],
stage=4,
block='c',
trainable=trainable)
print("x stage4 block c shape:", x.shape)
x = identity_block(x,
3, [256, 256, 1024],
stage=4,
block='d',
trainable=trainable)
print("x stage4 block d shape:", x.shape)
x = identity_block(x,
3, [256, 256, 1024],
stage=4,
block='e',
trainable=trainable)
print("x stage4 block e shape:", x.shape)
xstage4 = identity_block(x,
3, [256, 256, 1024],
stage=4,
block='f',
trainable=trainable)
print("x stage4 block f shape:", xstage4.shape)
#identity_block(input_tensor, kernel_size, filters, stage, block, trainable=True)
# resnet的结构解读参考https://blog.csdn.net/lanran2/article/details/79057994
# basenet只用到了resnet50的conv1到conv4_x,剪切掉了conv5_x
# 合并两个卷积层
stage3_lateral = Convolution2D(1024, (1, 1),
strides=(1, 1),
name='stage3_lateral',
trainable=trainable)(xstage3)
up4 = UpSampling2D(size=(2, 2))(xstage4)
print("up4 shape:", up4.shape)
print("up4.get_shape()=", up4.get_shape())
#up4 = Cropping2D(cropping=((up4.get_shape()[1]-stage3_lateral.get_shape()[1], 0), (up4.get_shape()[2]-stage3_lateral.get_shape()[2], 0)), data_format=None)(up4)
# 目前还缺一个crop层，导致不是所有情况下stage3_lateral,up4的shape都一样,从而不能Add
share_layer = Add()([stage3_lateral, up4])
print("share_layer shape:", share_layer.shape)
return share_layer, xstage3, xstage4