def attention_layer(x, ratio=4, scope=None):
with tf.name_scope(scope):
_, width, height, channel = x.get_shape().as_list()
x_shape = x.get_shape().as_list()
recalibrate1 = conv_sigmoid(x,
kernal=(1, 1, channel, 1),
scope=scope + 'spatial_squeeze')
squeeze = tf.reduce_mean(x,
axis=(1, 2),
name=scope + 'channel_squeeze')
exciation = full_connected_relu(squeeze,
kernal=(channel, channel // ratio),
activefuncation='relu',
scope=scope + '_fully_connected1')
exciation = full_connected_relu(exciation,
kernal=(channel // ratio, channel),
activefuncation='sigmoid',
scope=scope + '_fully_connected2')
recalibrate2 = tf.reshape(exciation, [-1, 1, 1, channel])
recalibrate3 = conv_sigmoid(x,
kernal=(3, 3, channel, channel // 2),
scope=scope + 'conv1')
recalibrate3 = max_pooling_2x2(recalibrate3)
recalibrate3 = conv_sigmoid(recalibrate3,
kernal=(3, 3, channel // 2, channel // 4),
scope=scope + 'conv2')
recalibrate3 = max_pooling_2x2(recalibrate3)
kernal = [3, 3, channel, channel // 4]
W = weight_xavier_init(shape=kernal,
n_inputs=kernal[0] * kernal[1] * kernal[-1],
n_outputs=kernal[-2],
activefunction='relu',
variable_name=scope + 'W')
B = bias_variable([kernal[-2]], variable_name=scope + 'B')
output_shape = tf.stack([2, x_shape[1], x_shape[2], channel])
deconv = tf.nn.conv2d_transpose(recalibrate3,
W,
output_shape,
strides=[1, 4, 4, 1],
padding="SAME") + B
recalibrate3 = tf.nn.leaky_relu(deconv)
# recalibrate3 = tf.image.resize_images(images=recalibrate3, size=[width, height])
recalibrate4 = tf.multiply(recalibrate1, recalibrate2)
alpha_attention = tf.Variable(tf.constant(1.0))
beta_attention = tf.Variable(tf.constant(1.0))
attention = tf.multiply(alpha_attention * recalibrate4,
beta_attention * recalibrate3)
out = tf.multiply(attention, x)
return out
def channel_squeeze_spatial_excitiation_layer(x, out_dim, scope=None):
with tf.name_scope(scope):
squeeze = conv_sigmoid(x,
kernal=(1, 1, out_dim, 1),
scope=scope + 'spatial_squeeze')
scale = x * squeeze
return scale
def subtract_refine(subtract, layer7, layer8, layer9, phase, drop_conv, n_class=1):
_, y_width, y_height, y_channel = subtract.get_shape().as_list() # 拿出 输入 tensor 的 最后一维:也就是通道数
seed = tf.reshape(subtract, [-1, y_width, y_height, y_channel]) # 将图片转换成tf识别格式
refine1 = conv_bn_relu_drop(x=layer7, kernal=[1, 1, 128, 64], phase=phase, drop=drop_conv, scope='refine1_1')
seed1 = tf.image.resize_images(images=seed, size=[int(y_width / 4), int(y_height / 4)], method=0)
refine1 = crop_and_concat(refine1, seed1)
refine1 = conv_bn_relu_drop(x=refine1, kernal=[3, 3, 65, 65], phase=phase, drop=drop_conv, scope='refine1_2')
refine1 = squeeze_excitation_model(refine1, out_dim=65, scope='sem2_1')
refine1 = deconv_relu(x=refine1, kernal=[3, 3, 32, 65], scope='refine1_deconv')
refine2 = conv_bn_relu_drop(x=layer8, kernal=[1, 1, 64, 32], phase=phase, drop=drop_conv, scope='refine2_1')
seed2 = tf.image.resize_images(images=seed, size=[int(y_width / 2), int(y_height / 2)], method=0)
refine2 = crop_and_concat(refine2, seed2)
refine2 = conv_bn_relu_drop(x=refine2, kernal=[3, 3, 33, 32], phase=phase, drop=drop_conv, scope='refine2_2')
refine2 = crop_and_concat(refine1, refine2)
refine2 = conv_bn_relu_drop(x=refine2, kernal=[3, 3, 64, 64], phase=phase, drop=drop_conv, scope='refine1_3')
refine2 = squeeze_excitation_model(refine2, out_dim=64, scope='sem2_2')
refine2 = deconv_relu(x=refine2, kernal=[3, 3, 32, 64], scope='refine2_deconv')
refine3 = conv_bn_relu_drop(x=layer9, kernal=[1, 1, 32, 1], phase=phase, drop=drop_conv, scope='refine3_1')
seed3 = tf.image.resize_images(images=seed, size=[int(y_width), int(y_height)], method=0)
refine3 = crop_and_concat(refine3, seed3)
refine3 = conv_bn_relu_drop(x=refine3, kernal=[3, 3, 2, 1], phase=phase, drop=drop_conv, scope='refine3_2')
refine3 = crop_and_concat(refine2, refine3)
refine3 = conv_bn_relu_drop(x=refine3, kernal=[3, 3, 33, 33], phase=phase, drop=drop_conv, scope='refine3_3')
refine3 = squeeze_excitation_model(refine3, out_dim=33, scope='sem2_3')
out1 = conv_sigmoid(x=refine3, kernal=[1, 1, 33, n_class], scope='out1')
return out1
def attention_layer(x, ratio=4, scope=None):
with tf.name_scope(scope):
_, width, height, channel = x.get_shape().as_list()
x_shape = x.get_shape().as_list()
recalibrate1 = conv_sigmoid(x,
kernal=(1, 1, channel, 1),
scope=scope + 'spatial_squeeze')
squeeze = tf.reduce_mean(x,
axis=(1, 2),
name=scope + 'channel_squeeze')
exciation = full_connected_relu(squeeze,
kernal=(channel, channel // ratio),
activefuncation='relu',
scope=scope + '_fully_connected1')
exciation = full_connected_relu(exciation,
kernal=(channel // ratio, channel),
activefuncation='sigmoid',
scope=scope + '_fully_connected2')
recalibrate2 = tf.reshape(exciation, [-1, 1, 1, channel])
recalibrate4 = tf.multiply(recalibrate1, recalibrate2)
out = tf.multiply(recalibrate4, x)
return out
def _create_conv_net(X, image_width, image_height, image_channel, phase, drop_conv, n_class=1):
inputX = tf.reshape(X, [-1, image_width, image_height, image_channel]) # shape=(?, 32, 32, 1)
# UNet model
# layer1->convolution
layer0 = conv_bn_relu_drop(x = inputX, kernal = [3, 3, image_channel, 32], phase= phase, drop = drop_conv, scope = 'layer0')
layer1 = conv_bn_relu_drop(x = layer0, kernal= [3, 3, 32, 32], phase = phase, drop=drop_conv, scope = 'layer1')
# print(layer1.get_shape().as_list())
pool1 = max_pooling_2x2(layer1)
# layer2->convolution
layer2 = conv_bn_relu_drop(x = pool1, kernal = [3, 3, 32, 64], phase= phase, drop = drop_conv, scope= 'layer2_1')
layer2 = conv_bn_relu_drop(x = layer2, kernal=[3, 3, 64, 64], phase = phase, drop = drop_conv, scope = 'layer2_2')
pool2 = max_pooling_2x2(layer2)
# layer3->convolution
layer3 = conv_bn_relu_drop(x = pool2, kernal= [3, 3, 64, 128], phase = phase, drop=drop_conv, scope = 'layer3_1')
layer3 = conv_bn_relu_drop(x = layer3, kernal= [3, 3, 128, 128], phase = phase, drop=drop_conv, scope = 'layer3_2')
pool3 = max_pooling_2x2(layer3)
# layer4->convolution
layer4 = conv_bn_relu_drop(x = pool3, kernal= [3, 3, 128, 256], phase=phase, drop=drop_conv, scope = 'layer4_1')
layer4 = conv_bn_relu_drop(x = layer4, kernal=[3, 3, 256, 256], phase=phase, drop=drop_conv, scope= 'layer4_2')
pool4 = max_pooling_2x2(layer4)
# layer5->convolution
layer5 = conv_bn_relu_drop(x = pool4, kernal= [3, 3, 256, 512], phase=phase, drop=drop_conv, scope='layer5_1')
layer5 = conv_bn_relu_drop(x = layer5, kernal=[3, 3, 512, 512], phase=phase, drop = drop_conv, scope='layer5_2')
# layer6->deconvolution
deconv1 = deconv_relu(x = layer5, kernal=[3, 3, 256, 512], scope = 'deconv1')
layer6 = crop_and_concat(layer4, deconv1)
# layer7->convolution
layer6 = conv_bn_relu_drop(x = layer6, kernal= [3, 3, 512, 256], phase=phase, drop=drop_conv, scope = 'layer6_1')
layer6 = conv_bn_relu_drop(x = layer6, kernal= [3, 3, 256, 256], phase= phase, drop=drop_conv, scope = 'layer6_2')
deconv2 = deconv_relu(layer6, kernal=[3, 3, 128, 256], scope = 'deconv2')
layer7 = crop_and_concat(layer3, deconv2)
# layer9->convolution
layer7 = conv_bn_relu_drop(x = layer7, kernal= [3, 3, 256, 128], phase=phase, drop=drop_conv, scope = 'layer7_1')
layer7 = conv_bn_relu_drop(x = layer7, kernal = [3, 3, 128, 128], phase=phase, drop=drop_conv, scope='layer7_2')
# layer10->deconvolution
deconv3 = deconv_relu(x = layer7, kernal= [3, 3, 64, 128], scope = 'deconv3')
layer8 = crop_and_concat(layer2, deconv3)
# layer11->convolution
layer8 = conv_bn_relu_drop(x = layer8, kernal= [3, 3, 128, 64], phase=phase, drop=drop_conv, scope='layer8_1')
layer8 = conv_bn_relu_drop(x = layer8, kernal= [3, 3, 64, 64], phase=phase, drop=drop_conv, scope='layer8_2')
deconv4 = deconv_relu(layer8, kernal=[3, 3, 32, 64], scope='deconv4')
layer9 = crop_and_concat(layer1, deconv4)
# layer 13->convolution
layer9 = conv_bn_relu_drop(x = layer9, kernal=[3, 3, 64, 32], phase=phase, drop=drop_conv, scope='layer9_1')
layer9 = conv_bn_relu_drop(x = layer9, kernal=[3, 3, 32, 32], phase=phase, drop=drop_conv, scope='layer9_2')
# layer14->output
out1 = conv_sigmoid(x = layer9, kernal=[1, 1, 32, n_class], scope = 'out1')
return out1