'''
covlution layer，pool layer，initialization。。。。
'''
from __future__ import division
import tensorflow as tf
import numpy as np
import cv2
# Weight initialization (Xavier's init)
def weight_xavier_init(shape, n_inputs, n_outputs, activefunction='sigomd', uniform=True, variable_name=None):
if activefunction == 'sigomd':
if uniform:
init_range = tf.sqrt(6.0 / (n_inputs + n_outputs))
initial = tf.random_uniform(shape, -init_range, init_range)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
else:
stddev = tf.sqrt(2.0 / (n_inputs + n_outputs))
initial = tf.truncated_normal(shape, mean=0.0, stddev=stddev)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
elif activefunction == 'relu':
if uniform:
init_range = tf.sqrt(6.0 / (n_inputs + n_outputs)) * np.sqrt(2)
initial = tf.random_uniform(shape, -init_range, init_range)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
else:
stddev = tf.sqrt(2.0 / (n_inputs + n_outputs)) * np.sqrt(2)
initial = tf.truncated_normal(shape, mean=0.0, stddev=stddev)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
elif activefunction == 'tan':
if uniform:
init_range = tf.sqrt(6.0 / (n_inputs + n_outputs)) * 4
initial = tf.random_uniform(shape, -init_range, init_range)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
else:
stddev = tf.sqrt(2.0 / (n_inputs + n_outputs)) * 4
initial = tf.truncated_normal(shape, mean=0.0, stddev=stddev)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
# Bias initialization
def bias_variable(shape, variable_name=None):
initial = tf.constant(0.1, shape=shape)
return tf.get_variable(name=variable_name, initializer=initial, trainable=True)
# 3D convolution
def conv3d(x, W, stride=1):
conv_3d = tf.nn.conv3d(x, W, strides=[1, stride, stride, stride, 1], padding='SAME')
return conv_3d
# 3D upsampling
def upsample3d(x, scale_factor, scope=None):
''''
X shape is [nsample,dim,rows, cols, channel]
out shape is[nsample,dim*scale_factor,rows*scale_factor, cols*scale_factor, channel]
'''
x_shape = tf.shape(x)
k = tf.ones([scale_factor, scale_factor, scale_factor, x_shape[-1], x_shape[-1]])
# note k.shape = [dim,rows, cols, depth_in, depth_output]
output_shape = tf.stack(
[x_shape[0], x_shape[1] * scale_factor, x_shape[2] * scale_factor, x_shape[3] * scale_factor, x_shape[4]])
upsample = tf.nn.conv3d_transpose(value=x, filter=k, output_shape=output_shape,
strides=[1, scale_factor, scale_factor, scale_factor, 1],
padding='SAME', name=scope)
return upsample
# 3D deconvolution
def deconv3d(x, W, samefeature=False, depth=False):
"""
depth flag:False is z axis is same between input and output,true is z axis is input is twice than output
"""
x_shape = tf.shape(x)
if depth:
if samefeature:
output_shape = tf.stack([x_shape[0], x_shape[1] * 2, x_shape[2] * 2, x_shape[3] * 2, x_shape[4]])
else:
output_shape = tf.stack([x_shape[0], x_shape[1] * 2, x_shape[2] * 2, x_shape[3] * 2, x_shape[4] // 2])
deconv = tf.nn.conv3d_transpose(x, W, output_shape, strides=[1, 2, 2, 2, 1], padding='SAME')
else:
if samefeature:
output_shape = tf.stack([x_shape[0], x_shape[1] * 2, x_shape[2] * 2, x_shape[3], x_shape[4]])
else:
output_shape = tf.stack([x_shape[0], x_shape[1] * 2, x_shape[2] * 2, x_shape[3], x_shape[4] // 2])
deconv = tf.nn.conv3d_transpose(x, W, output_shape, strides=[1, 2, 2, 1, 1], padding='SAME')
return deconv
# Max Pooling
def max_pool3d(x, depth=False):
"""
depth flag:False is z axis is same between input and output,true is z axis is input is twice than output
"""
if depth:
pool3d = tf.nn.max_pool3d(x, ksize=[1, 2, 2, 2, 1], strides=[1, 2, 2, 2, 1], padding='SAME')
else:
pool3d = tf.nn.max_pool3d(x, ksize=[1, 2, 2, 1, 1], strides=[1, 2, 2, 1, 1], padding='SAME')
return pool3d
# Unet crop and concat
def crop_and_concat(x1, x2):
"""
concat x1 and x2
:param x1:
:param x2:
:return:
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1] - x2_shape[1]) // 2,
(x1_shape[2] - x2_shape[2]) // 2, (x1_shape[3] - x2_shape[3]) // 2, 0]
size = [-1, x2_shape[1], x2_shape[2], x2_shape[3], -1]
x1_crop = tf.slice(x1, offsets, size)
return tf.concat([x1_crop, x2], 4)
# Batch Normalization
def normalizationlayer(x, is_train, height=None, width=None, image_z=None, norm_type=None, G=16, esp=1e-5, scope=None):
"""
normalizationlayer
:param x:input data with shap of[batch,height,width,channel]
:param is_train:flag of normalizationlayer,True is training,False is Testing
:param height:in some condition,the data height is in Runtime determined,such as through deconv layer and conv2d
:param width:in some condition,the data width is in Runtime determined
:param image_z:
:param norm_type:normalization type:support"batch","group","None"
:param G:in group normalization,channel is seperated with group number(G)
:param esp:Prevent divisor from being zero
:param scope:normalizationlayer scope
:return:
"""
with tf.name_scope(scope + norm_type):
if norm_type == None:
output = x
elif norm_type == 'batch':
output = tf.contrib.layers.batch_norm(x, center=True, scale=True, is_train=is_train)
elif norm_type == "group":
# tranpose:[bs,z,h,w,c]to[bs,c,z,h,w]following the paper
x = tf.transpose(x, [0, 4, 1, 2, 3])
N, C, Z, H, W = x.get_shape().as_list()
G = min(G, C)
if H == None and W == None and Z == None:
Z, H, W = image_z, height, width
x = tf.reshape(x, [-1, G, C // G, Z, H, W])
mean, var = tf.nn.moments(x, [2, 3, 4, 5], keep_dims=True)
x = (x - mean) / tf.sqrt(var + esp)
gama = tf.get_variable(scope + norm_type + 'group_gama', [C], initializer=tf.constant_initializer(1.0))
beta = tf.get_variable(scope + norm_type + 'group_beta', [C], initializer=tf.constant_initializer(0.0))
gama = tf.reshape(gama, [1, C, 1, 1, 1])
beta = tf.reshape(beta, [1, C, 1, 1, 1])
output = tf.reshape(x, [-1, C, Z, H, W]) * gama + beta
# tranpose:[bs,c,z,h,w]to[bs,z,h,w,c]following the paper
output = tf.transpose(output, [0, 2, 3, 4, 1])
return output
# resnet add_connect
def resnet_Add(x1, x2):
"""
add x1 and x2
:param x1:
:param x2:
:return:
"""
if x1.get_shape().as_list()[4] != x2.get_shape().as_list()[4]:
# Option A: Zero-padding
residual_connection = x2 + tf.pad(x1, [[0, 0], [0, 0], [0, 0], [0, 0],
[0, x2.get_shape().as_list()[4] -
x1.get_shape().as_list()[4]]])
else:
residual_connection = x2 + x1
return residual_connection
def save_images(images, size, path):
img = (images + 1.0) / 2.0
h, w = img.shape[1], img.shape[2]
merge_img = np.zeros((h * size[0], w * size[1]))
for idx, image in enumerate(images):
i = idx % size[1]
j = idx // size[1]
merge_img[j * h:j * h + h, i * w:i * w + w] = image
result = merge_img * 255.
result = np.clip(result, 0, 255).astype('uint8')
return cv2.imwrite(path, result)
def gatingsignal3d(x, kernal, phase, image_z=None, height=None, width=None, scope=None):
"""this is simply 1x1x1 convolution, bn, activation,Gating Signal(Query)
:param x:
:param kernal:(1,1,1,inputfilters,outputfilters)
:param phase:
:param drop:
:param image_z:
:param height:
:param width:
:param scope:
:return:
"""
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1], activefunction='relu', variable_name=scope + 'conv_W')
B = bias_variable([kernal[-1]], variable_name=scope + 'conv_B')
conv = conv3d(x, W) + B
conv = normalizationlayer(conv, is_train=phase, height=height, width=width, image_z=image_z, norm_type='group',
scope=scope)
conv = tf.nn.relu(conv)
return conv
def attngatingblock(x, g, inputfilters, outfilters, scale_factor, phase, image_z=None, height=None, width=None,
scope=None):
"""
take g which is the spatially smaller signal, do a conv to get the same number of feature channels as x (bigger spatially)
do a conv on x to also get same feature channels (theta_x)
then, upsample g to be same size as x add x and g (concat_xg) relu, 1x1x1 conv, then sigmoid then upsample the final -
this gives us attn coefficients
:param x:
:param g:
:param inputfilters:
:param outfilters:
:param scale_factor:2
:param scope:
:return:
"""
with tf.name_scope(scope):
kernalx = (1, 1, 1, inputfilters, outfilters)
Wx = weight_xavier_init(shape=kernalx, n_inputs=kernalx[0] * kernalx[1] * kernalx[2] * kernalx[3],
n_outputs=kernalx[-1], activefunction='relu', variable_name=scope + 'conv_Wx')
Bx = bias_variable([kernalx[-1]], variable_name=scope + 'conv_Bx')
theta_x = conv3d(x, Wx, scale_factor) + Bx
kernalg = (1, 1, 1, inputfilters, outfilters)
Wg = weight_xavier_init(shape=kernalg, n_inputs=kernalg[0] * kernalg[1] * kernalg[2] * kernalg[3],
n_outputs=kernalg[-1], activefunction='relu', variable_name=scope + 'conv_Wg')
Bg = bias_variable([kernalg[-1]], variable_name=scope + 'conv_Bg')
phi_g = conv3d(g, Wg) + Bg
add_xg = resnet_Add(theta_x, phi_g)
act_xg = tf.nn.relu(add_xg)
kernalpsi = (1, 1, 1, outfilters, 1)
Wpsi = weight_xavier_init(shape=kernalpsi, n_inputs=kernalpsi[0] * kernalpsi[1] * kernalpsi[2] * kernalpsi[3],
n_outputs=kernalpsi[-1], activefunction='relu', variable_name=scope + 'conv_Wpsi')
Bpsi = bias_variable([kernalpsi[-1]], variable_name=scope + 'conv_Bpsi')
psi = conv3d(act_xg, Wpsi) + Bpsi
sigmoid_psi = tf.nn.sigmoid(psi)
upsample_psi = upsample3d(sigmoid_psi, scale_factor=scale_factor, scope=scope + "resampler")
# Attention: upsample_psi * x
gat_x = tf.multiply(upsample_psi, x)
kernal_gat_x = (1, 1, 1, outfilters, outfilters)
Wgatx = weight_xavier_init(shape=kernal_gat_x,
n_inputs=kernal_gat_x[0] * kernal_gat_x[1] * kernal_gat_x[2] * kernal_gat_x[3],
n_outputs=kernal_gat_x[-1], activefunction='relu',
variable_name=scope + 'conv_Wgatx')
Bgatx = bias_variable([kernalpsi[-1]], variable_name=scope + 'conv_Bgatx')
gat_x_out = conv3d(gat_x, Wgatx) + Bgatx
gat_x_out = normalizationlayer(gat_x_out, is_train=phase, height=height, width=width, image_z=image_z,
norm_type='group', scope=scope)
return gat_x_out
def positionAttentionblock(x, inputfilters, outfilters, kernal_size=1, scope=None):
"""
Position attention module
:param x:
:param inputfilters:inputfilter number
:param outfilters:outputfilter number
:param scope:
:return:
"""
with tf.name_scope(scope):
m_batchsize, Z, H, W, C = x.get_shape().as_list()
kernalquery = (kernal_size, kernal_size, kernal_size, inputfilters, outfilters)
Wquery = weight_xavier_init(shape=kernalquery,
n_inputs=kernalquery[0] * kernalquery[1] * kernalquery[2] * kernalquery[3],
n_outputs=kernalquery[-1], activefunction='relu',
variable_name=scope + 'conv_Wquery')
Bquery = bias_variable([kernalquery[-1]], variable_name=scope + 'conv_Bquery')
query_conv = conv3d(x, Wquery) + Bquery
query_conv_new = tf.reshape(query_conv, [-1, Z * H * W, C])
kernalkey = (kernal_size, kernal_size, kernal_size, inputfilters, outfilters)
Wkey = weight_xavier_init(shape=kernalkey, n_inputs=kernalkey[0] * kernalkey[1] * kernalkey[2] * kernalkey[3],
n_outputs=kernalkey[-1], activefunction='relu', variable_name=scope + 'conv_Wkey')
Bkey = bias_variable([kernalkey[-1]], variable_name=scope + 'conv_Bkey')
key_conv = conv3d(x, Wkey) + Bkey
key_conv_new = tf.reshape(key_conv, [-1, Z * H * W, C])
# OOM,such as 512x512x32 then matric is 8388608x8388608
key_conv_new = tf.transpose(key_conv_new, [0, 2, 1])
# (2,2,2,3)*(2,2,3,4)=(2,2,2,4),(2,2,3)*(2,3,4)=(2,2,4)
energy = tf.matmul(query_conv_new, key_conv_new) # (m_batchsize,Z*H*W,Z*H*W)
attention = tf.nn.softmax(energy, -1)
kernalproj = (kernal_size, kernal_size, kernal_size, inputfilters, outfilters)
Wproj = weight_xavier_init(shape=kernalproj,
n_inputs=kernalproj[0] * kernalproj[1] * kernalproj[2] * kernalproj[3],
n_outputs=kernalproj[-1], activefunction='relu', variable_name=scope + 'conv_Wproj')
Bproj = bias_variable([kernalproj[-1]], variable_name=scope + 'conv_Bproj')
proj_value = conv3d(x, Wproj) + Bproj
proj_value_new = tf.reshape(proj_value, [-1, Z * H * W, C])
out = tf.matmul(attention, proj_value_new) # (m_batchsize,Z*H*W,C)
out_new = tf.reshape(out, [-1, Z, H, W, C])
out_new = resnet_Add(out_new, x)
return out_new
def channelAttentionblock(x, scope=None):
"""
Channel attention module
:param x:input
:param scope: scope name
:return:channelattention result
"""
with tf.name_scope(scope):
m_batchsize, Z, H, W, C = x.get_shape().as_list()
proj_query = tf.reshape(x, [-1, Z * H * W, C])
proj_key = tf.reshape(x, [-1, Z * H * W, C])
proj_query = tf.transpose(proj_query, [0, 2, 1])
energy = tf.matmul(proj_query, proj_key) # (-1,C,C)
attention = tf.nn.softmax(energy, -1) # (-1,C,C)
proj_value = tf.reshape(x, [-1, Z * H * W, C])
proj_value = tf.transpose(proj_value, [0, 2, 1])
out = tf.matmul(attention, proj_value) # (-1,C,Z*H*W)
out = tf.reshape(out, [-1, Z, H, W, C])
out = resnet_Add(out, x)
return out
def NonLocalBlock(input_x, phase, image_z=None, image_height=None, image_width=None, scope=None):
"""
Non-local netural network
:param input_x:
:param out_channels:
:param scope:
:return:
"""
batchsize, dimensizon, height, width, out_channels = input_x.get_shape().as_list()
with tf.name_scope(scope):
kernal_thela = (1, 1, 1, out_channels, out_channels // 2)
W_thela = weight_xavier_init(shape=kernal_thela,
n_inputs=kernal_thela[0] * kernal_thela[1] * kernal_thela[2] * kernal_thela[3],
n_outputs=kernal_thela[-1], activefunction='relu',
variable_name=scope + 'conv_W_thela')
B_thela = bias_variable([kernal_thela[-1]], variable_name=scope + 'conv_B_thela')
thela = conv3d(input_x, W_thela) + B_thela
thela = normalizationlayer(thela, is_train=phase, height=image_height, width=image_width, image_z=image_z,
norm_type='group', scope=scope + "NonLocalbn1")
kernal_phi = (1, 1, 1, out_channels, out_channels // 2)
W_phi = weight_xavier_init(shape=kernal_phi,
n_inputs=kernal_phi[0] * kernal_phi[1] * kernal_phi[2] * kernal_phi[3],
n_outputs=kernal_phi[-1], activefunction='relu',
variable_name=scope + 'conv_W_phi')
B_phi = bias_variable([kernal_phi[-1]], variable_name=scope + 'conv_B_phi')
phi = conv3d(input_x, W_phi) + B_phi
phi = normalizationlayer(phi, is_train=phase, height=image_height, width=image_width, image_z=image_z,
norm_type='group', scope=scope + "NonLocalbn2")
kernal_g = (1, 1, 1, out_channels, out_channels // 2)
W_g = weight_xavier_init(shape=kernal_g,
n_inputs=kernal_g[0] * kernal_g[1] * kernal_g[2] * kernal_g[3],
n_outputs=kernal_g[-1], activefunction='relu',
variable_name=scope + 'conv_W_g')
B_g = bias_variable([kernal_g[-1]], variable_name=scope + 'conv_B_g')
g = conv3d(input_x, W_g) + B_g
g = normalizationlayer(g, is_train=phase, height=image_height, width=image_width, image_z=image_z,
norm_type='group', scope=scope + "NonLocalbn3")
g_x = tf.reshape(g, [-1, dimensizon * height * width, out_channels // 2])
theta_x = tf.reshape(thela, [-1, dimensizon * height * width, out_channels // 2])
phi_x = tf.reshape(phi, [-1, dimensizon * height * width, out_channels // 2])
phi_x = tf.transpose(phi_x, [0, 2, 1])
f = tf.matmul(theta_x, phi_x)
f_softmax = tf.nn.softmax(f, -1)
y = tf.matmul(f_softmax, g_x)
y = tf.reshape(y, [-1, dimensizon, height, width, out_channels // 2])
kernal_y = (1, 1, 1, out_channels // 2, out_channels)
W_y = weight_xavier_init(shape=kernal_y,
n_inputs=kernal_y[0] * kernal_y[1] * kernal_y[2] * kernal_y[3],
n_outputs=kernal_y[-1], activefunction='relu',
variable_name=scope + 'conv_W_y')
B_y = bias_variable([kernal_y[-1]], variable_name=scope + 'conv_B_y')
w_y = conv3d(y, W_y) + B_y
w_y = normalizationlayer(w_y, is_train=phase, height=image_height, width=image_width, image_z=image_z,
norm_type='group', scope=scope + "NonLocalbn4")
z = resnet_Add(input_x, w_y)
return z
def conv_bn_relu_drop(x, kernal, phase, drop, image_z=None, height=None, width=None, scope=None):
"""
conv+bn+relu+drop
:param x:
:param kernal:
:param phase:
:param drop:
:param image_z:
:param height:
:param width:
:param scope:
:return:
"""
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1], activefunction='relu', variable_name=scope + 'conv_W')
B = bias_variable([kernal[-1]], variable_name=scope + 'conv_B')
conv = conv3d(x, W) + B
conv = normalizationlayer(conv, is_train=phase, height=height, width=width, image_z=image_z, norm_type='group',
scope=scope)
conv = tf.nn.dropout(tf.nn.leaky_relu(conv), drop)
return conv
def down_sampling(x, kernal, phase, drop, image_z=None, height=None, width=None, scope=None):
"""
downsampling with conv stride=2
:param x:
:param kernal:
:param phase:
:param drop:
:param image_z:
:param height:
:param width:
:param scope:
:return:
"""
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1],
activefunction='relu', variable_name=scope + 'W')
B = bias_variable([kernal[-1]], variable_name=scope + 'B')
conv = conv3d(x, W, 2) + B
conv = normalizationlayer(conv, is_train=phase, height=height, width=width, image_z=image_z, norm_type='group',
scope=scope)
conv = tf.nn.dropout(tf.nn.leaky_relu(conv), drop)
return conv
def deconv_relu(x, kernal, samefeture=False, scope=None):
"""
deconv+relu
:param x:
:param kernal:
:param samefeture:
:param scope:
:return:
"""
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[-1],
n_outputs=kernal[-2], activefunction='relu', variable_name=scope + 'W')
B = bias_variable([kernal[-2]], variable_name=scope + 'B')
conv = deconv3d(x, W, samefeture, True) + B
conv = tf.nn.leaky_relu(conv)
return conv
def conv_sigmod(x, kernal, scope=None):
"""
conv_sigmod
:param x:
:param kernal:
:param scope:
:return:
"""
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1], activefunction='sigomd', variable_name=scope + 'W')
B = bias_variable([kernal[-1]], variable_name=scope + 'B')
conv = conv3d(x, W) + B
conv = tf.nn.sigmoid(conv)
return conv
