def get_nn(self, input_placer, gold_placer, keep_prob,
learn_rate_placer, wd):
self.x = input_placer
self.y_ = gold_placer
self.l_rate = learn_rate_placer
self.keep_prob = keep_prob
self.para_dict = FbUpBN2D(is_training = self.train_flag).generate_flow(input_placer,
keep_prob,
wd = wd)
self.y_conv = self.para_dict['y_conv']
self.y_res = self.para_dict['y_res']
if self.train_flag:
self.y_conv_all = self.y_conv
self.y_conv_2x = dilation3D(self.y_conv_all)
self.y_res_all = self.y_res
self.shape = tf.shape(self.y_res)
width = self.shape[1]
self.y_conv = self.y_conv[:,2,2,2,:]
self.y_res = self.y_res[:,2,2,2]
cross_entropy_mean = -tf.reduce_mean(self.y_ * tf.log(self.y_conv))
tf.add_to_collection('losses', cross_entropy_mean)
cross_entropy = tf.add_n(tf.get_collection('losses'), name='total_loss')
self.train_step = tf.train.AdamOptimizer(self.l_rate).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(self.y_conv,1), tf.argmax(self.y_,1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
else:
self.y_conv_all = self.y_conv
self.y_conv_2x = dilation3D(self.y_conv_all,tf.shape(input_placer))
self.y_res_all = tf.expand_dims(self.y_res,4)
self.y_res_2x = tf.squeeze(dilation3D(self.y_res_all,tf.shape(input_placer)),[4])
def get_nn(self, input_placer, gold_placer, keep_prob, learn_rate_placer,
wd):
self.x = input_placer
self.y_ = gold_placer
self.l_rate = learn_rate_placer
self.keep_prob = keep_prob
self.para_dict = dict()
para_dict = FbUp().generate_flow(input_placer, keep_prob, wd=wd)
self.y_conv = para_dict['y_conv']
self.y_res = para_dict['y_res']
if self.train_flag:
self.y_conv_all = self.y_conv
self.y_conv_2x = dilation3D(self.y_conv_all)
self.y_res_all = self.y_res
self.shape = tf.shape(self.y_res)
width = self.shape[1]
self.y_conv = self.y_conv[:, 3, 3, 3, :]
self.y_res = self.y_res[:, 3, 3, 3]
cross_entropy_mean = -tf.reduce_mean(self.y_ * tf.log(self.y_conv))
tf.add_to_collection('losses', cross_entropy_mean)
cross_entropy = tf.add_n(tf.get_collection('losses'),
name='total_loss')
self.train_step = tf.train.AdamOptimizer(
self.l_rate).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(self.y_conv, 1),
tf.argmax(self.y_, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction,
"float"))
else:
self.y_conv_all = self.y_conv
self.y_conv_2x = dilation3D(self.y_conv_all)
self.y_res_all = self.y_res
self.y_res_2x = dilation3D(self.y_res_all)
self.para_dict['y_conv'] = self.y_conv_2x
self.para_dict['y_res'] = self.y_res_2x
self.para_dict['h_conv4'] = para_dict['h_conv4']
self.para_dict['h_pool2'] = para_dict['h_pool2']
self.para_dict['x'] = para_dict['x']
self.para_dict['h_conv3'] = para_dict['h_conv3']
self.para_dict['h_conv2'] = para_dict['h_conv2']
self.para_dict['h_conv_out'] = para_dict['h_conv_out']
def get_nn(self,
input_placer,
gold_placer,
learn_rate_placer,
keep_prob,
window_size,
conv_1_neuron_num=256,
conv_2_neuron_num=128,
conv_3_neuron_num=64,
conv_4_neuron_num=64,
conv_5_neuron_num=64,
conv_1=1,
conv_2=3,
conv_3=3,
conv_4=3,
conv_5=1,
wd=1e-9,
wdo=1e-9):
#create sub classes
self.x = input_placer
self.y_ = gold_placer
self.l_rate = learn_rate_placer
self.keep_prob = keep_prob
fg_dict = dict()
y_conv_combined = []
y_res_combined = []
for angle_xy in range(30, 181, 30):
for angle_xz in range(30, 181, 30):
print("Create FG NN %d %d" % (angle_xy, angle_xz))
fg_dict[(angle_xy,angle_xz)] = \
FbUpTrain(self.sess,input_placer,gold_placer,\
keep_prob,learn_rate_placer,\
angle_xy = angle_xy, \
angle_xz = angle_xz,
train_flag = False,
wd = 0e-9)
fg_variables = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope="fg_" + str(angle_xy) + '_' + str(angle_xz))
fg_dict[(angle_xy,angle_xz)].saver = \
tf.train.Saver(fg_variables)
#fg_dict[(angle_xy,angle_xz)].saver.restore(self.sess,\
# model_path + '/model_foreground_'+str(angle_xy)+'_'+str(angle_xz)+'.ckpt')
y_conv_combined.append(
tf.expand_dims(
fg_dict[(angle_xy, angle_xz)].para_dict['h_conv4'], 5))
#y_res_combined.append(
# tf.expand_dims(fg_dict[(angle_xy,angle_xz)].y_res,4))
FEAT_NUM = 32
self.fg_dict = fg_dict
y_conv_combined = tf.concat(5, y_conv_combined)
shape = tf.shape(y_conv_combined)
dir_out = tf.reshape(
y_conv_combined,
[shape[0], shape[1], shape[2], shape[3], 36 * FEAT_NUM])
#dir_out = tf.reshape(y_conv_combined,[-1,window_size,window_size,window_size,36 * FEAT_NUM])
self.dir_out = dir_out
with tf.variable_scope("transfer"):
#h_pool2 = self.pool_3x3(h_conv2)
with tf.variable_scope("transfer_layer_1"):
W_conv1 = self.nn.weight_variable(
[conv_1, conv_1, conv_1, 36 * FEAT_NUM, conv_1_neuron_num],
wd, 1**3 * 36 * FEAT_NUM, conv_1**3 * conv_1_neuron_num)
b_conv1 = self.nn.bias_variable([conv_1_neuron_num])
h_conv1 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(dir_out, W_conv1), b_conv1))
with tf.variable_scope("transfer_layer_2"):
W_conv2 = self.nn.weight_variable([
conv_2, conv_2, conv_2, conv_1_neuron_num,
conv_2_neuron_num
], wd, conv_1**3 * conv_1_neuron_num, conv_2**3 *
conv_2_neuron_num)
b_conv2 = self.nn.bias_variable([conv_2_neuron_num])
h_conv2 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(h_conv1, W_conv2), b_conv2))
with tf.variable_scope("transfer_layer_3"):
W_conv3 = self.nn.weight_variable([
conv_3, conv_3, conv_3, conv_2_neuron_num,
conv_3_neuron_num
], wd, conv_2**3 * conv_2_neuron_num, conv_3**3 *
conv_3_neuron_num)
b_conv3 = self.nn.bias_variable([conv_3_neuron_num])
h_conv3 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(h_conv2, W_conv3), b_conv3))
with tf.variable_scope("transfer_layer_4"):
W_conv4 = self.nn.weight_variable([
conv_4, conv_4, conv_4, conv_3_neuron_num,
conv_4_neuron_num
], wd, conv_3**3 * conv_3_neuron_num, conv_4**3 *
conv_4_neuron_num)
b_conv4 = self.nn.bias_variable([conv_4_neuron_num])
h_conv4 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(h_conv3, W_conv4), b_conv4))
with tf.variable_scope("transfer_layer_5"):
W_conv5 = self.nn.weight_variable([
conv_5, conv_5, conv_5, conv_4_neuron_num,
conv_5_neuron_num
], wd, conv_4**3 * conv_4_neuron_num, conv_5**3 *
conv_5_neuron_num)
b_conv5 = self.nn.bias_variable([conv_5_neuron_num])
h_conv5 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(h_conv4, W_conv5), b_conv5))
weight_decay = tf.mul(tf.nn.l2_loss(h_conv5),
wd,
name='weight_loss')
tf.add_to_collection('losses', weight_decay)
with tf.variable_scope("transfer_final"):
W_fc1 = self.nn.weight_variable(
[1, 1, 1, conv_5_neuron_num, 2], wd=wdo)
b_fc1 = self.nn.bias_variable([2])
h_fc1 = tf.nn.relu(
tf.nn.bias_add(self.nn.conv2d(h_conv5, W_fc1), b_fc1))
if self.predict_flag:
final_shape = tf.shape(h_fc1)
self.y_conv = tf.nn.softmax(h_fc1[:, (final_shape[1] - 1) / 2,
(final_shape[2] - 1) / 2,
(final_shape[3] - 1) / 2, :])
#self.y_conv = tf.nn.softmax(h_fc1[:,(window_size-1)//2,(window_size-1)//2,(window_size-1)//2,:])
self.y_res = tf.argmax(self.y_conv, 1)
else:
curr_shape = tf.shape(h_fc1)
reshaped_conv = tf.reshape(h_fc1, [-1, 2])
softmaxed_reshaped_conv = tf.nn.softmax(reshaped_conv)
self.y_conv = tf.reshape(softmaxed_reshaped_conv, curr_shape)
self.y_res = tf.argmax(self.y_conv, 4)
self.y_conv_2x = dilation3D(self.y_conv)
tmp = tf.expand_dims(self.y_res, 4)
self.y_res_2x = tf.squeeze(dilation3D(tmp))
if self.valid_flag:
cross_entropy_mean = -tf.reduce_mean(self.y_ * tf.log(self.y_conv))
tf.add_to_collection('losses', cross_entropy_mean)
cross_entropy = tf.add_n(tf.get_collection('losses'),
name='total_loss')
transfer_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
"transfer")
self.train_step = tf.train.AdamOptimizer(self.l_rate).minimize(
cross_entropy, var_list=transfer_vars)
correct_prediction = tf.equal(tf.argmax(self.y_conv, 1),
tf.argmax(self.y_, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction,
"float"))
else:
print("Transfer Not For Training Singly")
def get_nn(self,
input_placer,
gold_placer,
learn_rate_placer,
keep_prob,
window_size,
conv_1_neuron_num=256,
conv_2_neuron_num=128,
conv_3_neuron_num=64,
conv_4_neuron_num=64,
conv_5_neuron_num=64,
conv_6_neuron_num=64,
conv_7_neuron_num=64,
up_conv_1_neuron_num=256,
up_conv_1=3,
conv_1=1,
conv_2=3,
conv_3=3,
conv_4=3,
conv_5=3,
conv_6=3,
conv_7=1,
wd=1e-9,
wdo=1e-9):
if self.predict_flag:
padding = 'VALID'
else:
padding = 'SAME'
#create sub classes
self.x = input_placer
self.y_ = gold_placer
self.l_rate = learn_rate_placer
self.keep_prob = keep_prob
fg_dict = dict()
y_conv_combined = []
y_res_combined = []
for angle_xy in range(30, 181, 30):
for angle_xz in range(30, 181, 30):
print("Create FG NN %d %d" % (angle_xy, angle_xz))
fg_dict[(angle_xy,angle_xz)] = \
FbUpTrainBN(self.sess,input_placer,gold_placer,\
keep_prob,learn_rate_placer,\
angle_xy = angle_xy, \
angle_xz = angle_xz,
train_flag = False,
wd = 0e-9)
fg_variables = tf.get_collection(tf.GraphKeys.VARIABLES,
scope="fg_" + str(angle_xy) +
'_' + str(angle_xz))
fg_dict[(angle_xy,angle_xz)].saver = \
tf.train.Saver(fg_variables)
#fg_dict[(angle_xy,angle_xz)].saver.restore(self.sess,\
# model_path + '/model_foreground_'+str(angle_xy)+'_'+str(angle_xz)+'.ckpt')
y_conv_combined.append(
fg_dict[(angle_xy, angle_xz)].para_dict['h_conv3'])
#y_res_combined.append(
# tf.expand_dims(fg_dict[(angle_xy,angle_xz)].y_res,4))
FEAT_NUM = 16
self.fg_dict = fg_dict
y_conv_combined = tf.concat(4, y_conv_combined)
dir_out = y_conv_combined
self.dir_out = dir_out
with tf.variable_scope("transfer"):
with tf.variable_scope("transfer_layer_1"):
W_conv1 = self.nn.weight_variable(
[conv_1, conv_1, conv_1, 36 * FEAT_NUM, conv_1_neuron_num],
wd, 1**3 * 36 * FEAT_NUM, conv_1**3 * conv_1_neuron_num)
tmp_1 = self.nn.conv2d(dir_out, W_conv1, padding=padding)
h_conv1 = tf.nn.relu(
self.batch_norm_layer(tmp_1, self.predict_flag))
with tf.variable_scope("transfer_layer_2"):
W_conv2 = self.nn.weight_variable([
conv_2, conv_2, conv_2, conv_1_neuron_num,
conv_2_neuron_num
], wd, conv_1**3 * 2 * conv_1_neuron_num, conv_2**3 *
conv_2_neuron_num)
tmp_2 = self.nn.conv2d(h_conv1, W_conv2, padding=padding)
h_conv2 = tf.nn.relu(
self.batch_norm_layer(tmp_2, self.predict_flag))
"""
with tf.variable_scope("transfer_layer_3"):
W_conv3 = self.nn.weight_variable([conv_3,conv_3,conv_3,conv_2_neuron_num,conv_3_neuron_num],wd,conv_2 ** 3 * conv_2_neuron_num, conv_3 ** 3 * conv_3_neuron_num)
b_conv3 = self.nn.bias_variable([conv_3_neuron_num])
h_conv3 = tf.nn.relu(tf.nn.bias_add(self.nn.conv2d(h_conv2, W_conv3, padding = padding) , b_conv3))
with tf.variable_scope("transfer_layer_4"):
W_conv4 = self.nn.weight_variable([conv_4,conv_4,conv_4,conv_3_neuron_num,conv_4_neuron_num],wd,conv_3 ** 3 * conv_3_neuron_num,conv_4 ** 3 * conv_4_neuron_num)
b_conv4 = self.nn.bias_variable([conv_4_neuron_num])
if self.predict_flag:
residual_4 = h_conv3[:,1:8,1:8,1:8,:]
else:
residual_4 = h_conv3
h_conv4 = tf.nn.relu(tf.add(tf.nn.bias_add(self.nn.conv2d(h_conv3, W_conv4, padding = padding) , b_conv4),residual_4))
with tf.variable_scope("transfer_layer_5"):
W_conv5 = self.nn.weight_variable([conv_5,conv_5,conv_5,conv_4_neuron_num,conv_5_neuron_num],wd,conv_4 ** 3 * conv_4_neuron_num,conv_5 ** 3 * conv_5_neuron_num)
b_conv5 = self.nn.bias_variable([conv_5_neuron_num])
if self.predict_flag:
residual_5 = h_conv4[:,1:6,1:6,1:6,:]
else:
residual_5 = h_conv4
h_conv5 = tf.nn.relu(tf.add(tf.nn.bias_add(self.nn.conv2d(h_conv4, W_conv5, padding = padding) , b_conv5),residual_5))
with tf.variable_scope("transfer_layer_6"):
W_conv6 = self.nn.weight_variable([conv_6,conv_6,conv_6,conv_5_neuron_num,conv_6_neuron_num],wd,conv_5 ** 3 * conv_5_neuron_num,conv_6 ** 3 * conv_6_neuron_num)
b_conv6 = self.nn.bias_variable([conv_5_neuron_num])
if self.predict_flag:
residual_6 = h_conv5[:,1:4,1:4,1:4,:]
else:
residual_6 = h_conv5
h_conv6 = tf.nn.relu(tf.add(tf.nn.bias_add(self.nn.conv2d(h_conv5, W_conv6, padding = padding) , b_conv6),residual_6))
with tf.variable_scope("transfer_layer_7"):
W_conv7 = self.nn.weight_variable([conv_7,conv_7,conv_7,conv_6_neuron_num,conv_7_neuron_num],wd,conv_6 ** 3 * conv_6_neuron_num,conv_7 ** 3 * conv_7_neuron_num)
b_conv7 = self.nn.bias_variable([conv_7_neuron_num])
h_conv7 = tf.nn.relu(tf.nn.bias_add(self.nn.conv2d(h_conv6, W_conv7, padding = padding) , b_conv7))
weight_decay = tf.mul(tf.nn.l2_loss(h_conv7), wd, name='weight_loss')
tf.add_to_collection('losses', weight_decay)
"""
with tf.variable_scope("transfer_final"):
W_fc1 = self.nn.weight_variable(
[1, 1, 1, conv_2_neuron_num, 2], wd=wdo)
b_fc1 = self.nn.bias_variable([2])
h_fc1 = (tf.nn.bias_add(
self.nn.conv2d(h_conv2, W_fc1, padding=padding), b_fc1))
self.h_fc1 = h_fc1
if self.predict_flag:
final_shape = tf.shape(h_fc1)
self.final_shape = final_shape
#self.y_conv = tf.nn.softmax(h_fc1[:,(final_shape[1]-1)/2,(final_shape[2]-1)/2,(final_shape[3]-1)/2,:])
self.y_conv = tf.nn.softmax(h_fc1[:, 2, 2, 2, :])
self.y_res = tf.argmax(self.y_conv, 1)
conv_center3x3 = h_fc1
reshaped_conv_center3x3 = tf.reshape(conv_center3x3, [-1, 2])
softmaxed_reshaped_conv_center3x3 = tf.nn.softmax(
reshaped_conv_center3x3)
self.y_conv_center3x3 = tf.reshape(
softmaxed_reshaped_conv_center3x3, tf.shape(conv_center3x3))
else:
input_shape = tf.shape(input_placer)
curr_shape = tf.shape(h_fc1)
reshaped_conv = tf.reshape(h_fc1, [-1, 2])
softmaxed_reshaped_conv = tf.nn.softmax(reshaped_conv)
self.y_conv = tf.reshape(softmaxed_reshaped_conv, curr_shape)
self.y_res = tf.argmax(self.y_conv, 4)
self.y_conv_2x = dilation3D(self.y_conv, input_shape)
tmp = tf.expand_dims(self.y_res, 4)
self.y_res_2x = tf.squeeze(dilation3D(tmp, input_shape))
if self.valid_flag:
cross_entropy_mean = -tf.reduce_mean(self.y_ * tf.log(self.y_conv))
#For Centerline
center_sum = tf.reduce_sum(self.y_conv_center3x3, [1, 2, 3])
#self.line_penalty = tf.mul(wdo,tf.reduce_sum(tf.mul(self.y_[:,1],tf.abs(tf.sub(center_sum[:,1],3)))))
#self.ce = cross_entropy_mean
#tf.add_to_collection('losses', self.line_penalty)
tf.add_to_collection('losses', cross_entropy_mean)
cross_entropy = tf.add_n(tf.get_collection('losses'),
name='total_loss')
transfer_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
"transfer")
self.train_step = tf.train.AdamOptimizer(self.l_rate).minimize(
cross_entropy, var_list=transfer_vars)
correct_prediction = tf.equal(tf.argmax(self.y_conv, 1),
tf.argmax(self.y_, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction,
"float"))
else:
print("Transfer Not For Training Singly")