def conv_layer(inputT, shape, train_phase, activation=True, Before_ReLu=True, name=None):
width = shape[0]
hight = shape[1]
in_channel = shape[2]
out_channel = shape[3]
Input_channel = inputT.get_shape().as_list()[3]
def ReLu(inputs, train_phase, name):
x = tf.nn.relu(batch_norm_layer(inputs, train_phase, name))
return x
with tf.variable_scope(name) as scope:
if Before_ReLu is True:
biases_Input = _variable_on_cpu('biases_Input', [Input_channel], tf.constant_initializer(0.0))
biass = tf.nn.bias_add(inputT, biases_Input)
inputT = ReLu(biass, train_phase, scope.name +'_BeforeRelu')
kernel = _variable_with_weight_decay('ort_weights', [width, hight, in_channel, out_channel], initializer=orthogonal_initializer(), wd=None)
conv = tf.nn.conv2d(inputT, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [out_channel], tf.constant_initializer(0.0))
bias = tf.nn.bias_add(conv, biases)
if activation is True:
conv_out = ReLu(bias, train_phase, scope.name)
else:
conv_out = batch_norm_layer(bias, train_phase, scope.name)
return conv_out
def fully_connected(inputT, name, out_channel, activation_fn=tf.nn.relu):
in_channel = inputT.get_shape().as_list()[-1]
with tf.variable_scope(name):
kernel = _variable_with_weight_decay('weights', shape = [in_channel, out_channel], initializer=orthogonal_initializer(),wd=0.0005) #tf.contrib.layers.xavier_initializer() or msra_initializer(3, in_channel)
biases = _variable_on_cpu('biases', [out_channel], tf.constant_initializer(0.0))
out = tf.nn.bias_add(tf.matmul(inputT, kernel), biases)
out = activation_fn(out)
_activation_summary(out)
return out
def conv_layer_with_bn(inputT, shape, train_phase, activation=True, name=None):
in_channel = shape[2]
out_channel = shape[3]
k_size = shape[0]
with tf.variable_scope(name) as scope:
kernel = _variable_with_weight_decay('ort_weights', shape=shape, initializer=orthogonal_initializer(), wd=None)
conv = tf.nn.conv2d(inputT, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [out_channel], tf.constant_initializer(0.0))
bias = tf.nn.bias_add(conv, biases)
if activation is True:
conv_out = tf.nn.relu(batch_norm_layer(bias, train_phase, scope.name))
else:
conv_out = batch_norm_layer(bias, train_phase, scope.name)
return conv_out
def conv_layer_with_bn(inputT, shape, train_phase, name=None, activation_fn=tf.nn.relu):
#shape = [k_size, k_size, in_channel, out_channel]
out_channel = shape[3]
with tf.variable_scope(name) as scope:
kernel = _variable_with_weight_decay('weights', shape=shape, initializer=orthogonal_initializer(), wd=None) #orthogonal_initializer() or xaviar initialization
conv = tf.nn.conv2d(inputT, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [out_channel], tf.constant_initializer(0.0))
bias = tf.nn.bias_add(conv, biases)
#batch_norm = batch_norm_layer(bias, train_phase, scope.name)
#conv_out = tf.nn.relu(batch_norm)
conv_out = tf.nn.relu(bias)
_activation_summary(inputT)
return conv_out
def inference(images, labels, batch_size, phase_train):
# norm1
norm1 = tf.nn.lrn(images, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75,
name='norm1')
# conv1
conv1 = conv_layer_with_bn(norm1, [3, 3, images.get_shape().as_list()[3], 32], phase_train, name="conv1")
# pool1
pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool1')
# conv2
conv2 = conv_layer_with_bn(pool1, [3, 3, 32, 32], phase_train, name="conv2")
# pool2
pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool2')
""" End of encoder """
""" start upsample """
# upsample2
# upsample2 = upsample_with_pool_indices(conv_decode3, pool2_indices, conv_decode3.get_shape(), scale=2, name='upsample2')
upsample2= deconv_layer(pool2, [2, 2, 32, 32], [batch_size, 45, 60, 32], 2, "up2")
# decode 2
conv_decode2 = conv_layer_with_bn(upsample2, [3, 3, 32, 32], phase_train, False, name="conv_decode2")
# upsample1
# upsample1 = upsample_with_pool_indices(conv_decode2, pool1_indices, conv_decode2.get_shape(), scale=2, name='upsample1')
upsample1= deconv_layer(conv_decode2, [2, 2, 32, 32], [batch_size, 90, 120, 32], 2, "up1")
# decode4
conv_decode1 = conv_layer_with_bn(upsample1, [3, 3, 32, 32], phase_train, False, name="conv_decode1")
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 32, NUM_CLASSES],
initializer=msra_initializer(1, 32),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [NUM_CLASSES], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
loss = cal_loss(conv_classifier, labels)
return loss, logit
def inference_with_endpoints(images, labels, batch_size, phase_train):
end_points = {}
# norm1
norm1 = tf.nn.lrn(images,
depth_radius=5,
bias=1.0,
alpha=0.0001,
beta=0.75,
name='norm1')
# conv1
conv1 = conv_layer_with_bn(
norm1, [7, 7, images.get_shape().as_list()[3], 64],
phase_train,
name="conv1")
end_points['segnet/conv1'] = conv1
# pool1
# pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
# padding='SAME', name='pool1')
pool1 = tf.nn.max_pool(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# tf.nn.max_pool(value, ksize, strides, padding, data_format='NHWC', name=None)
# conv2
conv2 = conv_layer_with_bn(pool1, [7, 7, 64, 64],
phase_train,
name="conv2")
end_points['segnet/conv2'] = conv2
# pool2
# pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2, ksize=[1, 2, 2, 1],
# strides=[1, 2, 2, 1], padding='SAME', name='pool2')
pool2 = tf.nn.max_pool(conv2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# conv3
conv3 = conv_layer_with_bn(pool2, [7, 7, 64, 64],
phase_train,
name="conv3")
end_points['segnet/conv3'] = conv3
# pool3
# pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3, ksize=[1, 2, 2, 1],
# strides=[1, 2, 2, 1], padding='SAME', name='pool3')
pool3 = tf.nn.max_pool(conv3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# conv4
conv4 = conv_layer_with_bn(pool3, [7, 7, 64, 64],
phase_train,
name="conv4")
end_points['segnet/conv4'] = conv4
# pool4
# pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4, ksize=[1, 2, 2, 1],
# strides=[1, 2, 2, 1], padding='SAME', name='pool4')
pool4 = tf.nn.max_pool(conv4,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
""" End of encoder """
""" start upsample """
# upsample4
# Need to change when using different dataset out_w, out_h
# upsample4 = upsample_with_pool_indices(pool4, pool4_indices, pool4.get_shape(), out_w=45, out_h=60, scale=2, name='upsample4')
upsample4 = deconv_layer(pool4, [2, 2, 64, 64], [batch_size, 64, 64, 64],
2, "up4")
# decode 4
conv_decode4 = conv_layer_with_bn(upsample4, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode4")
end_points['segnet/conv_decode4'] = conv_decode4
# upsample 3
# upsample3 = upsample_with_pool_indices(conv_decode4, pool3_indices, conv_decode4.get_shape(), scale=2, name='upsample3')
upsample3 = deconv_layer(conv_decode4, [2, 2, 64, 64],
[batch_size, 128, 128, 64], 2, "up3")
# decode 3
conv_decode3 = conv_layer_with_bn(upsample3, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode3")
end_points['segnet/conv_decode3'] = conv_decode3
# upsample2
# upsample2 = upsample_with_pool_indices(conv_decode3, pool2_indices, conv_decode3.get_shape(), scale=2, name='upsample2')
upsample2 = deconv_layer(conv_decode3, [2, 2, 64, 64],
[batch_size, 256, 256, 64], 2, "up2")
# decode 2
conv_decode2 = conv_layer_with_bn(upsample2, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode2")
end_points['segnet/conv_decode2'] = conv_decode2
# upsample1
# upsample1 = upsample_with_pool_indices(conv_decode2, pool1_indices, conv_decode2.get_shape(), scale=2, name='upsample1')
upsample1 = deconv_layer(conv_decode2, [2, 2, 64, 64],
[batch_size, 512, 512, 64], 2, "up1")
# decode4
conv_decode1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode1")
end_points['segnet/conv_decode1'] = conv_decode1
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, NUM_CLASSES],
initializer=msra_initializer(
1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
end_points['segnet/conv_classifier'] = conv
biases = _variable_on_cpu('biases', [NUM_CLASSES],
tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
end_points['segnet/logits'] = logit
# loss = cal_loss(conv_classifier, labels)
return logit, end_points
def inference(images, labels, batch_size, phase_train):
# norm1
norm1 = tf.nn.lrn(images, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75,
name='norm1')
# conv1
conv1 = conv_layer_with_bn(norm1, [7, 7, images.get_shape().as_list()[3], 64], phase_train, name="conv1")
# pool1
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool1')
# conv2
conv2 = conv_layer_with_bn(pool1, [7, 7, 64, 64], phase_train, name="conv2")
# pool2
pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool2')
# conv3
conv3 = conv_layer_with_bn(pool2, [7, 7, 64, 64], phase_train, name="conv3")
# pool3
pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool3')
# conv4
conv4 = conv_layer_with_bn(pool3, [7, 7, 64, 64], phase_train, name="conv4")
# pool4
pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool4')
""" End of encoder """
""" start upsample """
# upsample4
# Need to change when using different dataset out_w, out_h
# upsample4 = upsample_with_pool_indices(pool4, pool4_indices, pool4.get_shape(), out_w=45, out_h=60, scale=2, name='upsample4')
upsample4 = deconv_layer(pool4, [2, 2, 64, 64], [batch_size, 45, 60, 64], 2, "up4")
# decode 4
conv_decode4 = conv_layer_with_bn(upsample4, [7, 7, 64, 64], phase_train, False, name="conv_decode4")
# upsample 3
# upsample3 = upsample_with_pool_indices(conv_decode4, pool3_indices, conv_decode4.get_shape(), scale=2, name='upsample3')
upsample3= deconv_layer(conv_decode4, [2, 2, 64, 64], [batch_size, 90, 120, 64], 2, "up3")
# decode 3
conv_decode3 = conv_layer_with_bn(upsample3, [7, 7, 64, 64], phase_train, False, name="conv_decode3")
# upsample2
# upsample2 = upsample_with_pool_indices(conv_decode3, pool2_indices, conv_decode3.get_shape(), scale=2, name='upsample2')
upsample2= deconv_layer(conv_decode3, [2, 2, 64, 64], [batch_size, 180, 240, 64], 2, "up2")
# decode 2
conv_decode2 = conv_layer_with_bn(upsample2, [7, 7, 64, 64], phase_train, False, name="conv_decode2")
# upsample1
# upsample1 = upsample_with_pool_indices(conv_decode2, pool1_indices, conv_decode2.get_shape(), scale=2, name='upsample1')
upsample1= deconv_layer(conv_decode2, [2, 2, 64, 64], [batch_size, 360, 480, 64], 2, "up1")
# decode4
conv_decode1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64], phase_train, False, name="conv_decode1")
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, NUM_CLASSES],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [NUM_CLASSES], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
loss = cal_loss(conv_classifier, labels)
return loss, logit
def inference(images, labels, batch_size, phase_train):
# norm1
with tf.name_scope(name="inference"):
with tf.variable_scope("norm-1"):
norm1 = tf.nn.lrn(images, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75,
name='norm1')
tf.summary.histogram('/norm-1', norm1)
# conv1
with tf.variable_scope("conv-1"):
conv1 = conv_layer_with_bn(norm1, [7, 7, images.get_shape().as_list()[3], 64], phase_train, name="conv1")
tf.summary.histogram('/conv-1', conv1)
## print(conv1.shape)
# pool1
## pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
## padding='SAME', name='pool1')
with tf.variable_scope("pool-1"):
pool1, pool1_indices = max_pool_2x2(conv1, name="pool1")
tf.summary.histogram('/pool-1', pool1)
## print(pool1.shape)
# conv2
with tf.variable_scope("conv-2"):
conv2 = conv_layer_with_bn(pool1, [7, 7, 64, 64], phase_train, name="conv2")
tf.summary.histogram('/conv-2', conv2)
# pool2
## pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2, ksize=[1, 2, 2, 1],
## strides=[1, 2, 2, 1], padding='SAME', name='pool2')
with tf.variable_scope("pool-2"):
pool2, pool2_indices = max_pool_2x2(conv2, name="pool2")
tf.summary.histogram('/pool-2', pool2)
## print(pool2.shape)
# conv3
with tf.variable_scope("conv-3"):
conv3 = conv_layer_with_bn(pool2, [7, 7, 64, 64], phase_train, name="conv3")
tf.summary.histogram('/conv-3', conv3)
# pool3
## pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3, ksize=[1, 2, 2, 1],
## strides=[1, 2, 2, 1], padding='SAME', name='pool3')
with tf.variable_scope("pool-3"):
pool3, pool3_indices = max_pool_2x2(conv3, name="pool3")
tf.summary.histogram('/pool-3', pool3)
## print(pool3.shape)
# conv4
with tf.variable_scope("conv-4"):
conv4 = conv_layer_with_bn(pool3, [7, 7, 64, 64], phase_train, name="conv4")
tf.summary.histogram('/conv-4', conv4)
# pool4
## pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4, ksize=[1, 2, 2, 1],
## strides=[1, 2, 2, 1], padding='SAME', name='pool4')
with tf.variable_scope("pool-4"):
pool4, pool4_indices = max_pool_2x2(conv4, name="pool4")
tf.summary.histogram('/pool-4', pool4)
print('pool4 = ',pool4.get_shape())
""" End of encoder """
""" Pyramid Pooling Module"""
with tf.variable_scope('PyP', dtype=tf.float32):
rcu01 = conv_layer(images, [3, 3, images.get_shape().as_list()[3], 64], phase_train, activation=True, Before_ReLu=True, name='RCU01')
rcu11 = tf.add(rcu01, conv_layer(rcu01, [3, 3, rcu01.get_shape().as_list()[3], 64], phase_train, name='RCU11'))
print('pool11 =' , images.get_shape() , ' and rcu11 =', rcu11.get_shape())
rcu02 = conv_layer(pool1, [3, 3, pool1.get_shape().as_list()[3], 32], phase_train, activation=True, Before_ReLu=True, name='RCU02')
rcu12 = tf.add(pool1, conv_layer(rcu02, [3, 3, rcu02.get_shape().as_list()[3], 64], phase_train, name='RCU12'))
rcu03 = conv_layer(pool2, [3, 3, pool2.get_shape().as_list()[3], 32], phase_train, activation=True, Before_ReLu=True, name='RCU03')
rcu13 = tf.add(pool2, conv_layer(rcu03, [3, 3, rcu03.get_shape().as_list()[3], 64], phase_train, name='RCU13'))
rcu04 = conv_layer(pool3, [3, 3, pool3.get_shape().as_list()[3], 32], phase_train, activation=True, Before_ReLu=True, name='RCU04')
rcu14 = tf.add(pool3, conv_layer(rcu04, [3, 3, rcu04.get_shape().as_list()[3], 64], phase_train, name='RCU14'))
rcu05 = conv_layer(pool4, [3, 3, pool4.get_shape().as_list()[3], 32], phase_train, activation=True, Before_ReLu=True, name='RCU05')
rcu15 = tf.add(pool4, conv_layer(rcu05, [3, 3, rcu05.get_shape().as_list()[3], 64], phase_train, name='RCU15'))
pypool1 = tf.image.resize_bilinear(rcu11, [120, 160])
pypool2 = tf.image.resize_bilinear(rcu12, [120, 160])
pypool3 = tf.image.resize_bilinear(rcu13, [120, 160])
pypool4 = tf.image.resize_bilinear(rcu14, [120, 160])
pypool5 = tf.image.resize_bilinear(rcu15, [120, 160])
print("PYpool4" , pypool4.get_shape())
pyp1 = tf.multiply(pypool1, pypool2, name='pyp1')
pyp2 = tf.multiply(pyp1, pypool3, name='pyp2')
pyp3 = tf.multiply(pyp2, pypool4, name='pyp3')
pyp4 = tf.multiply(pyp3, pypool5, name='pyp4')
## pyp = tf.stack(tf.reduce_mean([pypool1, pypool1, pypool1, pypool1, pypool1], 0))
print("pyp.shape = " , pyp1.get_shape())
print("pyp4.shape = " , pyp4.get_shape())
pyp_conv = conv_layer_with_bn(pyp4, [7, 7, pyp4.get_shape().as_list()[3], 64], phase_train, name='pyp_conv')
crp = tf.nn.relu(pyp_conv , name='crp')
pool_add_rcu = tf.add(tf.image.resize_bilinear(crp,[8, 10]), pool4)
print("pool_add_rcuF = " , pool_add_rcu.get_shape())
""" start upsample """
# upsample4
# Need to change when using different dataset out_w, out_h
with tf.variable_scope("upsample-4"):
upsample4 = deconv_layer(pool_add_rcu, [2, 2, 64, 64], [batch_size, 15, 20, 64], 2, "up4")
tf.summary.histogram('/upsample4', upsample4)
# decode 4
with tf.variable_scope("decode-4"):
conv_decode4 = conv_layer_with_bn(upsample4, [7, 7, 64, 64], phase_train, False, name="conv_decode4")
tf.summary.histogram('/conv_decode4', conv_decode4)
# upsample 3
with tf.variable_scope("upsample-3"):
upsample3= deconv_layer(conv_decode4, [2, 2, 64, 64], [batch_size, 30, 40, 64], 2, "up3")
tf.summary.histogram('/upsample3', upsample3)
# decode 3
with tf.variable_scope("decode-3"):
conv_decode3 = conv_layer_with_bn(upsample3, [7, 7, 64, 64], phase_train, False, name="conv_decode3")
tf.summary.histogram('/conv_decode3', conv_decode3)
# upsample2
with tf.variable_scope("updample-2"):
upsample2= deconv_layer(conv_decode3, [2, 2, 64, 64], [batch_size, 60, 80, 64], 2, "up2")
tf.summary.histogram('/upsample2', upsample2)
# decode 2
with tf.variable_scope("decode-2"):
conv_decode2 = conv_layer_with_bn(upsample2, [7, 7, 64, 64], phase_train, False, name="conv_decode2")
tf.summary.histogram('/conv_decode2', conv_decode2)
# upsample1
with tf.variable_scope("upsample-1"):
upsample1= deconv_layer(conv_decode2, [2, 2, 64, 64], [batch_size, 120, 160, 64], 2, "up1")
tf.summary.histogram('/upsample1', upsample1)
# decode1
with tf.variable_scope("decode-1"):
print('upsample = ', upsample1.get_shape())
conv_decode1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64], phase_train, False, name="conv_decode1")
tf.summary.histogram('/conv_decode1', conv_decode1)
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, NUM_CLASSES],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [NUM_CLASSES], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
loss = cal_loss(conv_classifier, labels)
tf.summary.histogram(scope.name, loss)
return loss, logit
def inference(images, labels, phase_train):
global num_classes, batch_size
#local normalization on input images
norm1 = tf.nn.lrn(images, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75,name='norm1')
#convolve input layer
conv1 = conv_layer_with_bn(norm1, [5, 5, images.get_shape().as_list()[3], 64], phase_train, name="conv1")
print conv1.shape
#residual set 1
conv2 = residual_block(conv1, "conv2", phase_train, "1")
conv3 = residual_block(conv2, "conv3", phase_train, "2")
conv4 = residual_block(conv3, "conv4", phase_train, "3")
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool1')
print pool1.shape
#residual block 2
conv5 = residual_block(pool1, "conv5", phase_train, "4")
conv6 = residual_block(conv5, "conv6", phase_train, "5")
conv7 = residual_block(conv6, "conv7", phase_train, "6")
pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv7, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool2')
print pool2.shape
#residual block 3
conv8 = residual_block(pool2, "conv8", phase_train, "7")
conv9 = residual_block(conv8, "conv9", phase_train, "8")
conv10 = residual_block(conv9, "conv10", phase_train, "9")
pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv9, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool3')
print pool3.shape
#residual block 4
conv11 = residual_block(pool3, "conv11", phase_train, "10")
conv12 = residual_block(conv11, "conv12", phase_train, "11")
conv13 = residual_block(conv12, "conv13", phase_train, "12")
pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv13, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool4')
print pool4.shape
#end of encoder - start of decoder
#residual block 5
upsample4 = deconv_layer(pool4, [2, 2, 64, 64], [batch_size, 30, 40, 64], 2, "up4")
conv_decode14 = residual_block_2(upsample4, "conv_decode14", phase_train, "13")
conv_decode15 = residual_block_2(conv_decode14, "conv_decode15", phase_train, "14")
conv_decode16 = residual_block_2(conv_decode15, "conv_decode16", phase_train, "15")
#residual block 6
upsample3 = deconv_layer(conv_decode16, [2, 2, 64, 64], [batch_size, 60, 80, 64], 2, "up3")
conv_decode17 = residual_block_2(upsample3, "conv_decode17", phase_train, "16")
conv_decode18 = residual_block_2(conv_decode17, "conv_decode18", phase_train, "17")
conv_decode19 = residual_block_2(conv_decode18, "conv_decode19", phase_train, "18")
#residual block 7
upsample2 = deconv_layer(conv_decode19, [2, 2, 64, 64], [batch_size, 120, 160, 64], 2, "up2")
conv_decode20 = residual_block_2(upsample2, "conv_decode20", phase_train, "19")
conv_decode21 = residual_block_2(conv_decode20, "conv_decode20", phase_train, "20")
conv_decode22 = residual_block_2(conv_decode21, "conv_decode21", phase_train, "21")
#residual block 8
upsample1 = deconv_layer(conv_decode22, [2, 2, 64, 64], [batch_size, 240, 320, 64], 2, "up1")
conv_decode23 = residual_block_2(upsample1, "conv_decode23", phase_train, "22")
conv_decode24 = residual_block_2(conv_decode23, "conv_decode24", phase_train, "23")
conv_decode25 = residual_block_2(conv_decode24, "conv_decode25", phase_train, "24")
print conv_decode25.shape
#conv_decode26 = conv_layer_with_bn(conv_decode25, [5, 5, 64, images.get_shape().as_list()[3]], phase_train, name="conv26")
#print conv_decode26.shape
#start classification
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights', shape=[1, 1, 64, num_classes], initializer=msra_initializer(1, 64),wd=0.0005)
print kernel.shape
conv = tf.nn.conv2d(conv_decode25, kernel, [1, 1, 1, 1], padding='SAME')
print "final" + str(conv.shape)
biases = _variable_on_cpu('biases', [num_classes], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
loss = cal_loss(conv_classifier, labels)
return loss, logit
def inference(FLAGS, images, labels, batch_size, phase_train):
# norm1
norm1 = tf.nn.lrn(images,
depth_radius=5,
bias=1.0,
alpha=0.0001,
beta=0.75,
name='norm1')
# conv1
conv1 = conv_layer_with_bn(
norm1, [7, 7, images.get_shape().as_list()[3], 64],
phase_train,
name="conv1")
# pool1
pool1, pool1_indices = tf.nn.max_pool_with_argmax(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# conv2
conv2 = conv_layer_with_bn(pool1, [7, 7, 64, 64],
phase_train,
name="conv2")
# pool2
pool2, pool2_indices = tf.nn.max_pool_with_argmax(conv2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# conv3
conv3 = conv_layer_with_bn(pool2, [7, 7, 64, 64],
phase_train,
name="conv3")
# pool3
pool3, pool3_indices = tf.nn.max_pool_with_argmax(conv3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# conv4
conv4 = conv_layer_with_bn(pool3, [7, 7, 64, 64],
phase_train,
name="conv4")
# pool4
pool4, pool4_indices = tf.nn.max_pool_with_argmax(conv4,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
""" End of encoder """
""" start upsample """
# upsample4
# Need to change when using different dataset out_w, out_h
# upsample4 = upsample_with_pool_indices(
# pool4, pool4_indices, pool4.get_shape(), out_w=45, out_h=60, scale=2, name='upsample4'
# )
upsample4 = deconv_layer(pool4, [2, 2, 64, 64], [batch_size, 45, 60, 64],
2, "up4")
# decode 4
conv_decode4 = conv_layer_with_bn(upsample4, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode4")
# upsample 3
# upsample3 = upsample_with_pool_indices(
# conv_decode4, pool3_indices, conv_decode4.get_shape(), scale=2, name='upsample3'
# )
upsample3 = deconv_layer(conv_decode4, [2, 2, 64, 64],
[batch_size, 90, 120, 64], 2, "up3")
# decode 3
conv_decode3 = conv_layer_with_bn(upsample3, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode3")
# upsample2
# upsample2 = upsample_with_pool_indices(
# conv_decode3, pool2_indices, conv_decode3.get_shape(), scale=2, name='upsample2'
# )
upsample2 = deconv_layer(conv_decode3, [2, 2, 64, 64],
[batch_size, 180, 240, 64], 2, "up2")
# decode 2
conv_decode2 = conv_layer_with_bn(upsample2, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode2")
# upsample1
# upsample1 = upsample_with_pool_indices(
# conv_decode2, pool1_indices, conv_decode2.get_shape(), scale=2, name='upsample1'
# )
upsample1 = deconv_layer(conv_decode2, [2, 2, 64, 64],
[batch_size, 360, 480, 64], 2, "up1")
# decode4
conv_decode1 = conv_layer_with_bn(upsample1, [7, 7, 64, 64],
phase_train,
False,
name="conv_decode1")
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay(
'weights',
shape=[1, 1, 64, FLAGS.num_classes],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [FLAGS.num_classes],
tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
loss = cal_loss(FLAGS, conv_classifier, labels)
return loss, logit
def prediction_myModel(images, batch_size=1, phase_train=False):
# norm1
norm1 = tf.nn.lrn(images, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75,
name='norm1')
# conv1
conv1_1 = conv_layer_with_bn(norm1, [3, 3, images.get_shape().as_list()[3], 64], phase_train, name="conv1_1")
conv1_2 = conv_layer_with_bn(conv1_1, [3, 3, 64, 64], phase_train, name="conv1_2")
# pool1
pool1 = tf.nn.max_pool(conv1_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],padding='SAME', name='pool1')
# conv2
conv2_1 = conv_layer_with_bn(pool1, [3, 3, 64, 128], phase_train, name="conv2_1")
conv2_2 = conv_layer_with_bn(conv2_1, [3, 3, 128, 128], phase_train, name="conv2_2")
# pool2
pool2 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME', name='pool2')
# conv3
conv3_1 = conv_layer_with_bn(pool2, [3, 3, 128, 256], phase_train, name="conv3_1")
conv3_2 = conv_layer_with_bn(conv3_1, [3, 3, 256, 256], phase_train, name="conv3_2")
conv3_3 = conv_layer_with_bn(conv3_2, [3, 3, 256, 256], phase_train, name="conv3_3")
# pool3
pool3 = tf.nn.max_pool(conv3_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool3')
# conv4
conv4_1 = conv_layer_with_bn(pool3, [3, 3, 256, 512], phase_train, name="conv4_1")
conv4_2 = conv_layer_with_bn(conv4_1, [3, 3, 512, 512], phase_train, name="conv4_2")
conv4_3 = conv_layer_with_bn(conv4_2, [3, 3, 512, 512], phase_train, name="conv4_3")
# pool4
pool4 = tf.nn.max_pool(conv4_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool4')
# conv5
conv5_1 = conv_layer_with_bn(pool4, [3, 3, 512, 512], phase_train, name="conv5_1")
conv5_2 = conv_layer_with_bn(conv5_1, [3, 3, 512, 512], phase_train, name="conv5_2")
conv5_3 = conv_layer_with_bn(conv5_2, [3, 3, 512, 512], phase_train, name="conv5_3")
# pool5
pool5 = tf.nn.max_pool(conv5_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool5')
""" End of encoder """
""" start upsample """
# upsample5
upsample5 = deconv_layer(pool5, [2, 2, 512, 512], [batch_size, 23, 30, 512], 2, "up5")
# decode 4
conv_decode5_3 = conv_layer_with_bn(upsample5, [3, 3, 512, 512], phase_train, False, name="conv_decode5_3")
conv_decode5_2 = conv_layer_with_bn(conv_decode5_3, [3, 3, 512, 512], phase_train, False, name="conv_decode5_2")
conv_decode5_1 = conv_layer_with_bn(conv_decode5_2, [3, 3, 512, 512], phase_train, False, name="conv_decode5_1")
# upsample4
# Need to change when using different dataset out_w, out_h
# upsample4 = upsample_with_pool_indices(pool4, pool4_indices, pool4.get_shape(), out_w=45, out_h=60, scale=2, name='upsample4')
upsample4 = deconv_layer(conv_decode5_1, [2, 2, 512, 512], [batch_size, 45, 60, 512], 2, "up4")
# decode 4
conv_decode4_3 = conv_layer_with_bn(upsample4, [3, 3, 512, 512], phase_train, False, name="conv_decode4_3")
conv_decode4_2 = conv_layer_with_bn(conv_decode4_3, [3, 3, 512, 512], phase_train, False, name="conv_decode4_2")
conv_decode4_1 = conv_layer_with_bn(conv_decode4_2, [3, 3, 512, 256], phase_train, False, name="conv_decode4_1")
# upsample 3
# upsample3 = upsample_with_pool_indices(conv_decode4, pool3_indices, conv_decode4.get_shape(), scale=2, name='upsample3')
upsample3= deconv_layer(conv_decode4_1, [2, 2, 256, 256], [batch_size, 90, 120, 256], 2, "up3")
# decode 3
conv_decode3_3 = conv_layer_with_bn(upsample3, [3, 3, 256, 256], phase_train, False, name="conv_decode3_3")
conv_decode3_2 = conv_layer_with_bn(conv_decode3_3, [3, 3, 256, 256], phase_train, False, name="conv_decode3_2")
conv_decode3_1 = conv_layer_with_bn(conv_decode3_2, [3, 3, 256, 128], phase_train, False, name="conv_decode3_1")
# upsample2
# upsample2 = upsample_with_pool_indices(conv_decode3, pool2_indices, conv_decode3.get_shape(), scale=2, name='upsample2')
upsample2= deconv_layer(conv_decode3_1, [2, 2, 128, 128], [batch_size, 180, 240, 128], 2, "up2")
# decode 2
conv_decode2_2 = conv_layer_with_bn(upsample2, [3, 3, 128, 128], phase_train, False, name="conv_decode2_2")
conv_decode2_1 = conv_layer_with_bn(conv_decode2_2, [3, 3, 128, 64], phase_train, False, name="conv_decode2_1")
# upsample1
# upsample1 = upsample_with_pool_indices(conv_decode2, pool1_indices, conv_decode2.get_shape(), scale=2, name='upsample1')
upsample1= deconv_layer(conv_decode2_1, [2, 2, 64, 64], [batch_size, 360, 480, 64], 2, "up1")
# decode4
conv_decode1_2 = conv_layer_with_bn(upsample1, [3, 3, 64, 64], phase_train, False, name="conv_decode1_2")
conv_decode1 = conv_layer_with_bn(conv_decode1_2, [3, 3, 64, 64], phase_train, False, name="conv_decode1")
""" end of Decode """
""" Start Classify """
# output predicted class number (6)
with tf.variable_scope('conv_classifier') as scope:
kernel = _variable_with_weight_decay('weights',
shape=[1, 1, 64, NUM_CLASSES],
initializer=msra_initializer(1, 64),
wd=0.0005)
conv = tf.nn.conv2d(conv_decode1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [NUM_CLASSES], tf.constant_initializer(0.0))
conv_classifier = tf.nn.bias_add(conv, biases, name=scope.name)
logit = conv_classifier
return logit
