def unet2D_bn_modified(images, training, nlabels):
images_padded = tf.pad(images, [[0,0], [92, 92], [92, 92], [0,0]], 'CONSTANT')
conv1_1 = layers.conv2D_layer_bn(images_padded, 'conv1_1', num_filters=64, training=training, padding='VALID')
conv1_2 = layers.conv2D_layer_bn(conv1_1, 'conv1_2', num_filters=64, training=training, padding='VALID')
pool1 = layers.max_pool_layer2d(conv1_2)
conv2_1 = layers.conv2D_layer_bn(pool1, 'conv2_1', num_filters=128, training=training, padding='VALID')
conv2_2 = layers.conv2D_layer_bn(conv2_1, 'conv2_2', num_filters=128, training=training, padding='VALID')
pool2 = layers.max_pool_layer2d(conv2_2)
conv3_1 = layers.conv2D_layer_bn(pool2, 'conv3_1', num_filters=256, training=training, padding='VALID')
conv3_2 = layers.conv2D_layer_bn(conv3_1, 'conv3_2', num_filters=256, training=training, padding='VALID')
pool3 = layers.max_pool_layer2d(conv3_2)
conv4_1 = layers.conv2D_layer_bn(pool3, 'conv4_1', num_filters=512, training=training, padding='VALID')
conv4_2 = layers.conv2D_layer_bn(conv4_1, 'conv4_2', num_filters=512, training=training, padding='VALID')
pool4 = layers.max_pool_layer2d(conv4_2)
conv5_1 = layers.conv2D_layer_bn(pool4, 'conv5_1', num_filters=1024, training=training, padding='VALID')
conv5_2 = layers.conv2D_layer_bn(conv5_1, 'conv5_2', num_filters=1024, training=training, padding='VALID')
upconv4 = layers.deconv2D_layer_bn(conv5_2, name='upconv4', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat4 = layers.crop_and_concat_layer([upconv4, conv4_2], axis=3)
conv6_1 = layers.conv2D_layer_bn(concat4, 'conv6_1', num_filters=512, training=training, padding='VALID')
conv6_2 = layers.conv2D_layer_bn(conv6_1, 'conv6_2', num_filters=512, training=training, padding='VALID')
upconv3 = layers.deconv2D_layer_bn(conv6_2, name='upconv3', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat3 = layers.crop_and_concat_layer([upconv3, conv3_2], axis=3)
conv7_1 = layers.conv2D_layer_bn(concat3, 'conv7_1', num_filters=256, training=training, padding='VALID')
conv7_2 = layers.conv2D_layer_bn(conv7_1, 'conv7_2', num_filters=256, training=training, padding='VALID')
upconv2 = layers.deconv2D_layer_bn(conv7_2, name='upconv2', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat2 = layers.crop_and_concat_layer([upconv2, conv2_2], axis=3)
conv8_1 = layers.conv2D_layer_bn(concat2, 'conv8_1', num_filters=128, training=training, padding='VALID')
conv8_2 = layers.conv2D_layer_bn(conv8_1, 'conv8_2', num_filters=128, training=training, padding='VALID')
upconv1 = layers.deconv2D_layer_bn(conv8_2, name='upconv1', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat1 = layers.crop_and_concat_layer([upconv1, conv1_2], axis=3)
conv9_1 = layers.conv2D_layer_bn(concat1, 'conv9_1', num_filters=64, training=training, padding='VALID')
conv9_2 = layers.conv2D_layer_bn(conv9_1, 'conv9_2', num_filters=64, training=training, padding='VALID')
pred = layers.conv2D_layer_bn(conv9_2, 'pred', num_filters=nlabels, kernel_size=(1,1), activation=tf.identity, training=training, padding='VALID')
return pred
def unet2D_bn_padding_same_modified(images, training, nlabels):
conv1_1 = layers.conv2D_layer_bn(images, 'conv1_1', num_filters=64, training=training)
conv1_2 = layers.conv2D_layer_bn(conv1_1, 'conv1_2', num_filters=64, training=training)
pool1 = layers.max_pool_layer2d(conv1_2)
conv2_1 = layers.conv2D_layer_bn(pool1, 'conv2_1', num_filters=128, training=training)
conv2_2 = layers.conv2D_layer_bn(conv2_1, 'conv2_2', num_filters=128, training=training)
pool2 = layers.max_pool_layer2d(conv2_2)
conv3_1 = layers.conv2D_layer_bn(pool2, 'conv3_1', num_filters=256, training=training)
conv3_2 = layers.conv2D_layer_bn(conv3_1, 'conv3_2', num_filters=256, training=training)
pool3 = layers.max_pool_layer2d(conv3_2)
conv4_1 = layers.conv2D_layer_bn(pool3, 'conv4_1', num_filters=512, training=training)
conv4_2 = layers.conv2D_layer_bn(conv4_1, 'conv4_2', num_filters=512, training=training)
pool4 = layers.max_pool_layer2d(conv4_2)
conv5_1 = layers.conv2D_layer_bn(pool4, 'conv5_1', num_filters=1024, training=training)
conv5_2 = layers.conv2D_layer_bn(conv5_1, 'conv5_2', num_filters=1024, training=training)
upconv4 = layers.deconv2D_layer_bn(conv5_2, name='upconv4', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat4 = tf.concat([conv4_2, upconv4], axis=3, name='concat4')
conv6_1 = layers.conv2D_layer_bn(concat4, 'conv6_1', num_filters=512, training=training)
conv6_2 = layers.conv2D_layer_bn(conv6_1, 'conv6_2', num_filters=512, training=training)
upconv3 = layers.deconv2D_layer_bn(conv6_2, name='upconv3', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat3 = tf.concat([conv3_2, upconv3], axis=3, name='concat3')
conv7_1 = layers.conv2D_layer_bn(concat3, 'conv7_1', num_filters=256, training=training)
conv7_2 = layers.conv2D_layer_bn(conv7_1, 'conv7_2', num_filters=256, training=training)
upconv2 = layers.deconv2D_layer_bn(conv7_2, name='upconv2', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat2 = tf.concat([conv2_2, upconv2], axis=3, name='concat2')
conv8_1 = layers.conv2D_layer_bn(concat2, 'conv8_1', num_filters=128, training=training)
conv8_2 = layers.conv2D_layer_bn(conv8_1, 'conv8_2', num_filters=128, training=training)
upconv1 = layers.deconv2D_layer_bn(conv8_2, name='upconv1', kernel_size=(4, 4), strides=(2, 2), num_filters=nlabels, training=training)
concat1 = tf.concat([conv1_2, upconv1], axis=3, name='concat1')
conv9_1 = layers.conv2D_layer_bn(concat1, 'conv9_1', num_filters=64, training=training)
conv9_2 = layers.conv2D_layer_bn(conv9_1, 'conv9_2', num_filters=64, training=training)
pred = layers.conv2D_layer_bn(conv9_2, 'pred', num_filters=nlabels, kernel_size=(1,1), activation=tf.identity, training=training)
return pred
def VGG16_FCN_8(images, training, nlabels):
conv1_1 = layers.conv2D_layer(images, 'conv1_1', num_filters=64)
conv1_2 = layers.conv2D_layer(conv1_1, 'conv1_2', num_filters=64)
pool1 = layers.max_pool_layer2d(conv1_2)
conv2_1 = layers.conv2D_layer(pool1, 'conv2_1', num_filters=128)
conv2_2 = layers.conv2D_layer(conv2_1, 'conv2_2', num_filters=128)
pool2 = layers.max_pool_layer2d(conv2_2)
conv3_1 = layers.conv2D_layer(pool2, 'conv3_1', num_filters=256)
conv3_2 = layers.conv2D_layer(conv3_1, 'conv3_2', num_filters=256)
conv3_3 = layers.conv2D_layer(conv3_2, 'conv3_3', num_filters=256)
pool3 = layers.max_pool_layer2d(conv3_3)
conv4_1 = layers.conv2D_layer(pool3, 'conv4_1', num_filters=512)
conv4_2 = layers.conv2D_layer(conv4_1, 'conv4_2', num_filters=512)
conv4_3 = layers.conv2D_layer(conv4_2, 'conv4_3', num_filters=512)
pool4 = layers.max_pool_layer2d(conv4_3)
conv5_1 = layers.conv2D_layer(pool4, 'conv5_1', num_filters=512)
conv5_2 = layers.conv2D_layer(conv5_1, 'conv5_2', num_filters=512)
conv5_3 = layers.conv2D_layer(conv5_2, 'conv5_3', num_filters=512)
pool5 = layers.max_pool_layer2d(conv5_3)
conv6 = layers.conv2D_layer(pool5, 'conv6', num_filters=4096, kernel_size=(3,3))
conv7= layers.conv2D_layer(conv6, 'conv7', num_filters=4096, kernel_size=(1,1))
score5 = layers.conv2D_layer(conv7, 'score5', num_filters=nlabels, kernel_size=(1,1))
score4 = layers.conv2D_layer(pool4, 'score4', num_filters=nlabels, kernel_size=(1,1))
score3 = layers.conv2D_layer(pool3, 'score3', num_filters=nlabels, kernel_size=(1,1))
upscore1 = layers.deconv2D_layer(score5, name='upscore1', kernel_size=(4,4), strides=(2,2), num_filters=nlabels, weight_init='bilinear')
sum1 = tf.add(upscore1, score4)
upscore2 = layers.deconv2D_layer(sum1, name='upscore2', kernel_size=(4,4), strides=(2,2), num_filters=nlabels, weight_init='bilinear')
sum2 = tf.add(upscore2, score3)
upscore3 = layers.deconv2D_layer(sum2, name='upscore3', kernel_size=(16,16), strides=(8,8), num_filters=nlabels, weight_init='bilinear', activation=tf.identity)
return upscore3
def unet2D_padding_same_shallow(images, training, nlabels):
conv1_1 = layers.conv2D_layer(images, 'conv1_1', num_filters=64)
conv1_2 = layers.conv2D_layer(conv1_1, 'conv1_2', num_filters=64)
pool1 = layers.max_pool_layer2d(conv1_2)
conv2_1 = layers.conv2D_layer(pool1, 'conv2_1', num_filters=128)
conv2_2 = layers.conv2D_layer(conv2_1, 'conv2_2', num_filters=128)
pool2 = layers.max_pool_layer2d(conv2_2)
conv3_1 = layers.conv2D_layer(pool2, 'conv3_1', num_filters=256)
conv3_2 = layers.conv2D_layer(conv3_1, 'conv3_2', num_filters=256)
pool3 = layers.max_pool_layer2d(conv3_2)
conv4_1 = layers.conv2D_layer(pool3, 'conv4_1', num_filters=512)
conv4_2 = layers.conv2D_layer(conv4_1, 'conv4_2', num_filters=512)
upconv3 = layers.deconv2D_layer(conv4_2, name='upconv3', kernel_size=(4, 4), strides=(2, 2), num_filters=256)
concat3 = tf.concat([conv3_2, upconv3], axis=3, name='concat3')
conv7_1 = layers.conv2D_layer(concat3, 'conv7_1', num_filters=256)
conv7_2 = layers.conv2D_layer(conv7_1, 'conv7_2', num_filters=256)
upconv2 = layers.deconv2D_layer(conv7_2, name='upconv2', kernel_size=(4, 4), strides=(2, 2), num_filters=128)
concat2 = tf.concat([conv2_2, upconv2], axis=3, name='concat2')
conv8_1 = layers.conv2D_layer(concat2, 'conv8_1', num_filters=128)
conv8_2 = layers.conv2D_layer(conv8_1, 'conv8_2', num_filters=128)
upconv1 = layers.deconv2D_layer(conv8_2, name='upconv1', kernel_size=(4, 4), strides=(2, 2), num_filters=64)
concat1 = tf.concat([conv1_2, upconv1], axis=3, name='concat1')
conv9_1 = layers.conv2D_layer(concat1, 'conv9_1', num_filters=64)
conv9_2 = layers.conv2D_layer(conv9_1, 'conv9_2', num_filters=64)
pred = layers.conv2D_layer(conv9_2, 'pred', num_filters=nlabels, kernel_size=(1,1), activation=tf.identity)
return pred
def discriminator(features,
training_pl,
scope_name = 'discriminator',
scope_reuse = False):
with tf.variable_scope(scope_name,
reuse = scope_reuse):
out = features
num_layers = 5
n0 = 16
for l in range(num_layers):
out = tf.layers.conv2d(inputs=out,
filters=(l+1)*n0,
kernel_size=3,
padding='SAME',
name='D_conv_'+str(l+1) + '_1',
use_bias=True,
activation=None)
out = tf.layers.batch_normalization(inputs=out, name = 'D_conv_' + str(l+1) + '_1' + '_bn', training = training_pl)
out = tf.nn.relu(out)
out = tf.layers.conv2d(inputs=out,
filters=(l+1)*n0,
kernel_size=3,
padding='SAME',
name='D_conv_'+str(l+1) + '_2',
use_bias=True,
activation=None)
out = tf.layers.batch_normalization(inputs=out, name = 'D_conv_' + str(l+1) + '_2' + '_bn', training = training_pl)
out = tf.nn.relu(out)
out = layers.max_pool_layer2d(out)
batch_size = out.get_shape()[0].value
out = tf.reshape(out, [batch_size, -1])
logits = tf.layers.dense(out, 1, name = 'D_logits')
# outputs = tf.nn.sigmoid(logits)
return logits
def forward(images, training, nlabels):
conv1_1 = layers.conv2D_layer_bn(images,
'conv1_1',
num_filters=64,
training=training,
padding='SAME')
conv1_2 = layers.conv2D_layer_bn(conv1_1,
'conv1_2',
num_filters=64,
training=training,
padding='SAME')
pool1 = layers.max_pool_layer2d(conv1_2, 'pool_1')
conv2_1 = layers.conv2D_layer_bn(pool1,
'conv2_1',
num_filters=128,
training=training,
padding='SAME')
conv2_2 = layers.conv2D_layer_bn(conv2_1,
'conv2_2',
num_filters=128,
training=training,
padding='SAME')
pool2 = layers.max_pool_layer2d(conv2_2, 'pool_2')
dout2 = layers.dropout_layer(pool2, 'dropout_2', training)
conv3_1 = layers.conv2D_layer_bn(dout2,
'conv3_1',
num_filters=256,
training=training,
padding='SAME')
conv3_2 = layers.conv2D_layer_bn(conv3_1,
'conv3_2',
num_filters=256,
training=training,
padding='SAME')
pool3 = layers.max_pool_layer2d(conv3_2, 'pool_3')
dout3 = layers.dropout_layer(pool3, 'dropout_3', training)
conv4_1 = layers.conv2D_layer_bn(dout3,
'conv4_1',
num_filters=512,
training=training,
padding='SAME')
conv4_2 = layers.conv2D_layer_bn(conv4_1,
'conv4_2',
num_filters=512,
training=training,
padding='SAME')
pool4 = layers.max_pool_layer2d(conv4_2, 'pool_4')
dout4 = layers.dropout_layer(pool4, 'dropout_4', training)
conv5_1 = layers.conv2D_layer_bn(dout4,
'conv5_1',
num_filters=1024,
training=training,
padding='SAME')
conv5_2 = layers.conv2D_layer_bn(conv5_1,
'conv5_2',
num_filters=1024,
training=training,
padding='SAME')
upconv4 = layers.deconv2D_layer_bn(conv5_2,
name='upconv4',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
training=training)
concat4 = layers.crop_and_concat_layer([upconv4, conv4_2],
'crop_concat_4',
axis=3)
dout5 = layers.dropout_layer(concat4, 'dropout_5', training)
conv6_1 = layers.conv2D_layer_bn(dout5,
'conv6_1',
num_filters=512,
training=training,
padding='SAME')
conv6_2 = layers.conv2D_layer_bn(conv6_1,
'conv6_2',
num_filters=512,
training=training,
padding='SAME')
upconv3 = layers.deconv2D_layer_bn(conv6_2,
name='upconv3',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
training=training)
concat3 = layers.crop_and_concat_layer([upconv3, conv3_2],
'crop_concat_3',
axis=3)
dout6 = layers.dropout_layer(concat3, 'dropout_6', training)
conv7_1 = layers.conv2D_layer_bn(dout6,
'conv7_1',
num_filters=256,
training=training,
padding='SAME')
conv7_2 = layers.conv2D_layer_bn(conv7_1,
'conv7_2',
num_filters=256,
training=training,
padding='SAME')
upconv2 = layers.deconv2D_layer_bn(conv7_2,
name='upconv2',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
training=training)
concat2 = layers.crop_and_concat_layer([upconv2, conv2_2],
'crop_concat_2',
axis=3)
dout7 = layers.dropout_layer(concat2, 'dropout_7', training)
conv8_1 = layers.conv2D_layer_bn(dout7,
'conv8_1',
num_filters=128,
training=training,
padding='SAME')
conv8_2 = layers.conv2D_layer_bn(conv8_1,
'conv8_2',
num_filters=128,
training=training,
padding='SAME')
upconv1 = layers.deconv2D_layer_bn(conv8_2,
name='upconv1',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
training=training)
concat1 = layers.crop_and_concat_layer([upconv1, conv1_2],
'crop_concat_1',
axis=3)
conv9_1 = layers.conv2D_layer_bn(concat1,
'conv9_1',
num_filters=64,
training=training,
padding='SAME')
conv9_2 = layers.conv2D_layer_bn(conv9_1,
'conv9_2',
num_filters=64,
training=training,
padding='SAME')
pred_1 = layers.conv2D_layer_bn(conv9_2,
'pred',
num_filters=nlabels,
kernel_size=(1, 1),
activation=tf.identity,
training=training,
padding='SAME')
# Deep supervision
ds1_1 = layers.conv2D_layer(conv7_2,
'ds_1',
num_filters=nlabels,
kernel_size=(1, 1),
activation=tf.identity,
padding='SAME')
ds1_2 = layers.deconv2D_layer(ds1_1,
'ds_2',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
padding='SAME')
ds2_1 = layers.conv2D_layer(conv8_2,
'ds_3',
num_filters=nlabels,
kernel_size=(1, 1),
activation=tf.identity,
padding='SAME')
ds1_ds2 = tf.add(ds1_2, ds2_1)
ds = layers.deconv2D_layer(ds1_ds2,
'ds_4',
kernel_size=(4, 4),
strides=(2, 2),
num_filters=nlabels,
padding='SAME')
pred_2 = tf.add(pred_1, ds)
return pred_2
def unet2D_i2l(images,
nlabels,
training_pl,
scope_reuse = False):
n0 = 16
n1, n2, n3, n4 = 1*n0, 2*n0, 4*n0, 8*n0
with tf.variable_scope('i2l_mapper') as scope:
if scope_reuse:
scope.reuse_variables()
# ====================================
# 1st Conv block - two conv layers, followed by max-pooling
# ====================================
conv1_1 = layers.conv2D_layer_bn(x=images, name='conv1_1', num_filters=n1, training = training_pl)
conv1_2 = layers.conv2D_layer_bn(x=conv1_1, name='conv1_2', num_filters=n1, training = training_pl)
pool1 = layers.max_pool_layer2d(conv1_2)
# ====================================
# 2nd Conv block
# ====================================
conv2_1 = layers.conv2D_layer_bn(x=pool1, name='conv2_1', num_filters=n2, training = training_pl)
conv2_2 = layers.conv2D_layer_bn(x=conv2_1, name='conv2_2', num_filters=n2, training = training_pl)
pool2 = layers.max_pool_layer2d(conv2_2)
# ====================================
# 3rd Conv block
# ====================================
conv3_1 = layers.conv2D_layer_bn(x=pool2, name='conv3_1', num_filters=n3, training = training_pl)
conv3_2 = layers.conv2D_layer_bn(x=conv3_1, name='conv3_2', num_filters=n3, training = training_pl)
pool3 = layers.max_pool_layer2d(conv3_1)
# ====================================
# 4th Conv block
# ====================================
conv4_1 = layers.conv2D_layer_bn(x=pool3, name='conv4_1', num_filters=n4, training = training_pl)
conv4_2 = layers.conv2D_layer_bn(x=conv4_1, name='conv4_2', num_filters=n4, training = training_pl)
# ====================================
# Upsampling via bilinear upsampling, concatenation (skip connection), followed by 2 conv layers
# ====================================
deconv3 = layers.bilinear_upsample2D(conv4_2, size = (tf.shape(conv3_2)[1],tf.shape(conv3_2)[2]), name='upconv3')
concat3 = tf.concat([deconv3, conv3_2], axis=-1)
conv5_1 = layers.conv2D_layer_bn(x=concat3, name='conv5_1', num_filters=n3, training = training_pl)
conv5_2 = layers.conv2D_layer_bn(x=conv5_1, name='conv5_2', num_filters=n3, training = training_pl)
# ====================================
# Upsampling via bilinear upsampling, concatenation (skip connection), followed by 2 conv layers
# ====================================
deconv2 = layers.bilinear_upsample2D(conv5_2, size = (tf.shape(conv2_2)[1],tf.shape(conv2_2)[2]), name='upconv2')
concat2 = tf.concat([deconv2, conv2_2], axis=-1)
conv6_1 = layers.conv2D_layer_bn(x=concat2, name='conv6_1', num_filters=n2, training = training_pl)
conv6_2 = layers.conv2D_layer_bn(x=conv6_1, name='conv6_2', num_filters=n2, training = training_pl)
# ====================================
# Upsampling via bilinear upsampling, concatenation (skip connection), followed by 2 conv layers
# ====================================
deconv1 = layers.bilinear_upsample2D(conv6_2, size = (tf.shape(conv1_2)[1],tf.shape(conv1_2)[2]), name='upconv1')
concat1 = tf.concat([deconv1, conv1_2], axis=-1)
conv7_1 = layers.conv2D_layer_bn(x=concat1, name='conv7_1', num_filters=n1, training = training_pl)
conv7_2 = layers.conv2D_layer_bn(x=conv7_1, name='conv7_2', num_filters=n1, training = training_pl)
# ====================================
# Final conv layer - without batch normalization or activation
# ====================================
pred = layers.conv2D_layer(x=conv7_2, name='pred', num_filters=nlabels, kernel_size=1)
return pool1, pool2, pool3, conv4_2, conv5_2, conv6_2, conv7_2, pred