def __init__(self, input_dim, output_dim=1, optimizer=None, use_dropout=True, lr=0.001, random_seed=123,
is_train=True, log_dir=None, name=None):
self.name = name
self.is_train = is_train
self.log_dir = log_dir
self.cur_lr = None
self.logger, self.file_handler, self.stream_handler = utils.init_logger(log_dir=self.log_dir,
name=self.name,
is_train=self.is_train)
with tf.variable_scope(self.name):
# Placeholders for inputs
self.X = tf.placeholder(dtype=tf.float32, shape=[None, input_dim], name='X')
self.y = tf.placeholder(dtype=tf.float32, shape=[None, output_dim], name='y')
self.keep_prob = tf.placeholder(tf.float32, name='keep_prob')
tf_utils.print_activations(self.X, logger=self.logger if self.is_train else None)
# Placeholders for TensorBoard
self.train_acc = tf.placeholder(tf.float32, name='train_acc')
self.val_acc = tf.placeholder(tf.float32, name='val_acc')
net = self.X
if use_dropout:
net = tf_utils.dropout(x=net,
keep_prob=self.keep_prob,
seed=random_seed,
name='dropout',
logger=self.logger if self.is_train else None)
# Network, loss, and optimizer
self.y_pred = tf_utils.linear(net, output_size=output_dim)
tf_utils.print_activations(self.y_pred, logger=self.logger if self.is_train else None)
self.loss = tf.math.reduce_mean(tf.nn.l2_loss(self.y_pred - self.y))
self.train_op, self.cur_lr = optimizer_fn(optimizer, lr=lr, loss=self.loss, name=self.name)
# Accuracy etc
self.y_pred_round = tf.math.round(x=self.y_pred, name='rounded_pred')
accuracy = tf.equal(tf.cast(x=self.y_pred_round, dtype=tf.int32), tf.cast(x=self.y, dtype=tf.int32))
self.accuracy = tf.reduce_mean(tf.cast(x=accuracy, dtype=tf.float32)) * 100.
self._tensorboard()
tf_utils.show_all_variables(logger=self.logger if self.is_train else None)
def forward_network(self, img, padding='SAME', reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
# Stage 0
s0_conv1 = tf_utils.conv2d(x=img,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s0_conv1')
s0_conv1 = tf_utils.relu(s0_conv1, name='relu_s0_conv1')
s0_conv2 = tf_utils.conv2d(x=s0_conv1,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s0_conv2')
s0_conv2 = tf_utils.norm(s0_conv2,
name='s0_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s0_conv2 = tf_utils.relu(s0_conv2, name='relu_s0_conv2')
# Stage 1
s1_maxpool = tf_utils.max_pool(x=s0_conv2, name='s1_maxpool2d')
s1_conv1 = tf_utils.conv2d(x=s1_maxpool,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s1_conv1')
s1_conv1 = tf_utils.norm(s1_conv1,
name='s1_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1')
s1_conv2 = tf_utils.conv2d(x=s1_conv1,
output_dim=self.conv_dims[1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s1_conv2')
s1_conv2 = tf_utils.norm(s1_conv2,
name='s1_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2')
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2, name='s2_maxpool2d')
s2_conv1 = tf_utils.conv2d(x=s2_maxpool,
output_dim=self.conv_dims[2],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s2_conv1')
s2_conv1 = tf_utils.norm(s2_conv1,
name='s2_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s2_conv1 = tf_utils.relu(s2_conv1, name='relu_s2_conv1')
s2_conv2 = tf_utils.conv2d(x=s2_conv1,
output_dim=self.conv_dims[3],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s2_conv2')
s2_conv2 = tf_utils.norm(s2_conv2,
name='s2_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s2_conv2 = tf_utils.relu(s2_conv2, name='relu_s2_conv2')
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2, name='s3_maxpool2d')
s3_conv1 = tf_utils.conv2d(x=s3_maxpool,
output_dim=self.conv_dims[4],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s3_conv1')
s3_conv1 = tf_utils.norm(s3_conv1,
name='s3_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s3_conv1 = tf_utils.relu(s3_conv1, name='relu_s3_conv1')
s3_conv2 = tf_utils.conv2d(x=s3_conv1,
output_dim=self.conv_dims[5],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s3_conv2')
s3_conv2 = tf_utils.norm(s3_conv2,
name='s3_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s3_conv2 = tf_utils.relu(s3_conv2, name='relu_s3_conv2')
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2, name='s4_maxpool2d')
s4_conv1 = tf_utils.conv2d(x=s4_maxpool,
output_dim=self.conv_dims[6],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s4_conv1')
s4_conv1 = tf_utils.norm(s4_conv1,
name='s4_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s4_conv1 = tf_utils.relu(s4_conv1, name='relu_s4_conv1')
s4_conv2 = tf_utils.conv2d(x=s4_conv1,
output_dim=self.conv_dims[7],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s4_conv2')
s4_conv2 = tf_utils.norm(s4_conv2,
name='s4_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s4_conv2 = tf_utils.relu(s4_conv2, name='relu_s4_conv2')
s4_conv2_drop = tf_utils.dropout(x=s4_conv2,
keep_prob=0.,
name='s4_dropout')
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop,
name='s5_maxpool2d')
s5_conv1 = tf_utils.conv2d(x=s5_maxpool,
output_dim=self.conv_dims[8],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s5_conv1')
s5_conv1 = tf_utils.norm(s5_conv1,
name='s5_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s5_conv1 = tf_utils.relu(s5_conv1, name='relu_s5_conv1')
s5_conv2 = tf_utils.conv2d(x=s5_conv1,
output_dim=self.conv_dims[9],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s5_conv2')
s5_conv2 = tf_utils.norm(s5_conv2,
name='s5_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s5_conv2 = tf_utils.relu(s5_conv2, name='relu_s5_conv2')
s5_conv2_drop = tf_utils.dropout(x=s5_conv2,
keep_prob=0.,
name='s5_dropout')
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop,
output_dim=self.conv_dims[10],
k_h=2,
k_w=2,
initializer='He',
name='s6_deconv1')
s6_deconv1 = tf_utils.norm(s6_deconv1,
name='s6_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s6_deconv1 = tf_utils.relu(s6_deconv1, name='relu_s6_deconv1')
# Cropping
w1 = s4_conv2_drop.get_shape().as_list()[2]
w2 = s6_deconv1.get_shape().as_list()[2] - s4_conv2_drop.get_shape(
).as_list()[2]
s6_deconv1_split, _ = tf.split(s6_deconv1,
num_or_size_splits=[w1, w2],
axis=2,
name='axis2_split')
tf_utils.print_activations(s6_deconv1_split)
# Concat
s6_concat = tf_utils.concat(
values=[s6_deconv1_split, s4_conv2_drop],
axis=3,
name='s6_axis3_concat')
s6_conv2 = tf_utils.conv2d(x=s6_concat,
output_dim=self.conv_dims[11],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s6_conv2')
s6_conv2 = tf_utils.norm(s6_conv2,
name='s6_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s6_conv2 = tf_utils.relu(s6_conv2, name='relu_s6_conv2')
s6_conv3 = tf_utils.conv2d(x=s6_conv2,
output_dim=self.conv_dims[12],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s6_conv3')
s6_conv3 = tf_utils.norm(s6_conv3,
name='s6_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s6_conv3 = tf_utils.relu(s6_conv3, name='relu_s6_conv3')
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3,
output_dim=self.conv_dims[13],
k_h=2,
k_w=2,
initializer='He',
name='s7_deconv1')
s7_deconv1 = tf_utils.norm(s7_deconv1,
name='s7_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s7_deconv1 = tf_utils.relu(s7_deconv1, name='relu_s7_deconv1')
# Concat
s7_concat = tf_utils.concat(values=[s7_deconv1, s3_conv2],
axis=3,
name='s7_axis3_concat')
s7_conv2 = tf_utils.conv2d(x=s7_concat,
output_dim=self.conv_dims[14],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s7_conv2')
s7_conv2 = tf_utils.norm(s7_conv2,
name='s7_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s7_conv2 = tf_utils.relu(s7_conv2, name='relu_s7_conv2')
s7_conv3 = tf_utils.conv2d(x=s7_conv2,
output_dim=self.conv_dims[15],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s7_conv3')
s7_conv3 = tf_utils.norm(s7_conv3,
name='s7_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s7_conv3 = tf_utils.relu(s7_conv3, name='relu_s7_conv3')
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3,
output_dim=self.conv_dims[16],
k_h=2,
k_w=2,
initializer='He',
name='s8_deconv1')
s8_deconv1 = tf_utils.norm(s8_deconv1,
name='s8_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s8_deconv1 = tf_utils.relu(s8_deconv1, name='relu_s8_deconv1')
# Concat
s8_concat = tf_utils.concat(values=[s8_deconv1, s2_conv2],
axis=3,
name='s8_axis3_concat')
s8_conv2 = tf_utils.conv2d(x=s8_concat,
output_dim=self.conv_dims[17],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s8_conv2')
s8_conv2 = tf_utils.norm(s8_conv2,
name='s8_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s8_conv2 = tf_utils.relu(s8_conv2, name='relu_s8_conv2')
s8_conv3 = tf_utils.conv2d(x=s8_conv2,
output_dim=self.conv_dims[18],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s8_conv3')
s8_conv3 = tf_utils.norm(s8_conv3,
name='s8_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s8_conv3 = tf_utils.relu(s8_conv3, name='relu_conv3')
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3,
output_dim=self.conv_dims[19],
k_h=2,
k_w=2,
initializer='He',
name='s9_deconv1')
s9_deconv1 = tf_utils.norm(s9_deconv1,
name='s9_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s9_deconv1 = tf_utils.relu(s9_deconv1, name='relu_s9_deconv1')
# Concat
s9_concat = tf_utils.concat(values=[s9_deconv1, s1_conv2],
axis=3,
name='s9_axis3_concat')
s9_conv2 = tf_utils.conv2d(x=s9_concat,
output_dim=self.conv_dims[20],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s9_conv2')
s9_conv2 = tf_utils.norm(s9_conv2,
name='s9_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s9_conv2 = tf_utils.relu(s9_conv2, name='relu_s9_conv2')
s9_conv3 = tf_utils.conv2d(x=s9_conv2,
output_dim=self.conv_dims[21],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s9_conv3')
s9_conv3 = tf_utils.norm(s9_conv3,
name='s9_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s9_conv3 = tf_utils.relu(s9_conv3, name='relu_s9_conv3')
s10_deconv1 = tf_utils.deconv2d(x=s9_conv3,
output_dim=self.conv_dims[-1],
k_h=2,
k_w=2,
initializer='He',
name='s10_deconv1')
s10_deconv1 = tf_utils.norm(s10_deconv1,
name='s10_norm0',
_type='batch',
_ops=self._ops,
is_train=False)
s10_deconv1 = tf_utils.relu(s10_deconv1, name='relu_s10_deconv1')
# Concat
s10_concat = tf_utils.concat(values=[s10_deconv1, s0_conv2],
axis=3,
name='s10_axis3_concat')
s10_conv2 = tf_utils.conv2d(s10_concat,
output_dim=self.conv_dims[-1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s10_conv2')
s10_conv2 = tf_utils.norm(s10_conv2,
name='s10_norm1',
_type='batch',
_ops=self._ops,
is_train=False)
s10_conv2 = tf_utils.relu(s10_conv2, name='relu_s10_conv2')
s10_conv3 = tf_utils.conv2d(x=s10_conv2,
output_dim=self.conv_dims[-1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='s10_conv3')
s10_conv3 = tf_utils.norm(s10_conv3,
name='s10_norm2',
_type='batch',
_ops=self._ops,
is_train=False)
s10_conv3 = tf_utils.relu(s10_conv3, name='relu_s10_conv3')
output = tf_utils.conv2d(s10_conv3,
output_dim=self.num_classes,
k_h=1,
k_w=1,
d_h=1,
d_w=1,
padding=padding,
initializer='He',
name='output')
return output
def __call__(self, x, keep_rate=0.5):
with tf.compat.v1.variable_scope(self.name, reuse=self.reuse):
tf_utils.print_activations(x, logger=self.logger)
# E0: (320, 200) -> (160, 100)
e0_conv2d = tf_utils.conv2d(x,
output_dim=self.gen_c[0],
initializer='He',
logger=self.logger,
name='e0_conv2d')
e0_lrelu = tf_utils.lrelu(e0_conv2d,
logger=self.logger,
name='e0_lrelu')
# E1: (160, 100) -> (80, 50)
e1_conv2d = tf_utils.conv2d(e0_lrelu,
output_dim=self.gen_c[1],
initializer='He',
logger=self.logger,
name='e1_conv2d')
e1_batchnorm = tf_utils.norm(e1_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e1_norm')
e1_lrelu = tf_utils.lrelu(e1_batchnorm,
logger=self.logger,
name='e1_lrelu')
# E2: (80, 50) -> (40, 25)
e2_conv2d = tf_utils.conv2d(e1_lrelu,
output_dim=self.gen_c[2],
initializer='He',
logger=self.logger,
name='e2_conv2d')
e2_batchnorm = tf_utils.norm(e2_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e2_norm')
e2_lrelu = tf_utils.lrelu(e2_batchnorm,
logger=self.logger,
name='e2_lrelu')
# E3: (40, 25) -> (20, 13)
e3_conv2d = tf_utils.conv2d(e2_lrelu,
output_dim=self.gen_c[3],
initializer='He',
logger=self.logger,
name='e3_conv2d')
e3_batchnorm = tf_utils.norm(e3_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e3_norm')
e3_lrelu = tf_utils.lrelu(e3_batchnorm,
logger=self.logger,
name='e3_lrelu')
# E4: (20, 13) -> (10, 7)
e4_conv2d = tf_utils.conv2d(e3_lrelu,
output_dim=self.gen_c[4],
initializer='He',
logger=self.logger,
name='e4_conv2d')
e4_batchnorm = tf_utils.norm(e4_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e4_norm')
e4_lrelu = tf_utils.lrelu(e4_batchnorm,
logger=self.logger,
name='e4_lrelu')
# E5: (10, 7) -> (5, 4)
e5_conv2d = tf_utils.conv2d(e4_lrelu,
output_dim=self.gen_c[5],
initializer='He',
logger=self.logger,
name='e5_conv2d')
e5_batchnorm = tf_utils.norm(e5_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e5_norm')
e5_lrelu = tf_utils.lrelu(e5_batchnorm,
logger=self.logger,
name='e5_lrelu')
# E6: (5, 4) -> (3, 2)
e6_conv2d = tf_utils.conv2d(e5_lrelu,
output_dim=self.gen_c[6],
initializer='He',
logger=self.logger,
name='e6_conv2d')
e6_batchnorm = tf_utils.norm(e6_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e6_norm')
e6_lrelu = tf_utils.lrelu(e6_batchnorm,
logger=self.logger,
name='e6_lrelu')
# E7: (3, 2) -> (2, 1)
e7_conv2d = tf_utils.conv2d(e6_lrelu,
output_dim=self.gen_c[7],
initializer='He',
logger=self.logger,
name='e7_conv2d')
e7_batchnorm = tf_utils.norm(e7_conv2d,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='e7_norm')
e7_relu = tf_utils.lrelu(e7_batchnorm,
logger=self.logger,
name='e7_relu')
# D0: (2, 1) -> (3, 2)
# Stage1: (2, 1) -> (4, 2)
d0_deconv = tf_utils.deconv2d(e7_relu,
output_dim=self.gen_c[8],
initializer='He',
logger=self.logger,
name='d0_deconv2d')
# Stage2: (4, 2) -> (3, 2)
shapeA = e6_conv2d.get_shape().as_list()[1]
shapeB = d0_deconv.get_shape().as_list()[1] - e6_conv2d.get_shape(
).as_list()[1]
d0_split, _ = tf.split(d0_deconv, [shapeA, shapeB],
axis=1,
name='d0_split')
tf_utils.print_activations(d0_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d0_batchnorm = tf_utils.norm(d0_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d0_norm')
d0_drop = tf_utils.dropout(d0_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d0_dropout')
d0_concat = tf.concat([d0_drop, e6_batchnorm],
axis=3,
name='d0_concat')
d0_relu = tf_utils.relu(d0_concat,
logger=self.logger,
name='d0_relu')
# D1: (3, 2) -> (5, 4)
# Stage1: (3, 2) -> (6, 4)
d1_deconv = tf_utils.deconv2d(d0_relu,
output_dim=self.gen_c[9],
initializer='He',
logger=self.logger,
name='d1_deconv2d')
# Stage2: (6, 4) -> (5, 4)
shapeA = e5_batchnorm.get_shape().as_list()[1]
shapeB = d1_deconv.get_shape().as_list(
)[1] - e5_batchnorm.get_shape().as_list()[1]
d1_split, _ = tf.split(d1_deconv, [shapeA, shapeB],
axis=1,
name='d1_split')
tf_utils.print_activations(d1_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d1_batchnorm = tf_utils.norm(d1_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d1_norm')
d1_drop = tf_utils.dropout(d1_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d1_dropout')
d1_concat = tf.concat([d1_drop, e5_batchnorm],
axis=3,
name='d1_concat')
d1_relu = tf_utils.relu(d1_concat,
logger=self.logger,
name='d1_relu')
# D2: (5, 4) -> (10, 7)
# Stage1: (5, 4) -> (10, 8)
d2_deconv = tf_utils.deconv2d(d1_relu,
output_dim=self.gen_c[10],
initializer='He',
logger=self.logger,
name='d2_deconv2d')
# Stage2: (10, 8) -> (10, 7)
shapeA = e4_batchnorm.get_shape().as_list()[2]
shapeB = d2_deconv.get_shape().as_list(
)[2] - e4_batchnorm.get_shape().as_list()[2]
d2_split, _ = tf.split(d2_deconv, [shapeA, shapeB],
axis=2,
name='d2_split')
tf_utils.print_activations(d2_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d2_batchnorm = tf_utils.norm(d2_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d2_norm')
d2_drop = tf_utils.dropout(d2_batchnorm,
keep_prob=keep_rate,
logger=self.logger,
name='d2_dropout')
d2_concat = tf.concat([d2_drop, e4_batchnorm],
axis=3,
name='d2_concat')
d2_relu = tf_utils.relu(d2_concat,
logger=self.logger,
name='d2_relu')
# D3: (10, 7) -> (20, 13)
# Stage1: (10, 7) -> (20, 14)
d3_deconv = tf_utils.deconv2d(d2_relu,
output_dim=self.gen_c[11],
initializer='He',
logger=self.logger,
name='d3_deconv2d')
# Stage2: (20, 14) -> (20, 13)
shapeA = e3_batchnorm.get_shape().as_list()[2]
shapeB = d3_deconv.get_shape().as_list(
)[2] - e3_batchnorm.get_shape().as_list()[2]
d3_split, _ = tf.split(d3_deconv, [shapeA, shapeB],
axis=2,
name='d3_split_2')
tf_utils.print_activations(d3_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d3_batchnorm = tf_utils.norm(d3_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d3_norm')
d3_concat = tf.concat([d3_batchnorm, e3_batchnorm],
axis=3,
name='d3_concat')
d3_relu = tf_utils.relu(d3_concat,
logger=self.logger,
name='d3_relu')
# D4: (20, 13) -> (40, 25)
# Stage1: (20, 13) -> (40, 26)
d4_deconv = tf_utils.deconv2d(d3_relu,
output_dim=self.gen_c[12],
initializer='He',
logger=self.logger,
name='d4_deconv2d')
# Stage2: (40, 26) -> (40, 25)
shapeA = e2_batchnorm.get_shape().as_list()[2]
shapeB = d4_deconv.get_shape().as_list(
)[2] - e2_batchnorm.get_shape().as_list()[2]
d4_split, _ = tf.split(d4_deconv, [shapeA, shapeB],
axis=2,
name='d4_split')
tf_utils.print_activations(d4_split, logger=self.logger)
# Stage3: Batch norm, concatenation, and relu
d4_batchnorm = tf_utils.norm(d4_split,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d4_norm')
d4_concat = tf.concat([d4_batchnorm, e2_batchnorm],
axis=3,
name='d4_concat')
d4_relu = tf_utils.relu(d4_concat,
logger=self.logger,
name='d4_relu')
# D5: (40, 25, 256) -> (80, 50, 128)
d5_deconv = tf_utils.deconv2d(d4_relu,
output_dim=self.gen_c[13],
initializer='He',
logger=self.logger,
name='d5_deconv2d')
d5_batchnorm = tf_utils.norm(d5_deconv,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d5_norm')
d5_concat = tf.concat([d5_batchnorm, e1_batchnorm],
axis=3,
name='d5_concat')
d5_relu = tf_utils.relu(d5_concat,
logger=self.logger,
name='d5_relu')
# D6: (80, 50, 128) -> (160, 100, 64)
d6_deconv = tf_utils.deconv2d(d5_relu,
output_dim=self.gen_c[14],
initializer='He',
logger=self.logger,
name='d6_deconv2d')
d6_batchnorm = tf_utils.norm(d6_deconv,
_type=self.norm,
_ops=self._ops,
logger=self.logger,
name='d6_norm')
d6_concat = tf.concat([d6_batchnorm, e0_conv2d],
axis=3,
name='d6_concat')
d6_relu = tf_utils.relu(d6_concat,
logger=self.logger,
name='d6_relu')
# D7: (160, 100, 64) -> (320, 200, 1)
d7_deconv = tf_utils.deconv2d(d6_relu,
output_dim=self.gen_c[15],
initializer='He',
logger=self.logger,
name='d7_deconv2d')
output = tf_utils.tanh(d7_deconv,
logger=self.logger,
name='output_tanh')
# Set reuse=True for next call
self.reuse = True
self.variables = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
return output
def forward_network(self, inputImg, padding='SAME', reuse=False):
with tf.compat.v1.variable_scope(self.name, reuse=reuse):
# This part is for compatible between input size [640, 400] and [320, 200]
if self.resize_factor == 1.0:
# Stage 0
tf_utils.print_activations(inputImg, logger=self.logger)
s0_conv1 = tf_utils.conv2d(x=inputImg, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s0_conv1', logger=self.logger)
s0_conv1 = tf_utils.relu(s0_conv1, name='relu_s0_conv1', logger=self.logger)
s0_conv2 = tf_utils.conv2d(x=s0_conv1, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s0_conv2', logger=self.logger)
if self.use_batch_norm:
s0_conv2 = tf_utils.norm(s0_conv2, name='s0_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s0_conv2 = tf_utils.relu(s0_conv2, name='relu_s0_conv2', logger=self.logger)
# Stage 1
s1_maxpool = tf_utils.max_pool(x=s0_conv2, name='s1_maxpool2d', logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=s1_maxpool, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv1', logger=self.logger)
if self.use_batch_norm:
s1_conv1 = tf_utils.norm(s1_conv1, name='s1_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1', logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1, output_dim=self.conv_dims[1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv2', logger=self.logger)
if self.use_batch_norm:
s1_conv2 = tf_utils.norm(s1_conv2, name='s1_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2', logger=self.logger)
else:
# Stage 1
tf_utils.print_activations(inputImg, logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=inputImg, output_dim=self.conv_dims[0], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv1', logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1, name='relu_s1_conv1', logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1, output_dim=self.conv_dims[1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s1_conv2', logger=self.logger)
if self.use_batch_norm:
s1_conv2 = tf_utils.norm(s1_conv2, name='s1_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2, name='relu_s1_conv2', logger=self.logger)
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2, name='s2_maxpool2d', logger=self.logger)
s2_conv1 = tf_utils.conv2d(x=s2_maxpool, output_dim=self.conv_dims[2], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s2_conv1', logger=self.logger)
if self.use_batch_norm:
s2_conv1 = tf_utils.norm(s2_conv1, name='s2_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s2_conv1 = tf_utils.relu(s2_conv1, name='relu_s2_conv1', logger=self.logger)
s2_conv2 = tf_utils.conv2d(x=s2_conv1, output_dim=self.conv_dims[3], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s2_conv2', logger=self.logger)
if self.use_batch_norm:
s2_conv2 = tf_utils.norm(s2_conv2, name='s2_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s2_conv2 = tf_utils.relu(s2_conv2, name='relu_s2_conv2', logger=self.logger)
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2, name='s3_maxpool2d', logger=self.logger)
s3_conv1 = tf_utils.conv2d(x=s3_maxpool, output_dim=self.conv_dims[4], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s3_conv1', logger=self.logger)
if self.use_batch_norm:
s3_conv1 = tf_utils.norm(s3_conv1, name='s3_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s3_conv1 = tf_utils.relu(s3_conv1, name='relu_s3_conv1', logger=self.logger)
s3_conv2 = tf_utils.conv2d(x=s3_conv1, output_dim=self.conv_dims[5], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s3_conv2', logger=self.logger)
if self.use_batch_norm:
s3_conv2 = tf_utils.norm(s3_conv2, name='s3_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s3_conv2 = tf_utils.relu(s3_conv2, name='relu_s3_conv2', logger=self.logger)
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2, name='s4_maxpool2d', logger=self.logger)
s4_conv1 = tf_utils.conv2d(x=s4_maxpool, output_dim=self.conv_dims[6], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s4_conv1', logger=self.logger)
if self.use_batch_norm:
s4_conv1 = tf_utils.norm(s4_conv1, name='s4_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s4_conv1 = tf_utils.relu(s4_conv1, name='relu_s4_conv1', logger=self.logger)
s4_conv2 = tf_utils.conv2d(x=s4_conv1, output_dim=self.conv_dims[7], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s4_conv2', logger=self.logger)
if self.use_batch_norm:
s4_conv2 = tf_utils.norm(s4_conv2, name='s4_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s4_conv2 = tf_utils.relu(s4_conv2, name='relu_s4_conv2', logger=self.logger)
s4_conv2_drop = tf_utils.dropout(x=s4_conv2, keep_prob=self.ratePh, name='s4_dropout',
logger=self.logger)
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop, name='s5_maxpool2d', logger=self.logger)
s5_conv1 = tf_utils.conv2d(x=s5_maxpool, output_dim=self.conv_dims[8], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s5_conv1', logger=self.logger)
if self.use_batch_norm:
s5_conv1 = tf_utils.norm(s5_conv1, name='s5_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s5_conv1 = tf_utils.relu(s5_conv1, name='relu_s5_conv1', logger=self.logger)
s5_conv2 = tf_utils.conv2d(x=s5_conv1, output_dim=self.conv_dims[9], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s5_conv2', logger=self.logger)
if self.use_batch_norm:
s5_conv2 = tf_utils.norm(s5_conv2, name='s5_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s5_conv2 = tf_utils.relu(s5_conv2, name='relu_s5_conv2', logger=self.logger)
s5_conv2_drop = tf_utils.dropout(x=s5_conv2, keep_prob=self.ratePh, name='s5_dropout',
logger=self.logger)
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop, output_dim=self.conv_dims[10], k_h=2, k_w=2,
initializer='He', name='s6_deconv1', logger=self.logger)
if self.use_batch_norm:
s6_deconv1 = tf_utils.norm(s6_deconv1, name='s6_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_deconv1 = tf_utils.relu(s6_deconv1, name='relu_s6_deconv1', logger=self.logger)
# Cropping
w1 = s4_conv2_drop.get_shape().as_list()[2]
w2 = s6_deconv1.get_shape().as_list()[2] - s4_conv2_drop.get_shape().as_list()[2]
s6_deconv1_split, _ = tf.split(s6_deconv1, num_or_size_splits=[w1, w2], axis=2, name='axis2_split')
tf_utils.print_activations(s6_deconv1_split, logger=self.logger)
# Concat
s6_concat = tf_utils.concat(values=[s6_deconv1_split, s4_conv2_drop], axis=3, name='s6_axis3_concat',
logger=self.logger)
s6_conv2 = tf_utils.conv2d(x=s6_concat, output_dim=self.conv_dims[11], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s6_conv2', logger=self.logger)
if self.use_batch_norm:
s6_conv2 = tf_utils.norm(s6_conv2, name='s6_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_conv2 = tf_utils.relu(s6_conv2, name='relu_s6_conv2', logger=self.logger)
s6_conv3 = tf_utils.conv2d(x=s6_conv2, output_dim=self.conv_dims[12], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s6_conv3', logger=self.logger)
if self.use_batch_norm:
s6_conv3 = tf_utils.norm(s6_conv3, name='s6_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s6_conv3 = tf_utils.relu(s6_conv3, name='relu_s6_conv3', logger=self.logger)
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3, output_dim=self.conv_dims[13], k_h=2, k_w=2, initializer='He',
name='s7_deconv1', logger=self.logger)
if self.use_batch_norm:
s7_deconv1 = tf_utils.norm(s7_deconv1, name='s7_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_deconv1 = tf_utils.relu(s7_deconv1, name='relu_s7_deconv1', logger=self.logger)
# Concat
s7_concat = tf_utils.concat(values=[s7_deconv1, s3_conv2], axis=3, name='s7_axis3_concat',
logger=self.logger)
s7_conv2 = tf_utils.conv2d(x=s7_concat, output_dim=self.conv_dims[14], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s7_conv2', logger=self.logger)
if self.use_batch_norm:
s7_conv2 = tf_utils.norm(s7_conv2, name='s7_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_conv2 = tf_utils.relu(s7_conv2, name='relu_s7_conv2', logger=self.logger)
s7_conv3 = tf_utils.conv2d(x=s7_conv2, output_dim=self.conv_dims[15], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s7_conv3', logger=self.logger)
if self.use_batch_norm:
s7_conv3 = tf_utils.norm(s7_conv3, name='s7_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s7_conv3 = tf_utils.relu(s7_conv3, name='relu_s7_conv3', logger=self.logger)
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3, output_dim=self.conv_dims[16], k_h=2, k_w=2, initializer='He',
name='s8_deconv1', logger=self.logger)
if self.use_batch_norm:
s8_deconv1 = tf_utils.norm(s8_deconv1, name='s8_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_deconv1 = tf_utils.relu(s8_deconv1, name='relu_s8_deconv1', logger=self.logger)
# Concat
s8_concat = tf_utils.concat(values=[s8_deconv1,s2_conv2], axis=3, name='s8_axis3_concat',
logger=self.logger)
s8_conv2 = tf_utils.conv2d(x=s8_concat, output_dim=self.conv_dims[17], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s8_conv2', logger=self.logger)
if self.use_batch_norm:
s8_conv2 = tf_utils.norm(s8_conv2, name='s8_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_conv2 = tf_utils.relu(s8_conv2, name='relu_s8_conv2', logger=self.logger)
s8_conv3 = tf_utils.conv2d(x=s8_conv2, output_dim=self.conv_dims[18], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s8_conv3', logger=self.logger)
if self.use_batch_norm:
s8_conv3 = tf_utils.norm(s8_conv3, name='s8_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s8_conv3 = tf_utils.relu(s8_conv3, name='relu_conv3', logger=self.logger)
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3, output_dim=self.conv_dims[19], k_h=2, k_w=2,
initializer='He', name='s9_deconv1', logger=self.logger)
if self.use_batch_norm:
s9_deconv1 = tf_utils.norm(s9_deconv1, name='s9_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_deconv1 = tf_utils.relu(s9_deconv1, name='relu_s9_deconv1', logger=self.logger)
# Concat
s9_concat = tf_utils.concat(values=[s9_deconv1, s1_conv2], axis=3, name='s9_axis3_concat',
logger=self.logger)
s9_conv2 = tf_utils.conv2d(x=s9_concat, output_dim=self.conv_dims[20], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s9_conv2', logger=self.logger)
if self.use_batch_norm:
s9_conv2 = tf_utils.norm(s9_conv2, name='s9_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_conv2 = tf_utils.relu(s9_conv2, name='relu_s9_conv2', logger=self.logger)
s9_conv3 = tf_utils.conv2d(x=s9_conv2, output_dim=self.conv_dims[21], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s9_conv3', logger=self.logger)
if self.use_batch_norm:
s9_conv3 = tf_utils.norm(s9_conv3, name='s9_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s9_conv3 = tf_utils.relu(s9_conv3, name='relu_s9_conv3', logger=self.logger)
if self.resize_factor == 1.0:
s10_deconv1 = tf_utils.deconv2d(x=s9_conv3, output_dim=self.conv_dims[-1], k_h=2, k_w=2,
initializer='He', name='s10_deconv1', logger=self.logger)
if self.use_batch_norm:
s10_deconv1 = tf_utils.norm(s10_deconv1, name='s10_norm0', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_deconv1 = tf_utils.relu(s10_deconv1, name='relu_s10_deconv1', logger=self.logger)
# Concat
s10_concat = tf_utils.concat(values=[s10_deconv1, s0_conv2], axis=3, name='s10_axis3_concat',
logger=self.logger)
s10_conv2 = tf_utils.conv2d(s10_concat, output_dim=self.conv_dims[-1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s10_conv2', logger=self.logger)
if self.use_batch_norm:
s10_conv2 = tf_utils.norm(s10_conv2, name='s10_norm1', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_conv2 = tf_utils.relu(s10_conv2, name='relu_s10_conv2', logger=self.logger)
s10_conv3 = tf_utils.conv2d(x=s10_conv2, output_dim=self.conv_dims[-1], k_h=3, k_w=3, d_h=1, d_w=1,
padding=padding, initializer='He', name='s10_conv3', logger=self.logger)
if self.use_batch_norm:
s10_conv3 = tf_utils.norm(s10_conv3, name='s10_norm2', _type='batch', _ops=self._ops,
is_train=self.trainMode, logger=self.logger)
s10_conv3 = tf_utils.relu(s10_conv3, name='relu_s10_conv3', logger=self.logger)
output = tf_utils.conv2d(s10_conv3, output_dim=self.numClasses, k_h=1, k_w=1, d_h=1, d_w=1,
padding=padding, initializer='He', name='output', logger=self.logger)
else:
output = tf_utils.conv2d(s9_conv3, output_dim=self.numClasses, k_h=1, k_w=1, d_h=1, d_w=1,
padding=padding, initializer='He', name='output', logger=self.logger)
return output
def u_net(self):
# Stage 1
tf_utils.print_activations(self.inp_img, logger=self.logger)
s1_conv1 = tf_utils.conv2d(x=self.inp_img,
output_dim=self.conv_dims[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s1_conv1',
logger=self.logger)
s1_conv1 = tf_utils.relu(s1_conv1,
name='relu_s1_conv1',
logger=self.logger)
s1_conv2 = tf_utils.conv2d(x=s1_conv1,
output_dim=self.conv_dims[1],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s1_conv2',
logger=self.logger)
s1_conv2 = tf_utils.relu(s1_conv2,
name='relu_s1_conv2',
logger=self.logger)
# Stage 2
s2_maxpool = tf_utils.max_pool(x=s1_conv2,
name='s2_maxpool',
logger=self.logger)
s2_conv1 = tf_utils.conv2d(x=s2_maxpool,
output_dim=self.conv_dims[2],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s2_conv1',
logger=self.logger)
s2_conv1 = tf_utils.relu(s2_conv1,
name='relu_s2_conv1',
logger=self.logger)
s2_conv2 = tf_utils.conv2d(x=s2_conv1,
output_dim=self.conv_dims[3],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s2_conv2',
logger=self.logger)
s2_conv2 = tf_utils.relu(s2_conv2,
name='relu_s2_conv2',
logger=self.logger)
# Stage 3
s3_maxpool = tf_utils.max_pool(x=s2_conv2,
name='s3_maxpool',
logger=self.logger)
s3_conv1 = tf_utils.conv2d(x=s3_maxpool,
output_dim=self.conv_dims[4],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s3_conv1',
logger=self.logger)
s3_conv1 = tf_utils.relu(s3_conv1,
name='relu_s3_conv1',
logger=self.logger)
s3_conv2 = tf_utils.conv2d(x=s3_conv1,
output_dim=self.conv_dims[5],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s3_conv2',
logger=self.logger)
s3_conv2 = tf_utils.relu(s3_conv2,
name='relu_s3_conv2',
logger=self.logger)
# Stage 4
s4_maxpool = tf_utils.max_pool(x=s3_conv2,
name='s4_maxpool',
logger=self.logger)
s4_conv1 = tf_utils.conv2d(x=s4_maxpool,
output_dim=self.conv_dims[6],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s4_conv1',
logger=self.logger)
s4_conv1 = tf_utils.relu(s4_conv1,
name='relu_s4_conv1',
logger=self.logger)
s4_conv2 = tf_utils.conv2d(x=s4_conv1,
output_dim=self.conv_dims[7],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s4_conv2',
logger=self.logger)
s4_conv2 = tf_utils.relu(s4_conv2,
name='relu_s4_conv2',
logger=self.logger)
s4_conv2_drop = tf_utils.dropout(x=s4_conv2,
keep_prob=self.keep_prob,
name='s4_conv2_dropout',
logger=self.logger)
# Stage 5
s5_maxpool = tf_utils.max_pool(x=s4_conv2_drop,
name='s5_maxpool',
logger=self.logger)
s5_conv1 = tf_utils.conv2d(x=s5_maxpool,
output_dim=self.conv_dims[8],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s5_conv1',
logger=self.logger)
s5_conv1 = tf_utils.relu(s5_conv1,
name='relu_s5_conv1',
logger=self.logger)
s5_conv2 = tf_utils.conv2d(x=s5_conv1,
output_dim=self.conv_dims[9],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s5_conv2',
logger=self.logger)
s5_conv2 = tf_utils.relu(s5_conv2,
name='relu_s5_conv2',
logger=self.logger)
s5_conv2_drop = tf_utils.dropout(x=s5_conv2,
keep_prob=self.keep_prob,
name='s5_conv2_dropout',
logger=self.logger)
# Stage 6
s6_deconv1 = tf_utils.deconv2d(x=s5_conv2_drop,
output_dim=self.conv_dims[10],
k_h=2,
k_w=2,
initializer='He',
name='s6_deconv1',
logger=self.logger)
s6_deconv1 = tf_utils.relu(s6_deconv1,
name='relu_s6_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s4_conv2_drop.get_shape().as_list()[1:3]
h2, w2 = s6_deconv1.get_shape().as_list()[1:3]
s4_conv2_crop = tf.image.crop_to_bounding_box(
image=s4_conv2_drop,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s4_conv2_crop, logger=self.logger)
s6_concat = tf_utils.concat(values=[s4_conv2_crop, s6_deconv1],
axis=3,
name='s6_concat',
logger=self.logger)
s6_conv2 = tf_utils.conv2d(x=s6_concat,
output_dim=self.conv_dims[11],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s6_conv2',
logger=self.logger)
s6_conv2 = tf_utils.relu(s6_conv2,
name='relu_s6_conv2',
logger=self.logger)
s6_conv3 = tf_utils.conv2d(x=s6_conv2,
output_dim=self.conv_dims[12],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s6_conv3',
logger=self.logger)
s6_conv3 = tf_utils.relu(s6_conv3,
name='relu_s6_conv3',
logger=self.logger)
# Stage 7
s7_deconv1 = tf_utils.deconv2d(x=s6_conv3,
output_dim=self.conv_dims[13],
k_h=2,
k_w=2,
initializer='He',
name='s7_deconv1',
logger=self.logger)
s7_deconv1 = tf_utils.relu(s7_deconv1,
name='relu_s7_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s3_conv2.get_shape().as_list()[1:3]
h2, w2 = s7_deconv1.get_shape().as_list()[1:3]
s3_conv2_crop = tf.image.crop_to_bounding_box(
image=s3_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s3_conv2_crop, logger=self.logger)
s7_concat = tf_utils.concat(values=[s3_conv2_crop, s7_deconv1],
axis=3,
name='s7_concat',
logger=self.logger)
s7_conv2 = tf_utils.conv2d(x=s7_concat,
output_dim=self.conv_dims[14],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s7_conv2',
logger=self.logger)
s7_conv2 = tf_utils.relu(s7_conv2,
name='relu_s7_conv2',
logger=self.logger)
s7_conv3 = tf_utils.conv2d(x=s7_conv2,
output_dim=self.conv_dims[15],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s7_conv3',
logger=self.logger)
s7_conv3 = tf_utils.relu(s7_conv3,
name='relu_s7_conv3',
logger=self.logger)
# Stage 8
s8_deconv1 = tf_utils.deconv2d(x=s7_conv3,
output_dim=self.conv_dims[16],
k_h=2,
k_w=2,
initializer='He',
name='s8_deconv1',
logger=self.logger)
s8_deconv1 = tf_utils.relu(s8_deconv1,
name='relu_s8_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s2_conv2.get_shape().as_list()[1:3]
h2, w2 = s8_deconv1.get_shape().as_list()[1:3]
s2_conv2_crop = tf.image.crop_to_bounding_box(
image=s2_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s2_conv2_crop, logger=self.logger)
s8_concat = tf_utils.concat(values=[s2_conv2_crop, s8_deconv1],
axis=3,
name='s8_concat',
logger=self.logger)
s8_conv2 = tf_utils.conv2d(x=s8_concat,
output_dim=self.conv_dims[17],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s8_conv2',
logger=self.logger)
s8_conv2 = tf_utils.relu(s8_conv2,
name='relu_s8_conv2',
logger=self.logger)
s8_conv3 = tf_utils.conv2d(x=s8_conv2,
output_dim=self.conv_dims[18],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s8_conv3',
logger=self.logger)
s8_conv3 = tf_utils.relu(s8_conv3,
name='relu_conv3',
logger=self.logger)
# Stage 9
s9_deconv1 = tf_utils.deconv2d(x=s8_conv3,
output_dim=self.conv_dims[19],
k_h=2,
k_w=2,
initializer='He',
name='s9_deconv1',
logger=self.logger)
s9_deconv1 = tf_utils.relu(s9_deconv1,
name='relu_s9_deconv1',
logger=self.logger)
# Cropping
h1, w1 = s1_conv2.get_shape().as_list()[1:3]
h2, w2 = s9_deconv1.get_shape().as_list()[1:3]
s1_conv2_crop = tf.image.crop_to_bounding_box(
image=s1_conv2,
offset_height=int(0.5 * (h1 - h2)),
offset_width=int(0.5 * (w1 - w2)),
target_height=h2,
target_width=w2)
tf_utils.print_activations(s1_conv2_crop, logger=self.logger)
s9_concat = tf_utils.concat(values=[s1_conv2_crop, s9_deconv1],
axis=3,
name='s9_concat',
logger=self.logger)
s9_conv2 = tf_utils.conv2d(x=s9_concat,
output_dim=self.conv_dims[20],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s9_conv2',
logger=self.logger)
s9_conv2 = tf_utils.relu(s9_conv2,
name='relu_s9_conv2',
logger=self.logger)
s9_conv3 = tf_utils.conv2d(x=s9_conv2,
output_dim=self.conv_dims[21],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
padding='VALID',
initializer='He',
name='s9_conv3',
logger=self.logger)
s9_conv3 = tf_utils.relu(s9_conv3,
name='relu_s9_conv3',
logger=self.logger)
self.pred = tf_utils.conv2d(x=s9_conv3,
output_dim=self.conv_dims[22],
k_h=1,
k_w=1,
d_h=1,
d_w=1,
padding='SAME',
initializer='He',
name='output',
logger=self.logger)
def __init__(self, input_dim, output_dim=[1000, 1000, 10], optimizer=None, use_dropout=True, lr=0.001,
weight_decay=1e-4, random_seed=123, is_train=True, log_dir=None, name=None):
self.name = name
self.is_train = is_train
self.log_dir = log_dir
self.cur_lr = None
self.logger, self.file_handler, self.stream_handler = utils.init_logger(log_dir=self.log_dir,
name=self.name,
is_train=self.is_train)
with tf.variable_scope(self.name):
# Placeholders for inputs
self.X = tf.placeholder(dtype=tf.float32, shape=[None, input_dim], name='X')
tf_utils.print_activations(self.X, logger=self.logger if self.is_train else None)
self.y = tf.placeholder(dtype=tf.float32, shape=[None, output_dim[-1]], name='y')
self.y_cls = tf.math.argmax(input=self.y, axis=1)
self.keep_prob = tf.placeholder(tf.float32, name='keep_prob')
# Placeholders for TensorBoard
self.train_acc = tf.placeholder(tf.float32, name='train_acc')
self.val_acc = tf.placeholder(tf.float32, name='val_acc')
net = self.X
for idx in range(len(output_dim) - 1):
net = tf_utils.linear(x=net,
output_size=output_dim[idx],
name='fc'+str(idx),
logger=self.logger if self.is_train else None)
if use_dropout:
net = tf_utils.dropout(x=net,
keep_prob=self.keep_prob,
seed=random_seed,
name='dropout'+str(idx),
logger=self.logger if self.is_train else None)
net = tf_utils.relu(x=net,
name='relu'+str(idx),
logger=self.logger if self.is_train else None)
# Last predict layer
self.y_pred = tf_utils.linear(net, output_size=output_dim[-1], name='last_fc')
tf_utils.print_activations(self.y_pred, logger=self.logger if self.is_train else None)
# Loss = data loss + regularization term
self.data_loss = tf.math.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=self.y_pred, labels=self.y))
self.reg_term = weight_decay * tf.reduce_sum(
[tf.nn.l2_loss(weight) for weight in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)])
self.loss = self.data_loss + self.reg_term
# Optimizer
self.train_op, self.cur_lr = optimizer_fn(optimizer, lr=lr, loss=self.loss, name=self.name)
# Accuracy etc
self.y_pred_cls = tf.math.argmax(input=self.y_pred, axis=1)
correct_prediction = tf.math.equal(self.y_pred_cls, self.y_cls)
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, dtype=tf.float32)) * 100.
self._tensorboard()
tf_utils.show_all_variables(logger=self.logger if self.is_train else None)