def __init__(self, layers=None, skip_connections=False, has_translator=True):
with tf.variable_scope('encoder-decoder'):
if layers == None:
layers = []
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_1_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_1_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_1', skip_connection=skip_connections))
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_2_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_2_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_2', skip_connection=skip_connections))
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_3_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_2'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_3'))
self.inputs = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS], name='inputs')
self.targets = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1], name='targets')
self.is_training = tf.placeholder_with_default(False, [], name='is_training')
self.description = ""
self.layers = {}
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net, is_training=False)
layers.reverse()
Conv2d.reverse_global_variables()
#midfield
if(has_translator == True):
old_shape = net.get_shape()
o_s = old_shape.as_list()
feature_len = o_s[1]*o_s[2]*o_s[3]
net = tf.reshape(net, [-1, feature_len])
for i in range(3):
net = slim.fully_connected(net, feature_len, scope="fc_{}".format(i+1))
self.fc_vars = tf.contrib.framework.get_variables("fc_{}".format(i+1))
net = tf.reshape(net, [-1, o_s[1], o_s[2], o_s[3]])
# DECODER
layers_len = len(layers)
for i, layer in enumerate(layers):
if i == (layers_len-1):
self.segmentation_result = layer.create_layer_reversed(net, prev_layer=self.layers[layer.name], last_layer=True, is_training=False)
else:
net = layer.create_layer_reversed(net, prev_layer=self.layers[layer.name], is_training=False)
self.final_result = self.segmentation_result
self.variables = tf.contrib.framework.get_variables(scope='encoder-decoder')
def __init__(self, layers=None, skip_connections=True):
if layers == None:
layers = []
layers.append(
Conv2d(kernel_size=3,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_1_1'))
layers.append(
Conv2d(kernel_size=3,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_1_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_1',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=3,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_2_1'))
layers.append(
Conv2d(kernel_size=3,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_2_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_2',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=3,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_3_1'))
layers.append(
Conv2d(kernel_size=3,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_3_2'))
self.inputs = tf.placeholder(
tf.float32,
[None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
self.targets = tf.placeholder(
tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1],
name='targets')
self.is_training = tf.placeholder_with_default(False, [],
name='is_training')
self.description = ""
self.layers = {}
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
layers.reverse()
Conv2d.reverse_global_variables()
#midfield
self.feature_set = tf.reshape(
net,
[-1,
net.get_shape()[1] * net.get_shape()[2] * net.get_shape()[3]])
# DECODER
layers_len = len(layers)
for i, layer in enumerate(layers):
if i == (layers_len - 1):
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name], last_layer=True)
self.final_result = slim.conv2d(net,
1,
1,
scope="last_conv",
activation_fn=None)
else:
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name])
# MSE loss
output = self.final_result
inputv = self.targets
mean = tf.reduce_mean(tf.square(output - inputv))
self.cost1 = mean
# Reconstruct with MS_SSIM loss function
self.train_op = tf.train.AdamOptimizer(
learning_rate=tf.train.polynomial_decay(
0.01, 1, 1000, 0.0001)).minimize(self.cost1)
with tf.name_scope('accuracy'):
correct_pred = tf.py_func(msssim.MultiScaleSSIM,
[self.final_result, self.targets],
tf.float32)
self.accuracy = correct_pred
self.mse = tf.reduce_mean(
tf.square(self.final_result - self.targets))
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
def __init__(self, layers=None, skip_connections=False):
with tf.variable_scope('encoder-decoder'):
if layers == None:
layers = []
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_1_1'))
#layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_1_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_1',
skip_connection=skip_connections))
#layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_2_1'))
#layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_2_2'))
#layers.append(MaxPool2d(kernel_size=2, name='max_2', skip_connection=skip_connections))
#layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_3_1'))
#layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_2'))
#layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_3'))
self.inputs = tf.placeholder(tf.float32, [
None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS
],
name='inputs')
self.targets = tf.placeholder(
tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1],
name='targets')
self.is_training = tf.placeholder_with_default(False, [],
name='is_training')
self.nepoch = tf.placeholder(tf.int32, [], name="nepoch")
self.description = ""
self.layers = {}
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
layers.reverse()
Conv2d.reverse_global_variables()
# DECODER
layers_len = len(layers)
for i, layer in enumerate(layers):
if i == (layers_len - 1):
self.segmentation_result = layer.create_layer_reversed(
net,
prev_layer=self.layers[layer.name],
last_layer=True)
else:
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name])
self.variables = tf.contrib.framework.get_variables(
scope='encoder-decoder')
self.final_result = self.segmentation_result
self.rec1 = Recognizer(self.final_result, reuse=tf.AUTO_REUSE)
self.rec2 = Recognizer(self.inputs, reuse=tf.AUTO_REUSE)
self.rec3 = Recognizer(self.targets, reuse=tf.AUTO_REUSE)
# MSE loss
# Expression Removal with MSE loss function
output = self.segmentation_result
inputv = self.targets
mean = tf.reduce_mean(tf.square(output - inputv))
# Recognition feature sets
rec1_loss = self.rec1.modelo
rec2_loss = self.rec2.modelo
rec3_loss = self.rec3.modelo
output_weight = tf.constant(3, shape=[], dtype=tf.float32)
cost_rec1 = tf.multiply(
output_weight,
tf.reduce_sum(tf.reduce_mean(tf.abs(rec1_loss - rec3_loss), 0)))
cost_rec2 = tf.reduce_sum(
tf.reduce_mean(tf.abs(rec1_loss - rec2_loss), 0))
self.cost_rec = cost_rec1 + cost_rec2
self.cost_mse = mean
self.train_op_rec = tf.train.AdamOptimizer(
learning_rate=tf.train.polynomial_decay(
10e-5, self.nepoch, 10000, 10e-7)).minimize(self.cost_rec)
self.train_op_mse = tf.train.AdamOptimizer(
learning_rate=tf.train.polynomial_decay(
10e-4, self.nepoch, 10000, 10e-5)).minimize(self.cost_mse)
def __init__(self, layers=None, per_image_standardization=True, batch_norm=True, skip_connections=True):
# Define network - ENCODER (decoder will be symmetric).
s=3
if layers == None:
layers = []
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_1_1')) # 61 x 61 x 64
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_1_2')) # 55 x 55 x 64
layers.append(MaxPool2d(kernel_size=2, name='max_1', skip_connection=skip_connections)) # 54 x 54
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_2_1')) # 24 x 24 x 64
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_2_2')) # 18
layers.append(MaxPool2d(kernel_size=2, name='max_2', skip_connection=skip_connections)) # 17
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_1')) # 24 x 24 x 64
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_2')) # 18
layers.append(MaxPool2d(kernel_size=2, name='max_3', skip_connection=skip_connections)) # 17
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_4_1')) # 6
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_4_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_4', skip_connection=skip_connections))
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_5_1')) # 6
layers.append(Conv2d(kernel_size=s, strides=[1, 1, 1, 1], output_channels=64, name='conv_5_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_5'))
self.inputs = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
self.targets = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1], name='targets')
self.is_training = tf.placeholder_with_default(False, [], name='is_training')
self.description = ""
self.layers = {}
if per_image_standardization:
list_of_images_norm = tf.map_fn(tf.image.per_image_standardization, self.inputs)
net = tf.stack(list_of_images_norm)
else:
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
print("Current input shape: ", net.get_shape())
layers.reverse()
Conv2d.reverse_global_variables()
# DECODER
for layer in layers:
net = layer.create_layer_reversed(net, prev_layer=self.layers[layer.name])
self.segmentation_result = tf.sigmoid(net)
print('segmentation_result.shape: {}, targets.shape: {}'.format(self.segmentation_result.get_shape(),
self.targets.get_shape()))
# MSE loss
self.cost = tf.sqrt(tf.reduce_mean(tf.square(self.segmentation_result - self.targets)))
self.train_op = tf.train.AdamOptimizer().minimize(self.cost)
with tf.name_scope('accuracy'):
argmax_probs = tf.round(self.segmentation_result) # 0x1
correct_pred = tf.cast(tf.equal(argmax_probs, self.targets), tf.float32)
self.accuracy = tf.reduce_mean(correct_pred)
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
def __init__(self, layers = None, batch_norm=True, skip_connections=True, pretrain= False):
# Define network - ENCODER (decoder will be symmetric).
self.IMAGE_HEIGHT = 2048
self.IMAGE_WIDTH = 1
self.IMAGE_CHANNELS = 1
self.is_training = tf.placeholder_with_default(False, [], name='is_training')
if pretrain:
# load pretrained weights from TICNN
pre_train_path = '/home/dawei/cyril/1d_conv_encoder_decoder/pretrained_TICNN/'
W_conv1 = tf.constant(np.load(pre_train_path + 'W_conv1.npy'))
W_conv2 = tf.constant(np.load(pre_train_path + 'W_conv2.npy'))
W_conv3 = tf.constant(np.load(pre_train_path + 'W_conv3.npy'))
W_conv4 = tf.constant(np.load(pre_train_path + 'W_conv4.npy'))
W_conv5 = tf.constant(np.load(pre_train_path + 'W_conv5.npy'))
else:
W_conv1 = None
W_conv2 = None
W_conv3 = None
W_conv4 = None
W_conv5 = None
if layers == None:
layers = []
layers.append(Conv2d(kernel_size=64, strides=[1, 8, 1, 1], output_channels=16, name='conv_1', is_training = self.is_training, initializer = W_conv1))
layers.append(MaxPool2d(kernel_size=[2, 1], name='max_1',skip_connection=True and skip_connections))
#
layers.append(Conv2d(kernel_size=3, strides=[1, 1, 1, 1], output_channels=32, name='conv_2', is_training = self.is_training, initializer = W_conv2))
layers.append(MaxPool2d(kernel_size=[2, 1], name='max_2',skip_connection=True and skip_connections))
#
layers.append(Conv2d(kernel_size=3, strides=[1, 1, 1, 1], output_channels=64, name='conv_3', is_training = self.is_training, initializer = W_conv3))
layers.append(MaxPool2d(kernel_size=[2, 1], name='max_3',skip_connection=True and skip_connections))
layers.append(Conv2d(kernel_size=3, strides=[1, 1, 1, 1], output_channels=64, name='conv_4', is_training = self.is_training, initializer = W_conv4))
layers.append(MaxPool2d(kernel_size=[2, 1], name='max_4',skip_connection=True and skip_connections))
layers.append(Conv2d(kernel_size=3, strides=[1, 1, 1, 1], output_channels=64, name='conv_5', is_training = self.is_training, initializer = W_conv5))
layers.append(MaxPool2d(kernel_size=[2, 1], name='max_5'))
self.inputs = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
self.targets = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1], name='targets')
self.description = ""
self.layers = {}
net = self.inputs #(None, 2048, 1, 1)
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
print("Current input shape: ", self.inputs.get_shape())
layers.reverse() # reverse the list of layers
Conv2d.reverse_global_variables()
# DECODER
for layer in layers:
net = layer.create_layer_reversed(net, prev_layer=self.layers[layer.name])
#self.segmentation_result = tf.sigmoid(net)
self.segmentation_result = net
print('segmentation_result.shape: {}, targets.shape: {}'.format(self.segmentation_result.get_shape(),
self.targets.get_shape()))
# RMSE loss
self.cost = tf.sqrt(tf.reduce_mean(tf.square(self.segmentation_result - self.targets)))
self.train_op = tf.train.AdamOptimizer(0.001).minimize(self.cost)
tf.summary.scalar('rmse_cost', self.cost)
self.summary = tf.summary.merge_all()
def __init__(self, layers=None, skip_connections=False):
with tf.variable_scope('encoder-decoder'):
if layers == None:
layers = []
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_1_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_1_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_1',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_2_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_2_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_2',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_3_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_3_2'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_3_3'))
self.inputs = tf.placeholder(tf.float32, [
None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS
],
name='inputs')
self.targets = tf.placeholder(
tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1],
name='targets')
self.is_training = tf.placeholder_with_default(False, [],
name='is_training')
self.description = ""
self.layers = {}
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
layers.reverse()
Conv2d.reverse_global_variables()
#midfield
'''old_shape = net.get_shape()
o_s = old_shape.as_list()
feature_len = o_s[1]*o_s[2]*o_s[3]
net = tf.reshape(net, [-1, feature_len])
for i in range(3):
net = slim.fully_connected(net, feature_len, scope="fc_{}".format(i+1))
self.fc_vars = tf.contrib.framework.get_variables("fc_{}".format(i+1))
net = tf.reshape(net, [-1, o_s[1], o_s[2], o_s[3]])'''
# DECODER
layers_len = len(layers)
for i, layer in enumerate(layers):
if i == (layers_len - 1):
self.segmentation_result = layer.create_layer_reversed(
net,
prev_layer=self.layers[layer.name],
last_layer=True)
else:
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name])
self.variables = tf.contrib.framework.get_variables(
scope='encoder-decoder')
self.final_result = self.segmentation_result
self.rec1 = Recognizer(self.final_result, reuse=tf.AUTO_REUSE)
self.rec2 = Recognizer(self.inputs, reuse=tf.AUTO_REUSE)
self.rec3 = Recognizer(self.targets, reuse=tf.AUTO_REUSE)
# MSE loss
# Expression Removal with MSE loss function
output = self.segmentation_result
inputv = self.targets
mean = tf.reduce_mean(tf.square(output - inputv))
# Recognition feature sets
rec1_loss = self.rec1.modelo
rec2_loss = self.rec2.modelo
rec3_loss = self.rec3.modelo
output_weight = tf.constant(3, shape=[], dtype=tf.float32)
cost_rec1 = tf.multiply(
output_weight,
tf.reduce_sum(tf.reduce_mean(tf.abs(rec1_loss - rec3_loss), 0)))
cost_rec2 = tf.reduce_sum(
tf.reduce_mean(tf.abs(rec1_loss - rec2_loss), 0))
# Cost based on recognition
final_weight = tf.constant(10, shape=[], dtype=tf.float32)
self.cost_rec = tf.multiply(final_weight, mean) + cost_rec1 + cost_rec2
#self.cost_rec = tf.constant(0, dtype=tf.float32)
self.cost_mse = mean
self.train_op_rec = tf.train.AdamOptimizer(
learning_rate=tf.train.polynomial_decay(
0.00001, 1, 10000, 0.0000001)).minimize(self.cost_rec)
self.train_op_mse = tf.train.AdamOptimizer(
learning_rate=tf.train.polynomial_decay(
0.0001, 1, 10000, 0.00001)).minimize(self.cost_mse)
with tf.name_scope('accuracy'):
correct_pred = tf.py_func(msssim.MultiScaleSSIM,
[self.final_result, self.targets],
tf.float32)
self.accuracy = correct_pred
self.mse = tf.reduce_mean(
tf.square(self.final_result - self.targets))
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
def __init__(self,
layers=None,
per_image_standardization=True,
batch_norm=True,
skip_connections=True):
# Define network - ENCODER (decoder will be symmetric).
if layers == None:
layers = []
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_1_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_1_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_1',
skip_connection=True and skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_2_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_2_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_2',
skip_connection=True and skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_3_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_3_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_3'))
self.inputs = tf.placeholder(
tf.float32,
[None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
# when output image has multiple dimensions, and each pixel has one of n class predictions
#self.targets = tf.placeholder(tf.int32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS], name='targets')
self.targets = tf.placeholder(
tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1],
name='targets')
self.is_training = tf.placeholder_with_default(False, [],
name='is_training')
self.description = ""
self.layers = {}
if per_image_standardization:
list_of_images_norm = tf.map_fn(tf.image.per_image_standardization,
self.inputs)
net = tf.stack(list_of_images_norm)
else:
net = self.inputs
# ENCODER
for layer in layers:
# layer is Conv2d type, Conv2d has create_layer method
# note how we are passing output of graph net as input to next layer
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
print("Current input shape: ", net.get_shape())
# just reversing the array
layers.reverse()
Conv2d.reverse_global_variables()
# DECODER
for layer in layers:
# use prev_layer to reverse
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name])
self.segmentation_result = tf.sigmoid(net)
# cross entropy loss
# can't use cross entropy since output image is not image mask, rather an image with float values
print('segmentation_result.shape: {}, targets.shape: {}'.format(
self.segmentation_result.get_shape(), self.targets.get_shape()))
# targets_as_classes = tf.reshape(self.targets, [-1, self.IMAGE_HEIGHT, self.IMAGE_WIDTH])
#self.cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.segmentation_result, labels=targets_as_classes))
# MSE loss
self.cost = tf.sqrt(
tf.reduce_mean(tf.square(self.segmentation_result - self.targets)))
self.train_op = tf.train.AdamOptimizer().minimize(self.cost)
with tf.name_scope('accuracy'):
argmax_probs = tf.round(self.segmentation_result) # 0x1
correct_pred = tf.cast(tf.equal(argmax_probs, self.targets),
tf.float32)
self.accuracy = tf.reduce_mean(correct_pred)
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
def __init__(self, layers = None, per_image_standardization=True, batch_norm=True, skip_connections=True):
# Define network - ENCODER (decoder will be symmetric).
if layers == None:
layers = []
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_1_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_1_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_1', skip_connection=True and skip_connections))
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_2_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_2_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_2', skip_connection=True and skip_connections))
layers.append(Conv2d(kernel_size=7, strides=[1, 2, 2, 1], output_channels=64, name='conv_3_1'))
layers.append(Conv2d(kernel_size=7, strides=[1, 1, 1, 1], output_channels=64, name='conv_3_2'))
layers.append(MaxPool2d(kernel_size=2, name='max_3'))
self.inputs = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
# when output image has multiple dimensions, and each pixel has one of n class predictions
#self.targets = tf.placeholder(tf.int32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS], name='targets')
self.targets = tf.placeholder(tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1], name='targets')
self.is_training = tf.placeholder_with_default(False, [], name='is_training')
self.description = ""
self.layers = {}
if per_image_standardization:
list_of_images_norm = tf.map_fn(tf.image.per_image_standardization, self.inputs)
net = tf.stack(list_of_images_norm)
else:
net = self.inputs
# ENCODER
for layer in layers:
# layer is Conv2d type, Conv2d has create_layer method
# note how we are passing output of graph net as input to next layer
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
print("Current input shape: ", net.get_shape())
# just reversing the array
layers.reverse()
Conv2d.reverse_global_variables()
# DECODER
for layer in layers:
# use prev_layer to reverse
net = layer.create_layer_reversed(net, prev_layer=self.layers[layer.name])
self.segmentation_result = tf.sigmoid(net)
# cross entropy loss
# can't use cross entropy since output image is not image mask, rather an image with float values
print('segmentation_result.shape: {}, targets.shape: {}'.format(self.segmentation_result.get_shape(),
self.targets.get_shape()))
# targets_as_classes = tf.reshape(self.targets, [-1, self.IMAGE_HEIGHT, self.IMAGE_WIDTH])
#self.cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.segmentation_result, labels=targets_as_classes))
# MSE loss
self.cost = tf.sqrt(tf.reduce_mean(tf.square(self.segmentation_result - self.targets)))
self.train_op = tf.train.AdamOptimizer().minimize(self.cost)
with tf.name_scope('accuracy'):
argmax_probs = tf.round(self.segmentation_result) # 0x1
correct_pred = tf.cast(tf.equal(argmax_probs, self.targets), tf.float32)
self.accuracy = tf.reduce_mean(correct_pred)
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
def __init__(self,
layers=None,
per_image_standardization=False,
batch_norm=True,
skip_connections=True):
# Define network - ENCODER (decoder will be symmetric).
if layers == None:
layers = []
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_1_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_1_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_1',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_2_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_2_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_2',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_3_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_3_2'))
layers.append(
MaxPool2d(kernel_size=2,
name='max_3',
skip_connection=skip_connections))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_4_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_4_2'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 2, 2, 1],
output_channels=64,
name='conv_5_1'))
layers.append(
Conv2d(kernel_size=7,
strides=[1, 1, 1, 1],
output_channels=64,
name='conv_5_2'))
self.inputs = tf.placeholder(
tf.float32,
[None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, self.IMAGE_CHANNELS],
name='inputs')
self.targets = tf.placeholder(
tf.float32, [None, self.IMAGE_HEIGHT, self.IMAGE_WIDTH, 1],
name='targets')
self.is_training = tf.placeholder_with_default(False, [],
name='is_training')
self.description = ""
self.layers = {}
if per_image_standardization:
list_of_images_norm = tf.map_fn(tf.image.per_image_standardization,
self.inputs)
net = tf.stack(list_of_images_norm)
else:
net = self.inputs
# ENCODER
for layer in layers:
self.layers[layer.name] = net = layer.create_layer(net)
self.description += "{}".format(layer.get_description())
layers.reverse()
Conv2d.reverse_global_variables()
# DECODER
layers_len = len(layers)
for i, layer in enumerate(layers):
if i == (layers_len - 1):
self.segmentation_result = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name], last_layer=True)
else:
net = layer.create_layer_reversed(
net, prev_layer=self.layers[layer.name])
# segmentation_as_classes = tf.reshape(self.y, [50 * self.IMAGE_HEIGHT * self.IMAGE_WIDTH, 1])
# targets_as_classes = tf.reshape(self.targets, [50 * self.IMAGE_HEIGHT * self.IMAGE_WIDTH])
# print(self.y.get_shape())
# self.cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(segmentation_as_classes, targets_as_classes))
print('segmentation_result.shape: {}, targets.shape: {}'.format(
self.segmentation_result.get_shape(), self.targets.get_shape()))
# MSE loss
self.cost = self.batchmsssim(self.segmentation_result, self.targets)
#self.cost = tf.square(self.segmentation_result - self.targets)
self.train_op = tf.train.AdamOptimizer(
learning_rate=0.001).minimize(1 - self.cost)
with tf.name_scope('accuracy'):
# argmax_probs = tf.round(self.segmentation_result) # 0x1
# correct_pred = tf.cast(tf.equal(argmax_probs, self.targets), tf.float32)
correct_pred = tf.square(self.segmentation_result - self.targets)
self.accuracy = tf.reduce_mean(correct_pred)
tf.summary.scalar('accuracy', self.accuracy)
self.summaries = tf.summary.merge_all()
