def build_graph_placeholder():
# Create placeholders (I/O for our model/graph)
x = tf.placeholder(tf.float32, shape=[None, 66, 200, 3], name='IMAGE_IN')
y_ = tf.placeholder(tf.float32, shape=[None, 1], name='LABEL_IN')
dropout_prob = tf.placeholder(tf.float32, name='drop_prob')
x_image = x
# Add input image/steering angle on summary
tf.summary.image("input_image", x_image, 10)
tf.summary.histogram("steer_angle", y_)
# CONV 1 (Mark that want visualization)
conv1 = util.conv2d(x_image, 5, 5, 3, 24, 2, "conv1", True)
# CONV 2
conv2 = util.conv2d(conv1, 5, 5, 24, 36, 2, "conv2")
# CONV 3
conv3 = util.conv2d(conv2, 5, 5, 36, 48, 2, "conv3")
# CONV 4
conv4 = util.conv2d(conv3, 3, 3, 48, 64, 1, "conv4")
# CONV 5
conv5 = util.conv2d(conv4, 3, 3, 64, 64, 1, "conv5")
# Fully Connect 1
# Needs calculation... (-1 means any batch size)
conv5_flat = tf.reshape(conv5, [-1, 1152])
fc1 = util.fc_layer(conv5_flat, 1152, 1164, "fc1")
# Add dropout to the fully connected layer
fc1_drop = tf.nn.dropout(fc1, dropout_prob)
# Fully Connect 2
fc2 = util.fc_layer(fc1_drop, 1164, 100, "fc2")
# Add dropout to the fully connected layer
fc2_drop = tf.nn.dropout(fc2, dropout_prob)
# Fully Connect 3
fc3 = util.fc_layer(fc2_drop, 100, 50, "fc3")
# Add dropout to the fully connected layer
fc3_drop = tf.nn.dropout(fc3, dropout_prob)
# Fully Connect 4
fc4 = util.fc_layer(fc3_drop, 50, 10, "fc4")
# Add dropout to the fully connected layer
fc4_drop = tf.nn.dropout(fc4, dropout_prob)
# Output
y = util.output_layer(fc4_drop, 10, 1, "output_layer")
return x, y, y_, dropout_prob
def build_graph_no_placeholder(input):
dropout_prob = tf.placeholder(tf.float32, name='drop_prob')
# Add input image/steering angle on summary
tf.summary.image("input_image", input, 10)
# CONV 1 (Mark that want visualization)
conv1 = util.conv2d(input, 5, 5, 3, 24, 2, "conv1", True)
# CONV 2
conv2 = util.conv2d(conv1, 5, 5, 24, 36, 2, "conv2")
# CONV 3
conv3 = util.conv2d(conv2, 5, 5, 36, 48, 2, "conv3")
# CONV 4
conv4 = util.conv2d(conv3, 3, 3, 48, 64, 1, "conv4")
# CONV 5
conv5 = util.conv2d(conv4, 3, 3, 64, 64, 1, "conv5")
# Fully Connect 1
# Needs calculation... (-1 means any batch size)
conv5_flat = tf.reshape(conv5, [-1, 1152])
fc1 = util.fc_layer(conv5_flat, 1152, 1164, "fc1")
# Add dropout to the fully connected layer
fc1_drop = tf.nn.dropout(fc1, dropout_prob)
# Fully Connect 2
fc2 = util.fc_layer(fc1_drop, 1164, 100, "fc2")
# Add dropout to the fully connected layer
fc2_drop = tf.nn.dropout(fc2, dropout_prob)
# Fully Connect 3
fc3 = util.fc_layer(fc2_drop, 100, 50, "fc3")
# Add dropout to the fully connected layer
fc3_drop = tf.nn.dropout(fc3, dropout_prob)
# Fully Connect 4
fc4 = util.fc_layer(fc3_drop, 50, 10, "fc4")
# Add dropout to the fully connected layer
fc4_drop = tf.nn.dropout(fc4, dropout_prob)
# Output
y = util.output_layer(fc4_drop, 10, 1, "output_layer")
return y, dropout_prob
def __init__(self, input=None, use_placeholder=True, training_mode=True, img_size=224, num_classes=2):
self.__x = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='IMAGE_IN')
self.__size_input_x = img_size
self.__size_input_y = img_size
self.__label = tf.placeholder(tf.int32, shape=[None, img_size, img_size, 1], name='LABEL_IN')
self.__use_placeholder = use_placeholder
##### ENCODER
# Calculating the convolution output:
# https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/convolutional_neural_networks.html
# H_out = 1 + (H_in+(2*pad)-K)/S
# W_out = 1 + (W_in+(2*pad)-K)/S
# CONV1: Input 100x100x3 after CONV 3x3 P:0 S:1 H_out: 1 + (100-3)/1 = 98, W_out= 1 + (100-3)/1 = 98
if use_placeholder:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(self.__x, 3, 3, 3, 16, 1, "conv1", viewWeights=True, do_summary=False)
else:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(input, 3, 3, 3, 16, 1, "conv1", viewWeights=True, do_summary=False)
self.__conv1_bn = util.batch_norm(self.__conv1, training_mode, name='bn_c1')
self.__conv1_act = util.relu(self.__conv1_bn, do_summary=False)
# CONV2: Input 98x98x16 after CONV 3x3 P:0 S:2 H_out: 1 + (98-3)/2 = 48, W_out= 1 + (98-3)/2 = 48
self.__conv2 = util.conv2d(self.__conv1_act, 3, 3, 16, 16, 1, "conv2", do_summary=False)
self.__conv2_bn = util.batch_norm(self.__conv2, training_mode, name='bn_c2')
self.__conv2_act = util.relu(self.__conv2_bn, do_summary=False)
# Add Maxpool
self.__conv2_mp_act = util.max_pool(self.__conv2_act, 2,2,2,name="maxpool1")
# CONV3: Input 48x48x16 after CONV 3x3 P:0 S:1 H_out: 1 + (48-3)/1 = 46, W_out= 1 + (48-3)/1 = 46
self.__conv3 = util.conv2d(self.__conv2_mp_act, 3, 3, 16, 32, 1, "conv3", do_summary=False)
self.__conv3_bn = util.batch_norm(self.__conv3, training_mode, name='bn_c3')
self.__conv3_act = util.relu(self.__conv3_bn, do_summary=False)
# CONV4: Input 46x46x32 after CONV 3x3 P:0 S:2 H_out: 1 + (46-3)/2 = 22, W_out= 1 + (46-3)/2 = 22
self.__conv4 = util.conv2d(self.__conv3_act, 3, 3, 32, 64, 1, "conv4", do_summary=False)
self.__conv4_bn = util.batch_norm(self.__conv4, training_mode, name='bn_c4')
self.__conv4_act = util.relu(self.__conv4_bn, do_summary=False)
# Add Maxpool
self.__conv4_mp_act = util.max_pool(self.__conv4_act, 2, 2, 2, name="maxpool2")
# CONV5: Input 22x22x64 after CONV 3x3 P:0 S:1 H_out: 1 + (22-3)/1 = 20, W_out= 1 + (22-3)/1 = 20
self.__conv5 = util.conv2d(self.__conv4_mp_act, 3, 3, 64, 64, 1, "conv5", do_summary=False)
self.__conv5_bn = util.batch_norm(self.__conv5, training_mode, name='bn_c5')
self.__conv5_act = util.relu(self.__conv5_bn, do_summary=False)
##### DECODER (At this point we have 1x18x64
# Kernel, output size, in_volume, out_volume, stride
self.__conv_t5_out = util.conv2d_transpose(self.__conv5_act, (3, 3), (22, 22), 64, 64, 1, name="dconv1",
do_summary=False)
self.__conv_t5_out_bn = util.batch_norm(self.__conv_t5_out, training_mode, name='bn_t_c5')
self.__conv_t5_out_act = util.relu(self.__conv_t5_out_bn, do_summary=False)
self.__conv_t4_out = util.conv2d_transpose(
self.__conv_t5_out_act,
(3, 3), (46, 46), 64, 32, 2, name="dconv2", do_summary=False)
self.__conv_t4_out_bn = util.batch_norm(self.__conv_t4_out, training_mode, name='bn_t_c4')
self.__conv_t4_out_act = util.relu(self.__conv_t4_out_bn, do_summary=False)
self.__conv_t3_out = util.conv2d_transpose(
self.__conv_t4_out_act,
(3, 3), (48, 48), 32, 16, 1, name="dconv3", do_summary=False)
self.__conv_t3_out_bn = util.batch_norm(self.__conv_t3_out, training_mode, name='bn_t_c3')
self.__conv_t3_out_act = util.relu(self.__conv_t3_out_bn, do_summary=False)
self.__conv_t2_out = util.conv2d_transpose(
self.__conv_t3_out_act,
(3, 3), (98, 98), 16, 16, 2, name="dconv4", do_summary=False)
self.__conv_t2_out_bn = util.batch_norm(self.__conv_t2_out, training_mode, name='bn_t_c2')
self.__conv_t2_out_act = util.relu(self.__conv_t2_out_bn, do_summary=False)
# Observe that the last deconv depth is the same as the number of classes
self.__conv_t1_out = util.conv2d_transpose(
self.__conv_t2_out_act,
(3, 3), (img_size, img_size), 16, num_classes, 1, name="dconv5", do_summary=False)
self.__conv_t1_out_bn = util.batch_norm(self.__conv_t1_out, training_mode, name='bn_t_c1')
# Model output (It's not the segmentation yet...)
self.__y = self.__conv_t1_out_bn
with tf.name_scope('anotation_pred'):
# Create the predicted annotation
# Now we filter on the third dimension the the strogest pixels from a particular class
# Returns the index with the largest value across axes of a tensor, on our case the index will be one of the
# classes represented by the depth of our output (150 classes)
self.__anotation = tf.argmax(self.__conv_t1_out_bn, dimension=3, name="prediction")
def __init__(self, input=None, use_placeholder=True, training_mode=True, img_size = 224):
self.__x = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='IMAGE_IN')
self.__label = tf.placeholder(tf.int32, shape=[None, img_size, img_size, 1], name='LABEL_IN')
self.__use_placeholder = use_placeholder
self.__size_input_x = img_size
self.__size_input_y = img_size
##### ENCODER
# Calculating the convolution output:
# https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/convolutional_neural_networks.html
# H_out = 1 + (H_in+(2*pad)-K)/S
# W_out = 1 + (W_in+(2*pad)-K)/S
# CONV1: Input 224x224x3 after CONV 5x5 P:0 S:2 H_out: 1 + (224-5)/2 = 110, W_out= 1 + (224-5)/2 = 110
if use_placeholder:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(self.__x, 5, 5, 3, 32, 2, "conv1", viewWeights=True, do_summary=False)
else:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(input, 5, 5, 3, 32, 2, "conv1", viewWeights=True, do_summary=False)
self.__conv1_bn = util.batch_norm(self.__conv1, training_mode, name='bn_c1')
self.__conv1_act = util.relu(self.__conv1_bn, do_summary=False)
# CONV2: Input 110x110x24 after CONV 5x5 P:0 S:2 H_out: 1 + (110-5)/2 = 53, W_out= 1 + (110-5)/2 = 53
self.__conv2 = util.conv2d(self.__conv1_act, 5, 5, 32, 64, 2, "conv2", do_summary=False)
self.__conv2_bn = util.batch_norm(self.__conv2, training_mode, name='bn_c2')
self.__conv2_act = util.relu(self.__conv2_bn, do_summary=False)
# CONV3: Input 53x53x36 after CONV 5x5 P:0 S:2 H_out: 1 + (53-5)/2 = 25, W_out= 1 + (53-5)/2 = 25
self.__conv3 = util.conv2d(self.__conv2_act, 5, 5, 64, 64, 2, "conv3", do_summary=False)
self.__conv3_bn = util.batch_norm(self.__conv3, training_mode, name='bn_c3')
self.__conv3_act = util.relu(self.__conv3_bn, do_summary=False)
# CONV4: Input 25x25x48 after CONV 3x3 P:0 S:1 H_out: 1 + (25-3)/1 = 23, W_out= 1 + (25-3)/1 = 23
self.__conv4 = util.conv2d(self.__conv3_act, 3, 3, 64, 64, 1, "conv4", do_summary=False)
self.__conv4_bn = util.batch_norm(self.__conv4, training_mode, name='bn_c4')
self.__conv4_act = util.relu(self.__conv4_bn, do_summary=False)
# CONV5: Input 23x23x64 after CONV 3x3 P:0 S:1 H_out: 1 + (23-3)/1 = 21, W_out= 1 + (23-3)/1 = 21
self.__conv5 = util.conv2d(self.__conv4_act, 3, 3, 64, 32, 1, "conv5", do_summary=False)
self.__conv5_bn = util.batch_norm(self.__conv5, training_mode, name='bn_c5')
self.__conv5_act = util.relu(self.__conv5_bn, do_summary=False)
# CONV6: Input 21x21x32 after CONV 3x3 P:0 S:1 H_out: 1 + (21-3)/1 = 19, W_out= 1 + (21-3)/1 = 19
self.__conv6 = util.conv2d(self.__conv5_act, 3, 3, 32, 32, 1, "conv6", do_summary=False)
self.__conv6_bn = util.batch_norm(self.__conv6, training_mode, name='bn_c6')
self.__conv6_act = util.relu(self.__conv6_bn, do_summary=False)
# CONV7: Input 19x19x64 after CONV 3x3 P:0 S:1 H_out: 1 + (19-3)/1 = 17, W_out= 1 + (19-3)/1 = 17
self.__conv7 = util.conv2d(self.__conv6_act, 3, 3, 32, 32, 2, "conv7", do_summary=False)
self.__conv7_bn = util.batch_norm(self.__conv7, training_mode, name='bn_c6')
self.__conv7_act = util.relu(self.__conv7_bn, do_summary=False)
##### DECODER (At this point we have 1x18x64
# Kernel, output size, in_volume, out_volume, stride
self.__conv_t7_out = util.conv2d_transpose(self.__conv7_act, (3, 3), (19, 19), 32, 32, 2, name="dconv1",
do_summary=False)
self.__conv_t7_out_bn = util.batch_norm(self.__conv_t7_out, training_mode, name='bn_t_c7')
self.__conv_t7_out_act = util.relu(self.__conv_t7_out_bn, do_summary=False)
self.__conv_t6_out = util.conv2d_transpose(self.__conv_t7_out_act, (3, 3), (21, 21), 32, 32, 1, name="dconv2",
do_summary=False)
self.__conv_t6_out_bn = util.batch_norm(self.__conv_t6_out, training_mode, name='bn_t_c6')
self.__conv_t6_out_act = util.relu(self.__conv_t6_out, do_summary=False)
self.__conv_t5_out = util.conv2d_transpose(self.__conv_t6_out_act, (3, 3), (23, 23), 32, 64, 1, name="dconv3",
do_summary=False)
self.__conv_t5_out_bn = util.batch_norm(self.__conv_t5_out, training_mode, name='bn_t_c5')
self.__conv_t5_out_act = util.relu(self.__conv_t5_out_bn, do_summary=False)
self.__conv_t4_out = util.conv2d_transpose(
self.__conv_t5_out_act,
(3, 3), (25, 25), 64, 64, 1, name="dconv4",do_summary=False)
self.__conv_t4_out_bn = util.batch_norm(self.__conv_t4_out, training_mode, name='bn_t_c4')
self.__conv_t4_out_act = util.relu(self.__conv_t4_out_bn, do_summary=False)
self.__conv_t3_out = util.conv2d_transpose(
self.__conv_t4_out_act,
(5, 5), (53, 53), 64, 64, 2, name="dconv5",do_summary=False)
self.__conv_t3_out_bn = util.batch_norm(self.__conv_t3_out, training_mode, name='bn_t_c3')
self.__conv_t3_out_act = util.relu(self.__conv_t3_out_bn, do_summary=False)
self.__conv_t2_out = util.conv2d_transpose(
self.__conv_t3_out_act,
(5, 5), (110, 110), 64, 32, 2, name="dconv6",do_summary=False)
self.__conv_t2_out_bn = util.batch_norm(self.__conv_t2_out, training_mode, name='bn_t_c2')
self.__conv_t2_out_act = util.relu(self.__conv_t2_out_bn, do_summary=False)
# Observe that the last deconv depth is the same as the number of classes
self.__conv_t1_out = util.conv2d_transpose(
self.__conv_t2_out_act,
(5, 5), (img_size, img_size), 32, 3, 2, name="dconv7",do_summary=False)
self.__conv_t1_out_bn = util.batch_norm(self.__conv_t1_out, training_mode, name='bn_t_c1')
# Model output (It's not the segmentation yet...)
self.__y = util.relu(self.__conv_t1_out_bn, do_summary = False)
def __init__(self, input=None, use_placeholder=True, training_mode=True, img_size = 224, num_classes=2):
self.__x = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 3], name='IMAGE_IN')
self.__label = tf.placeholder(tf.int32, shape=[None, img_size, img_size, 1], name='LABEL_IN')
self.__use_placeholder = use_placeholder
self.__size_input_x = img_size
self.__size_input_y = img_size
##### ENCODER
# Calculating the convolution output:
# https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/convolutional_neural_networks.html
# H_out = 1 + (H_in+(2*pad)-K)/S
# W_out = 1 + (W_in+(2*pad)-K)/S
# CONV1: Input 224x224x3 after CONV 5x5 P:0 S:2 H_out: 1 + (224-5)/2 = 110, W_out= 1 + (224-5)/2 = 110
if use_placeholder:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(self.__x, 5, 5, 3, 32, 2, "conv1", viewWeights=True, do_summary=False)
else:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(input, 5, 5, 3, 32, 2, "conv1", viewWeights=True, do_summary=False)
self.__conv1_bn = util.batch_norm(self.__conv1, training_mode, name='bn_c1')
self.__conv1_act = util.relu(self.__conv1_bn, do_summary=False)
# CONV2: Input 110x110x24 after CONV 5x5 P:0 S:2 H_out: 1 + (110-5)/2 = 53, W_out= 1 + (110-5)/2 = 53
self.__conv2 = util.conv2d(self.__conv1_act, 5, 5, 32, 64, 2, "conv2", do_summary=False)
self.__conv2_bn = util.batch_norm(self.__conv2, training_mode, name='bn_c2')
self.__conv2_act = util.relu(self.__conv2_bn, do_summary=False)
# CONV3: Input 53x53x36 after CONV 5x5 P:0 S:2 H_out: 1 + (53-5)/2 = 25, W_out= 1 + (53-5)/2 = 25
self.__conv3 = util.conv2d(self.__conv2_act, 5, 5, 64, 64, 2, "conv3", do_summary=False)
self.__conv3_bn = util.batch_norm(self.__conv3, training_mode, name='bn_c3')
self.__conv3_act = util.relu(self.__conv3_bn, do_summary=False)
# CONV4: Input 25x25x48 after CONV 3x3 P:0 S:1 H_out: 1 + (25-3)/1 = 23, W_out= 1 + (25-3)/1 = 23
self.__conv4 = util.conv2d(self.__conv3_act, 3, 3, 64, 64, 1, "conv4", do_summary=False)
self.__conv4_bn = util.batch_norm(self.__conv4, training_mode, name='bn_c4')
self.__conv4_act = util.relu(self.__conv4_bn, do_summary=False)
# CONV5: Input 23x23x64 after CONV 3x3 P:0 S:1 H_out: 1 + (23-3)/1 = 21, W_out= 1 + (23-3)/1 = 21
self.__conv5 = util.conv2d(self.__conv4_act, 3, 3, 64, 64, 1, "conv5", do_summary=False)
self.__conv5_bn = util.batch_norm(self.__conv5, training_mode, name='bn_c5')
self.__conv5_act = util.relu(self.__conv5_bn, do_summary=False)
##### DECODER (At this point we have 1x18x64
# Kernel, output size, in_volume, out_volume, stride
self.__conv_t5_out = util.conv2d_transpose(self.__conv5_act, (3, 3), (23, 23), 64, 64, 1, name="dconv1",
do_summary=False)
self.__conv_t5_out_bn = util.batch_norm(self.__conv_t5_out, training_mode, name='bn_t_c5')
self.__conv_t5_out_act = util.relu(self.__conv_t5_out_bn, do_summary=False)
self.__conv_t4_out = util.conv2d_transpose(
self.__conv_t5_out_act+util.gate_tensor(self.__conv4_act, name='gate4'),
(3, 3), (25, 25), 64, 64, 1, name="dconv2",do_summary=False)
self.__conv_t4_out_bn = util.batch_norm(self.__conv_t4_out, training_mode, name='bn_t_c4')
self.__conv_t4_out_act = util.relu(self.__conv_t4_out_bn, do_summary=False)
self.__conv_t3_out = util.conv2d_transpose(
self.__conv_t4_out_act+util.gate_tensor(self.__conv3_act, name='gate3'),
(5, 5), (53, 53), 64, 64, 2, name="dconv3",do_summary=False)
self.__conv_t3_out_bn = util.batch_norm(self.__conv_t3_out, training_mode, name='bn_t_c3')
self.__conv_t3_out_act = util.relu(self.__conv_t3_out_bn, do_summary=False)
self.__conv_t2_out = util.conv2d_transpose(
self.__conv_t3_out_act+util.gate_tensor(self.__conv2_act, name='gate2'),
(5, 5), (110, 110), 64, 32, 2, name="dconv4",do_summary=False)
self.__conv_t2_out_bn = util.batch_norm(self.__conv_t2_out, training_mode, name='bn_t_c2')
self.__conv_t2_out_act = util.relu(self.__conv_t2_out_bn, do_summary=False)
# Observe that the last deconv depth is the same as the number of classes
self.__conv_t1_out = util.conv2d_transpose(
self.__conv_t2_out_act+util.gate_tensor(self.__conv1_act, name='gate1'),
(5, 5), (img_size, img_size), 32, num_classes, 2, name="dconv5",do_summary=False)
self.__conv_t1_out_bn = util.batch_norm(self.__conv_t1_out, training_mode, name='bn_t_c1')
# Model output (It's not the segmentation yet...)
self.__y = self.__conv_t1_out_bn
with tf.name_scope('anotation_pred'):
# Create the predicted annotation
# Now we filter on the third dimension the the strogest pixels from a particular class
# Returns the index with the largest value across axes of a tensor, on our case the index will be one of the
# classes represented by the depth of our output (150 classes)
self.__anotation = tf.argmax(self.__conv_t1_out_bn, dimension=3, name="prediction")
def __init__(self,
input=None,
use_placeholder=True,
training_mode=True,
img_size=100,
multiplier=1):
self.__x = tf.placeholder(tf.float32,
shape=[None, img_size, img_size, 3],
name='IMAGE_IN')
self.__label = tf.placeholder(tf.int32,
shape=[None, img_size, img_size, 1],
name='LABEL_IN')
self.__use_placeholder = use_placeholder
##### ENCODER
# Calculating the convolution output:
# https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/convolutional_neural_networks.html
# H_out = 1 + (H_in+(2*pad)-K)/S
# W_out = 1 + (W_in+(2*pad)-K)/S
# CONV1: Input 100x100x3 after CONV 3x3 P:0 S:1 H_out: 1 + (100-3)/1 = 98, W_out= 1 + (100-3)/1 = 98
if use_placeholder:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(self.__x,
3,
3,
3,
16,
1,
"conv1",
viewWeights=True,
do_summary=False)
else:
# CONV 1 (Mark that want visualization)
self.__conv1 = util.conv2d(input,
3,
3,
3,
16,
1,
"conv1",
viewWeights=True,
do_summary=False)
self.__conv1_bn = util.batch_norm(self.__conv1,
training_mode,
name='bn_c1')
self.__conv1_act = util.relu(self.__conv1_bn, do_summary=False)
# CONV2: Input 98x98x16 after CONV 3x3 P:0 S:2 H_out: 1 + (98-3)/2 = 48, W_out= 1 + (98-3)/2 = 48
#self.__conv2 = util.conv2d(self.__conv1_act, 3, 3, 16, 16, 1, "conv2", do_summary=False)
#self.__conv2_bn = util.batch_norm(self.__conv2, training_mode, name='bn_c2')
#self.__conv2_act = util.relu(self.__conv2_bn, do_summary=False)
self.__conv2_act = util.conv2d_separable(self.__conv1_act,
3,
3,
16,
16,
2,
training_mode,
"conv2",
do_summary=False,
multiplier=multiplier)
# Add Maxpool
#self.__conv2_mp_act = util.max_pool(self.__conv2_act, 2,2,2,name="maxpool1")
# CONV3: Input 48x48x16 after CONV 3x3 P:0 S:1 H_out: 1 + (48-3)/1 = 46, W_out= 1 + (48-3)/1 = 46
#self.__conv3 = util.conv2d(self.__conv2_mp_act, 3, 3, 16, 32, 1, "conv3", do_summary=False)
#self.__conv3_bn = util.batch_norm(self.__conv3, training_mode, name='bn_c3')
#self.__conv3_act = util.relu(self.__conv3_bn, do_summary=False)
self.__conv3_act = util.conv2d_separable(self.__conv2_act,
3,
3,
16,
32,
1,
training_mode,
"conv3",
do_summary=False,
multiplier=multiplier)
# CONV4: Input 46x46x32 after CONV 3x3 P:0 S:2 H_out: 1 + (46-3)/2 = 22, W_out= 1 + (46-3)/2 = 22
#self.__conv4 = util.conv2d(self.__conv3_act, 3, 3, 32, 64, 1, "conv4", do_summary=False)
#self.__conv4_bn = util.batch_norm(self.__conv4, training_mode, name='bn_c4')
#self.__conv4_act = util.relu(self.__conv4_bn, do_summary=False)
self.__conv4_act = util.conv2d_separable(self.__conv3_act,
3,
3,
32,
64,
2,
training_mode,
"conv4",
do_summary=False,
multiplier=multiplier)
# Add Maxpool
#self.__conv4_mp_act = util.max_pool(self.__conv4_act, 2, 2, 2, name="maxpool2")
# CONV5: Input 22x22x64 after CONV 3x3 P:0 S:1 H_out: 1 + (22-3)/1 = 20, W_out= 1 + (22-3)/1 = 20
#self.__conv5 = util.conv2d(self.__conv4_mp_act, 3, 3, 64, 64, 1, "conv5", do_summary=False)
#self.__conv5_bn = util.batch_norm(self.__conv5, training_mode, name='bn_c5')
#self.__conv5_act = util.relu(self.__conv5_bn, do_summary=False)
self.__conv5_act = util.conv2d_separable(self.__conv4_act,
3,
3,
64,
64,
1,
training_mode,
"conv5",
do_summary=False,
multiplier=multiplier)
##### DECODER (At this point we have 1x18x64
# Kernel, output size, in_volume, out_volume, stride
self.__conv_t5_out = util.conv2d_transpose(self.__conv5_act, (3, 3),
(22, 22),
64,
64,
1,
name="dconv1",
do_summary=False)
self.__conv_t5_out_bn = util.batch_norm(self.__conv_t5_out,
training_mode,
name='bn_t_c5')
self.__conv_t5_out_act = util.relu(self.__conv_t5_out_bn,
do_summary=False)
self.__conv_t4_out = util.conv2d_transpose(self.__conv_t5_out_act,
(3, 3), (46, 46),
64,
32,
2,
name="dconv2",
do_summary=False)
self.__conv_t4_out_bn = util.batch_norm(self.__conv_t4_out,
training_mode,
name='bn_t_c4')
self.__conv_t4_out_act = util.relu(self.__conv_t4_out_bn,
do_summary=False)
self.__conv_t3_out = util.conv2d_transpose(self.__conv_t4_out_act,
(3, 3), (48, 48),
32,
16,
1,
name="dconv3",
do_summary=False)
self.__conv_t3_out_bn = util.batch_norm(self.__conv_t3_out,
training_mode,
name='bn_t_c3')
self.__conv_t3_out_act = util.relu(self.__conv_t3_out_bn,
do_summary=False)
self.__conv_t2_out = util.conv2d_transpose(self.__conv_t3_out_act,
(3, 3), (98, 98),
16,
16,
2,
name="dconv4",
do_summary=False)
self.__conv_t2_out_bn = util.batch_norm(self.__conv_t2_out,
training_mode,
name='bn_t_c2')
self.__conv_t2_out_act = util.relu(self.__conv_t2_out_bn,
do_summary=False)
# Observe that the last deconv depth is the same as the number of classes
self.__conv_t1_out = util.conv2d_transpose(self.__conv_t2_out_act,
(3, 3),
(img_size, img_size),
16,
3,
1,
name="dconv5",
do_summary=False)
self.__conv_t1_out_bn = util.batch_norm(self.__conv_t1_out,
training_mode,
name='bn_t_c1')
# Model output (It's not the segmentation yet...)
self.__y = util.relu(self.__conv_t1_out_bn, do_summary=False)
# Calculate flat tensor for Binary Cross entropy loss
self.__y_flat = tf.reshape(
self.__y, [tf.shape(self.__x)[0], img_size * img_size * 3])
self.__x_flat = tf.reshape(
self.__x, [tf.shape(self.__x)[0], img_size * img_size * 3])