def __init__(self, conf_file="config.json"):
with open(conf_file) as f:
self.config = json.load(f)
self.num_classes = self.config["NUM_CLASSES"]
self.use_vgg = self.config["USE_VGG"]
if self.use_vgg is False:
self.vgg_param_dict = None
print("No VGG path in config, so learning from scratch")
else:
self.vgg16_npy_path = self.config["VGG_FILE"]
self.vgg_param_dict = np.load(self.vgg16_npy_path,
encoding='latin1').item()
print("VGG parameter loaded")
self.train_file = self.config["TRAIN_FILE"]
self.val_file = self.config["VAL_FILE"]
self.test_file = self.config["TEST_FILE"]
self.img_prefix = self.config["IMG_PREFIX"]
self.label_prefix = self.config["LABEL_PREFIX"]
self.opt = self.config["OPT"]
self.input_w = self.config["INPUT_WIDTH"]
self.input_h = self.config["INPUT_HEIGHT"]
self.input_c = self.config["INPUT_CHANNELS"]
self.tb_logs = self.config["TB_LOGS"]
self.saved_dir = self.config["SAVE_MODEL_DIR"]
self.images_tr, self.labels_tr = None, None
self.images_val, self.labels_val = None, None
self.train_loss, self.train_accuracy = [], []
self.val_loss, self.val_acc = [], []
self.model_version = 0 # used for saving the model
self.saver = None
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.Session()
self.batch_size_pl = tf.placeholder(tf.int64,
shape=[],
name="batch_size")
#self.is_training_pl = tf.placeholder(tf.bool, name="is_training")
self.inputs_pl = tf.placeholder(
tf.float32, [None, self.input_h, self.input_w, self.input_c])
self.labels_pl = tf.placeholder(
tf.int64, [None, self.input_h, self.input_w, 1])
# perform local response normalization
self.norm1 = tf.nn.lrn(self.inputs_pl,
depth_radius=5,
bias=1.0,
alpha=0.0001,
beta=0.75,
name='norm1')
# first box of convolution layer,each part we do convolution two times, so we have conv1_1, and conv1_2
self.conv1_1 = conv_layer(self.norm1, "conv1_1", [3, 3, 3, 64],
self.use_vgg, self.vgg_param_dict)
self.conv1_2 = conv_layer(self.conv1_1, "conv1_2", [3, 3, 64, 64],
self.use_vgg, self.vgg_param_dict)
self.pool1, self.pool1_index, self.shape_1 = max_pool(
self.conv1_2, 'pool1')
# Second box of convolution layer(4)
self.conv2_1 = conv_layer(self.pool1, "conv2_1", [3, 3, 64, 128],
self.use_vgg, self.vgg_param_dict)
self.conv2_2 = conv_layer(self.conv2_1, "conv2_2",
[3, 3, 128, 128], self.use_vgg,
self.vgg_param_dict)
self.pool2, self.pool2_index, self.shape_2 = max_pool(
self.conv2_2, 'pool2')
# Third box of convolution layer(7)
self.conv3_1 = conv_layer(self.pool2, "conv3_1", [3, 3, 128, 256],
self.use_vgg, self.vgg_param_dict)
self.conv3_2 = conv_layer(self.conv3_1, "conv3_2",
[3, 3, 256, 256], self.use_vgg,
self.vgg_param_dict)
self.conv3_3 = conv_layer(self.conv3_2, "conv3_3",
[3, 3, 256, 256], self.use_vgg,
self.vgg_param_dict)
self.pool3, self.pool3_index, self.shape_3 = max_pool(
self.conv3_3, 'pool3')
# Fourth box of convolution layer(10)
self.conv4_1 = conv_layer(self.pool3, "conv4_1", [3, 3, 256, 512],
self.use_vgg, self.vgg_param_dict)
self.conv4_2 = conv_layer(self.conv4_1, "conv4_2",
[3, 3, 512, 512], self.use_vgg,
self.vgg_param_dict)
self.conv4_3 = conv_layer(self.conv4_2, "conv4_3",
[3, 3, 512, 512], self.use_vgg,
self.vgg_param_dict)
self.pool4, self.pool4_index, self.shape_4 = max_pool(
self.conv4_3, 'pool4')
# Fifth box of convolution layers(13)
self.conv5_1 = conv_layer(self.pool4, "conv5_1", [3, 3, 512, 512],
self.use_vgg, self.vgg_param_dict)
self.conv5_2 = conv_layer(self.conv5_1, "conv5_2",
[3, 3, 512, 512], self.use_vgg,
self.vgg_param_dict)
self.conv5_3 = conv_layer(self.conv5_2, "conv5_3",
[3, 3, 512, 512], self.use_vgg,
self.vgg_param_dict)
self.pool5, self.pool5_index, self.shape_5 = max_pool(
self.conv5_3, 'pool5')
# ---------------------So Now the encoder process has been Finished--------------------------------------#
# ------------------Then Let's start Decoder Process-----------------------------------------------------#
# First box of deconvolution layers(3)
self.deconv5_1 = up_sampling(self.pool5,
self.pool5_index,
self.shape_5,
self.batch_size_pl,
name="unpool_5")
self.deconv5_2 = conv_layer(self.deconv5_1, "deconv5_2",
[3, 3, 512, 512])
self.deconv5_3 = conv_layer(self.deconv5_2, "deconv5_3",
[3, 3, 512, 512])
self.deconv5_4 = conv_layer(self.deconv5_3, "deconv5_4",
[3, 3, 512, 512])
# Second box of deconvolution layers(6)
self.deconv4_1 = up_sampling(self.deconv5_4,
self.pool4_index,
self.shape_4,
self.batch_size_pl,
name="unpool_4")
self.deconv4_2 = conv_layer(self.deconv4_1, "deconv4_2",
[3, 3, 512, 512])
self.deconv4_3 = conv_layer(self.deconv4_2, "deconv4_3",
[3, 3, 512, 512])
self.deconv4_4 = conv_layer(self.deconv4_3, "deconv4_4",
[3, 3, 512, 256])
# Third box of deconvolution layers(9)
self.deconv3_1 = up_sampling(self.deconv4_4,
self.pool3_index,
self.shape_3,
self.batch_size_pl,
name="unpool_3")
self.deconv3_2 = conv_layer(self.deconv3_1, "deconv3_2",
[3, 3, 256, 256])
self.deconv3_3 = conv_layer(self.deconv3_2, "deconv3_3",
[3, 3, 256, 256])
self.deconv3_4 = conv_layer(self.deconv3_3, "deconv3_4",
[3, 3, 256, 128])
# Fourth box of deconvolution layers(11)
self.deconv2_1 = up_sampling(self.deconv3_4,
self.pool2_index,
self.shape_2,
self.batch_size_pl,
name="unpool_2")
self.deconv2_2 = conv_layer(self.deconv2_1, "deconv2_2",
[3, 3, 128, 128])
self.deconv2_3 = conv_layer(self.deconv2_2, "deconv2_3",
[3, 3, 128, 64])
# Fifth box of deconvolution layers(13)
self.deconv1_1 = up_sampling(self.deconv2_3,
self.pool1_index,
self.shape_1,
self.batch_size_pl,
name="unpool_1")
self.deconv1_2 = conv_layer(self.deconv1_1, "deconv1_2",
[3, 3, 64, 64])
self.deconv1_3 = conv_layer(self.deconv1_2, "deconv1_3",
[3, 3, 64, 64])
with tf.variable_scope('conv_classifier') as scope:
self.kernel = tf.get_variable(
'weights',
shape=[1, 1, 64, self.num_classes],
initializer=initialization(1, 64))
self.conv = tf.nn.conv2d(self.deconv1_3,
self.kernel, [1, 1, 1, 1],
padding='SAME')
self.biases = tf.get_variable(
'biases',
shape=[self.num_classes],
initializer=tf.constant_initializer(0.0))
self.logits = tf.nn.bias_add(self.conv,
self.biases,
name=scope.name)
def visual_results(self,
dataset_type="TEST",
images_index=3,
FLAG_MAX_VOTE=False):
image_w = self.config["INPUT_WIDTH"]
image_h = self.config["INPUT_HEIGHT"]
image_c = self.config["INPUT_CHANNELS"]
train_dir = self.config["SAVE_MODEL_DIR"]
FLAG_BAYES = self.config["BAYES"]
with self.sess as sess:
# Restore saved session
saver = tf.train.Saver()
saver.restore(sess, train_dir)
kernel = variable_with_weight_decay('weights',
initializer=initialization(
1, 64),
shape=[1, 1, 64, 3],
wd=False)
conv = tf.nn.conv2d(self.deconv1_3,
kernel, [1, 1, 1, 1],
padding='SAME')
biases = variable_with_weight_decay('biases',
tf.constant_initializer(0.0),
shape=[3],
wd=False)
logits = tf.nn.bias_add(conv, biases, name="scope.name")
#exit()
sess.run(tf.global_variables_initializer())
#sess.run(logits)
_, _, prediction = cal_loss(logits=logits, labels=self.labels_pl)
prob = tf.nn.softmax(logits, dim=-1)
print(
"==================================================================================="
)
print(prediction)
#exit()
if (dataset_type == 'TRAIN'):
test_type_path = self.config["TRAIN_FILE"]
if type(images_index) == list:
indexes = images_index
else:
indexes = random.sample(range(367), images_index)
#indexes = [0,75,150,225,300]
elif (dataset_type == 'VAL'):
test_type_path = self.config["VAL_FILE"]
if type(images_index) == list:
indexes = images_index
else:
indexes = random.sample(range(101), images_index)
#indexes = [0,25,50,75,100]
elif (dataset_type == 'TEST'):
test_type_path = self.config["TEST_FILE"]
if type(images_index) == list:
indexes = images_index
else:
indexes = random.sample(range(233), images_index)
#indexes = [0,50,100,150,200]
# Load images
image_filename, label_filename = get_filename_list(
test_type_path, self.config)
images, labels = get_all_test_data(image_filename, label_filename)
# Keep images subset of length images_index
images = [images[i] for i in indexes]
labels = [labels[i] for i in indexes]
num_sample_generate = 30
pred_tot = []
var_tot = []
for image_batch, label_batch in zip(images, labels):
image_batch = np.reshape(image_batch,
[1, image_h, image_w, image_c])
label_batch = np.reshape(label_batch, [1, image_h, image_w, 1])
if FLAG_BAYES is False:
fetches = [prediction]
feed_dict = {
self.inputs_pl: image_batch,
self.labels_pl: label_batch,
self.is_training_pl: False,
self.keep_prob_pl: 0.5,
self.batch_size_pl: 1
}
pred = sess.run(fetches=fetches, feed_dict=feed_dict)
pred = np.reshape(pred, [image_h, image_w])
var_one = []
else:
feed_dict = {
self.inputs_pl: image_batch,
self.labels_pl: label_batch,
self.is_training_pl: False,
self.keep_prob_pl: 0.5,
self.with_dropout_pl: True,
self.batch_size_pl: 1
}
prob_iter_tot = []
pred_iter_tot = []
for iter_step in range(num_sample_generate):
prob_iter_step = sess.run(fetches=[prob],
feed_dict=feed_dict)
prob_iter_tot.append(prob_iter_step)
pred_iter_tot.append(
np.reshape(np.argmax(prob_iter_step, axis=-1),
[-1]))
if FLAG_MAX_VOTE is True:
prob_variance, pred = MAX_VOTE(
pred_iter_tot, prob_iter_tot,
self.config["NUM_CLASSES"])
#acc_per = np.mean(np.equal(pred,np.reshape(label_batch,[-1])))
var_one = var_calculate(pred, prob_variance)
pred = np.reshape(pred, [image_h, image_w])
else:
prob_mean = np.nanmean(prob_iter_tot, axis=0)
prob_variance = np.var(prob_iter_tot, axis=0)
pred = np.reshape(
np.argmax(prob_mean, axis=-1), [-1]
) #pred is the predicted label with the mean of generated samples
#THIS TIME I DIDN'T INCLUDE TAU
var_one = var_calculate(pred, prob_variance)
pred = np.reshape(pred, [image_h, image_w])
pred_tot.append(pred)
var_tot.append(var_one)
draw_plots_bayes(images, labels, pred_tot, var_tot)
def __init__(self, conf_file="config2.json"):
with open(conf_file) as f:
self.config = json.load(f)
self.num_classes = self.config["NUM_CLASSES"]
self.use_vgg = self.config["USE_VGG"]
if self.use_vgg is False:
self.vgg_param_dict = None
print("No VGG path in config, so learning from scratch")
else:
self.vgg16_npy_path = self.config["VGG_FILE"]
self.vgg_param_dict = np.load(self.vgg16_npy_path, encoding='latin1').item()
print("VGG parameter loaded")
self.train_file = self.config["TRAIN_FILE"]
self.val_file = self.config["VAL_FILE"]
self.test_file = self.config["TEST_FILE"]
self.img_prefix = self.config["IMG_PREFIX"]
self.label_prefix = self.config["LABEL_PREFIX"]
self.bayes = self.config["BAYES"]
self.opt = self.config["OPT"]
self.saved_dir = self.config["SAVE_MODEL_DIR"]
self.input_w = self.config["INPUT_WIDTH"]
self.input_h = self.config["INPUT_HEIGHT"]
self.input_c = self.config["INPUT_CHANNELS"]
self.tb_logs = self.config["TB_LOGS"]
self.batch_size = self.config["BATCH_SIZE"]
self.train_loss, self.train_accuracy = [], []
self.val_loss, self.val_acc = [], []
self.model_version = 0 # used for saving the model
self.saver = None
self.images_tr, self.labels_tr = None, None
self.images_val, self.labels_val = None, None
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.Session()
self.batch_size_pl = tf.placeholder(tf.int64, shape=[], name="batch_size")
self.is_training_pl = tf.placeholder(tf.bool, name="is_training")
self.with_dropout_pl = tf.placeholder(tf.bool, name="with_dropout")
self.keep_prob_pl = tf.placeholder(tf.float32, shape=None, name="keep_rate")
self.inputs_pl = tf.placeholder(tf.float32, [None, self.input_h, self.input_w, self.input_c])
self.labels_pl = tf.placeholder(tf.int64, [None, self.input_h, self.input_w, 1])
# Before enter the images into the architecture, we need to do Local Contrast Normalization
# But it seems a bit complicated, so we use Local Response Normalization which implement in Tensorflow
# Reference page:https://www.tensorflow.org/api_docs/python/tf/nn/local_response_normalization
self.norm1 = tf.nn.lrn(self.inputs_pl, depth_radius=5, bias=1.0, alpha=0.0001, beta=0.75, name='norm1')
# first box of convolution layer,each part we do convolution two times, so we have conv1_1, and conv1_2
self.conv1_1 = conv_layer(self.norm1, "conv1_1", [3, 3, 3, 64], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv1_2 = conv_layer(self.conv1_1, "conv1_2", [3, 3, 64, 64], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.pool1, self.pool1_index, self.shape_1 = max_pool(self.conv1_2, 'pool1')
# Second box of convolution layer(4)
self.conv2_1 = conv_layer(self.pool1, "conv2_1", [3, 3, 64, 128], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv2_2 = conv_layer(self.conv2_1, "conv2_2", [3, 3, 128, 128], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.pool2, self.pool2_index, self.shape_2 = max_pool(self.conv2_2, 'pool2')
# Third box of convolution layer(7)
self.conv3_1 = conv_layer(self.pool2, "conv3_1", [3, 3, 128, 256], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv3_2 = conv_layer(self.conv3_1, "conv3_2", [3, 3, 256, 256], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv3_3 = conv_layer(self.conv3_2, "conv3_3", [3, 3, 256, 256], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.pool3, self.pool3_index, self.shape_3 = max_pool(self.conv3_3, 'pool3')
# Fourth box of convolution layer(10)
if self.bayes:
self.dropout1 = tf.layers.dropout(self.pool3, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout1")
self.conv4_1 = conv_layer(self.dropout1, "conv4_1", [3, 3, 256, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
else:
self.conv4_1 = conv_layer(self.pool3, "conv4_1", [3, 3, 256, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv4_2 = conv_layer(self.conv4_1, "conv4_2", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv4_3 = conv_layer(self.conv4_2, "conv4_3", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.pool4, self.pool4_index, self.shape_4 = max_pool(self.conv4_3, 'pool4')
# Fifth box of convolution layers(13)
if self.bayes:
self.dropout2 = tf.layers.dropout(self.pool4, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout2")
self.conv5_1 = conv_layer(self.dropout2, "conv5_1", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
else:
self.conv5_1 = conv_layer(self.pool4, "conv5_1", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv5_2 = conv_layer(self.conv5_1, "conv5_2", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.conv5_3 = conv_layer(self.conv5_2, "conv5_3", [3, 3, 512, 512], self.is_training_pl, self.use_vgg,
self.vgg_param_dict)
self.pool5, self.pool5_index, self.shape_5 = max_pool(self.conv5_3, 'pool5')
# ---------------------So Now the encoder process has been Finished--------------------------------------#
# ------------------Then Let's start Decoder Process-----------------------------------------------------#
# First box of deconvolution layers(3)
if self.bayes:
self.dropout3 = tf.layers.dropout(self.pool5, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout3")
self.deconv5_1 = up_sampling(self.dropout3, self.pool5_index, self.shape_5, self.batch_size_pl,
name="unpool_5")
else:
self.deconv5_1 = up_sampling(self.pool5, self.pool5_index, self.shape_5, self.batch_size_pl,
name="unpool_5")
self.deconv5_2 = conv_layer(self.deconv5_1, "deconv5_2", [3, 3, 512, 512], self.is_training_pl)
self.deconv5_3 = conv_layer(self.deconv5_2, "deconv5_3", [3, 3, 512, 512], self.is_training_pl)
self.deconv5_4 = conv_layer(self.deconv5_3, "deconv5_4", [3, 3, 512, 512], self.is_training_pl)
# Second box of deconvolution layers(6)
if self.bayes:
self.dropout4 = tf.layers.dropout(self.deconv5_4, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout4")
self.deconv4_1 = up_sampling(self.dropout4, self.pool4_index, self.shape_4, self.batch_size_pl,
name="unpool_4")
else:
self.deconv4_1 = up_sampling(self.deconv5_4, self.pool4_index, self.shape_4, self.batch_size_pl,
name="unpool_4")
self.deconv4_2 = conv_layer(self.deconv4_1, "deconv4_2", [3, 3, 512, 512], self.is_training_pl)
self.deconv4_3 = conv_layer(self.deconv4_2, "deconv4_3", [3, 3, 512, 512], self.is_training_pl)
self.deconv4_4 = conv_layer(self.deconv4_3, "deconv4_4", [3, 3, 512, 256], self.is_training_pl)
# Third box of deconvolution layers(9)
if self.bayes:
self.dropout5 = tf.layers.dropout(self.deconv4_4, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout5")
self.deconv3_1 = up_sampling(self.dropout5, self.pool3_index, self.shape_3, self.batch_size_pl,
name="unpool_3")
else:
self.deconv3_1 = up_sampling(self.deconv4_4, self.pool3_index, self.shape_3, self.batch_size_pl,
name="unpool_3")
self.deconv3_2 = conv_layer(self.deconv3_1, "deconv3_2", [3, 3, 256, 256], self.is_training_pl)
self.deconv3_3 = conv_layer(self.deconv3_2, "deconv3_3", [3, 3, 256, 256], self.is_training_pl)
self.deconv3_4 = conv_layer(self.deconv3_3, "deconv3_4", [3, 3, 256, 128], self.is_training_pl)
# Fourth box of deconvolution layers(11)
if self.bayes:
self.dropout6 = tf.layers.dropout(self.deconv3_4, rate=(1 - self.keep_prob_pl),
training=self.with_dropout_pl, name="dropout6")
self.deconv2_1 = up_sampling(self.dropout6, self.pool2_index, self.shape_2, self.batch_size_pl,
name="unpool_2")
else:
self.deconv2_1 = up_sampling(self.deconv3_4, self.pool2_index, self.shape_2, self.batch_size_pl,
name="unpool_2")
self.deconv2_2 = conv_layer(self.deconv2_1, "deconv2_2", [3, 3, 128, 128], self.is_training_pl)
self.deconv2_3 = conv_layer(self.deconv2_2, "deconv2_3", [3, 3, 128, 64], self.is_training_pl)
# Fifth box of deconvolution layers(13)
self.deconv1_1 = up_sampling(self.deconv2_3, self.pool1_index, self.shape_1, self.batch_size_pl,
name="unpool_1")
self.deconv1_2 = conv_layer(self.deconv1_1, "deconv1_2", [3, 3, 64, 64], self.is_training_pl)
self.deconv1_3 = conv_layer(self.deconv1_2, "deconv1_3", [3, 3, 64, 64], self.is_training_pl)
with tf.variable_scope('conv_classifier') as scope:
self.kernel = variable_with_weight_decay('weights', initializer=initialization(1, 64),
shape=[1, 1, 64, self.num_classes], wd=False)
self.conv = tf.nn.conv2d(self.deconv1_3, self.kernel, [1, 1, 1, 1], padding='SAME')
self.biases = variable_with_weight_decay('biases', tf.constant_initializer(0.0),
shape=[self.num_classes], wd=False)
self.logits = tf.nn.bias_add(self.conv, self.biases, name=scope.name)
def __init__(self):
# Number classes possible per pixel
self.n_classes = FLAGS.n_classes
# Paths to dataset (training/test/validation) summary files
self.train_file = os.path.join(FLAGS.data_dir, 'train.txt')
self.val_file = os.path.join(FLAGS.data_dir, 'validate.txt')
self.test_file = os.path.join(FLAGS.data_dir, 'test.txt')
self.input_w, self.input_h, self.input_c = FLAGS.input_dims
self.images_tr, self.labels_tr = None, None
self.images_val, self.labels_val = None, None
# Create placeholders
self.batch_size_pl = tf.placeholder(tf.int64,
shape=[],
name="batch_size")
self.is_training_pl = tf.placeholder(tf.bool, name="is_training")
self.with_dropout_pl = tf.placeholder(tf.bool, name="with_dropout")
self.keep_prob_pl = tf.placeholder(tf.float32,
shape=None,
name="keep_rate")
self.inputs_pl = tf.placeholder(
tf.float32, [None, self.input_h, self.input_w, self.input_c])
self.labels_pl = tf.placeholder(tf.int64,
[None, self.input_h, self.input_w, 1])
##################
# SegNet Encoder #
##################
# SegNet includes Local Contrast Normalization - Substituted for Local Response Normalization
self.norm1 = tf.nn.lrn(self.inputs_pl,
depth_radius=5,
bias=1.0,
alpha=0.0001,
beta=0.75,
name='norm1')
# First set of convolution layers
self.conv1_1 = conv_layer(self.norm1, "conv1_1", [3, 3, 3, 64],
self.is_training_pl)
self.conv1_2 = conv_layer(self.conv1_1, "conv1_2", [3, 3, 64, 64],
self.is_training_pl)
self.pool1, self.pool1_index, self.shape_1 = max_pool(
self.conv1_2, 'pool1')
# Second set of convolution layers
self.conv2_1 = conv_layer(self.pool1, "conv2_1", [3, 3, 64, 128],
self.is_training_pl)
self.conv2_2 = conv_layer(self.conv2_1, "conv2_2", [3, 3, 128, 128],
self.is_training_pl)
self.pool2, self.pool2_index, self.shape_2 = max_pool(
self.conv2_2, 'pool2')
# Third set of convolution layers
self.conv3_1 = conv_layer(self.pool2, "conv3_1", [3, 3, 128, 256],
self.is_training_pl)
self.conv3_2 = conv_layer(self.conv3_1, "conv3_2", [3, 3, 256, 256],
self.is_training_pl)
self.conv3_3 = conv_layer(self.conv3_2, "conv3_3", [3, 3, 256, 256],
self.is_training_pl)
self.pool3, self.pool3_index, self.shape_3 = max_pool(
self.conv3_3, 'pool3')
# Fourth set of convolution layers
self.conv4_1 = conv_layer(self.pool3, "conv4_1", [3, 3, 256, 512],
self.is_training_pl)
self.conv4_2 = conv_layer(self.conv4_1, "conv4_2", [3, 3, 512, 512],
self.is_training_pl)
self.conv4_3 = conv_layer(self.conv4_2, "conv4_3", [3, 3, 512, 512],
self.is_training_pl)
self.pool4, self.pool4_index, self.shape_4 = max_pool(
self.conv4_3, 'pool4')
# Fifth set of convolution layers
self.conv5_1 = conv_layer(self.pool4, "conv5_1", [3, 3, 512, 512],
self.is_training_pl)
self.conv5_2 = conv_layer(self.conv5_1, "conv5_2", [3, 3, 512, 512],
self.is_training_pl)
self.conv5_3 = conv_layer(self.conv5_2, "conv5_3", [3, 3, 512, 512],
self.is_training_pl)
self.pool5, self.pool5_index, self.shape_5 = max_pool(
self.conv5_3, 'pool5')
##################
# SegNet Decoder #
##################
# First set of deconvolution layers
self.deconv5_1 = up_sampling(self.pool5,
self.pool5_index,
self.shape_5,
self.batch_size_pl,
name="unpool_5")
self.deconv5_2 = conv_layer(self.deconv5_1, "deconv5_2",
[3, 3, 512, 512], self.is_training_pl)
self.deconv5_3 = conv_layer(self.deconv5_2, "deconv5_3",
[3, 3, 512, 512], self.is_training_pl)
self.deconv5_4 = conv_layer(self.deconv5_3, "deconv5_4",
[3, 3, 512, 512], self.is_training_pl)
# Second set of deconvolution layers
self.deconv4_1 = up_sampling(self.deconv5_4,
self.pool4_index,
self.shape_4,
self.batch_size_pl,
name="unpool_4")
self.deconv4_2 = conv_layer(self.deconv4_1, "deconv4_2",
[3, 3, 512, 512], self.is_training_pl)
self.deconv4_3 = conv_layer(self.deconv4_2, "deconv4_3",
[3, 3, 512, 512], self.is_training_pl)
self.deconv4_4 = conv_layer(self.deconv4_3, "deconv4_4",
[3, 3, 512, 256], self.is_training_pl)
# Third set of deconvolution layers
self.deconv3_1 = up_sampling(self.deconv4_4,
self.pool3_index,
self.shape_3,
self.batch_size_pl,
name="unpool_3")
self.deconv3_2 = conv_layer(self.deconv3_1, "deconv3_2",
[3, 3, 256, 256], self.is_training_pl)
self.deconv3_3 = conv_layer(self.deconv3_2, "deconv3_3",
[3, 3, 256, 256], self.is_training_pl)
self.deconv3_4 = conv_layer(self.deconv3_3, "deconv3_4",
[3, 3, 256, 128], self.is_training_pl)
# Fourth set of deconvolution layers
self.deconv2_1 = up_sampling(self.deconv3_4,
self.pool2_index,
self.shape_2,
self.batch_size_pl,
name="unpool_2")
self.deconv2_2 = conv_layer(self.deconv2_1, "deconv2_2",
[3, 3, 128, 128], self.is_training_pl)
self.deconv2_3 = conv_layer(self.deconv2_2, "deconv2_3",
[3, 3, 128, 64], self.is_training_pl)
# Fifth set of deconvolution layers
self.deconv1_1 = up_sampling(self.deconv2_3,
self.pool1_index,
self.shape_1,
self.batch_size_pl,
name="unpool_1")
self.deconv1_2 = conv_layer(self.deconv1_1, "deconv1_2",
[3, 3, 64, 64], self.is_training_pl)
self.deconv1_3 = conv_layer(self.deconv1_2, "deconv1_3",
[3, 3, 64, 64], self.is_training_pl)
with tf.variable_scope('conv_classifier') as scope:
self.kernel = variable_with_weight_decay(
'weights',
initializer=initialization(1, 64),
shape=[1, 1, 64, self.n_classes],
wd=False)
self.conv = tf.nn.conv2d(self.deconv1_3,
self.kernel, [1, 1, 1, 1],
padding='SAME')
self.biases = variable_with_weight_decay(
'biases',
tf.constant_initializer(0.0),
shape=[self.n_classes],
wd=False)
self.logits = tf.nn.bias_add(self.conv,
self.biases,
name=scope.name)