source = load_data_source(args.data_source)
source.load_data(args.data_dir, 0.1)
print('[i] # training samples: ', source.num_training)
print('[i] # validation samples: ', source.num_validation)
print('[i] # classes: ', source.num_classes)
print('[i] Image size: ', source.image_size)
train_generator = source.train_generator
valid_generator = source.valid_generator
label_colors = source.label_colors
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess)
net.build_from_vgg(args.vgg_dir, source.num_classes, progress_hook='tqdm')
labels = tf.placeholder(tf.float32,
shape=[None, None, None, source.num_classes])
optimizer, loss = net.get_optimizer(labels)
summary_writer = tf.summary.FileWriter(args.tensorboard_dir, sess.graph)
saver = tf.train.Saver(max_to_keep=10)
label_mapper = tf.argmax(labels, axis=3)
n_train_batches = int(math.ceil(source.num_training / args.batch_size))
initialize_uninitialized_variables(sess)
print('[i] Training...')
#-------------------------------------------------------------------------------
print('[i] Project name: ', args.name)
print('[i] Network checkpoint:', checkpoint_file)
print('[i] Metagraph file: ', metagraph_file)
print('[i] Number of samples: ', len(samples))
print('[i] Output directory: ', args.output_dir)
print('[i] Image size: ', source.image_size)
print('[i] # classes: ', source.num_classes)
print('[i] Batch size: ', args.batch_size)
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess, source.num_classes)
net.build_from_metagraph(metagraph_file, checkpoint_file)
#---------------------------------------------------------------------------
# Process the images
#---------------------------------------------------------------------------
generator = sample_generator(samples, source.image_size, args.batch_size)
n_sample_batches = int(math.ceil(len(samples) / args.batch_size))
description = '[i] Processing samples'
for x, names in tqdm(generator,
total=n_sample_batches,
desc=description,
unit='batches'):
feed = {net.image_input: x, net.keep_prob: 1, net.is_training: False}
img_labels = sess.run(net.classes, feed_dict=feed)
except (ImportError, AttributeError, RuntimeError) as e:
print('[!] Unable to load data source:', str(e))
sys.exit(1)
loss_list = []
val_loss_list = []
psnr_list = []
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
rd_feature = tf.placeholder(
tf.float32, shape=[None, None, None, 3], name='rd_feature_train')
net = FCNVGG(sess, rd_feature)
net.build_from_vgg(args.vgg_dir, source.num_classes, progress_hook='tqdm')
labels = tf.placeholder(tf.int32,
shape=[None, None, None, source.num_classes], name='labels_train')
raw_image = tf.placeholder(
tf.float32, shape=[None, None, None, 3], name='raw_img_train')
optimizer, loss_mse, loss_ce, psnr, fm = net.get_optimizer(labels)
summary_writer = tf.summary.FileWriter(args.tensorboard_dir, sess.graph)
saver = tf.train.Saver(max_to_keep=10)
label_mapper = tf.argmax(labels, axis=3)
n_train_batches = int(math.ceil(source.num_training/args.batch_size))
#-------------------------------------------------------------------------------
print('[i] Project name: ', args.name)
print('[i] Network checkpoint:', checkpoint_file)
print('[i] Metagraph file: ', metagraph_file)
print('[i] Number of samples: ', len(samples))
print('[i] Output directory: ', args.output_dir)
print('[i] Image size: ', source.image_size)
print('[i] # classes: ', source.num_classes)
print('[i] Batch size: ', args.batch_size)
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess)
net.build_from_metagraph(metagraph_file, checkpoint_file)
#---------------------------------------------------------------------------
# Process the images
#---------------------------------------------------------------------------
generator = sample_generator(samples, source.image_size, args.batch_size)
n_sample_batches = int(math.ceil(len(samples)/args.batch_size))
description = '[i] Processing samples'
for x, names in tqdm(generator, total=n_sample_batches,
desc=description, unit='batches'):
feed = {net.image_input: x,
net.keep_prob: 1}
img_labels = sess.run(net.classes, feed_dict=feed)
imgs = draw_labels_batch(x, img_labels, label_colors, False)
#-------------------------------------------------------------------------------
print('[i] Project name: ', args.name)
print('[i] Network checkpoint:', checkpoint_file)
print('[i] Metagraph file: ', metagraph_file)
print('[i] Number of samples: ', len(samples))
print('[i] Output directory: ', args.output_dir)
print('[i] Image size: ', source.image_size)
print('[i] # classes: ', source.num_classes)
print('[i] Batch size: ', args.batch_size)
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess)
net.build_from_metagraph(metagraph_file, checkpoint_file)
#---------------------------------------------------------------------------
# Process the images
#---------------------------------------------------------------------------
generator = sample_generator(samples, source.image_size, args.batch_size)
n_sample_batches = int(math.ceil(len(samples) / args.batch_size))
description = '[i] Processing samples'
for x, names in tqdm(generator,
total=n_sample_batches,
desc=description,
unit='batches'):
feed = {net.image_input: x, net.keep_prob: 1}
img_labels = sess.run(net.classes, feed_dict=feed)
def run(self):
self.__disp_config()
source = self.__load_data_source()
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess, source.num_classes)
net.build_from_vgg(args.vgg_dir,
source.num_classes,
progress_hook='tqdm')
net.get_optimizer(net.labels)
net.add_summary_nodes(args.tensorboard_dir)
validation_imgs = tf.placeholder(tf.float32,
shape=[None, None, None, 3])
validation_img_summary_op = tf.summary.image(
'validation_img', validation_imgs)
train_img_summary_op = tf.summary.image('train_img',
validation_imgs)
saver = tf.train.Saver(max_to_keep=100)
print('[i] Training...')
utils.initialize_uninitialized_variables(sess)
n_train_batches = int(
math.ceil(source.num_training / args.batch_size))
cur_step = 0
step_num = n_train_batches * args.epochs
for e in range(args.epochs):
train_generator = source.train_generator(args.batch_size)
for x, y in train_generator:
cur_step += 1
feed = {
net.image_input: x,
net.labels: y,
net.keep_prob: 0.5,
net.is_training: True
}
sess.run(net.reset_iou_op)
summary, _, loss_batch, _, label_mapper, img_classes, f1, accuracy = sess.run(
[
net.merged, net.update_iou_op, net.loss,
net.optimizer, net.label_mapper, net.classes,
net.f1, net.accuracy
],
feed_dict=feed)
net.train_writer.add_summary(summary, cur_step)
iou, summary = sess.run(
[net.metric_iou__op, net.merged_update])
net.train_writer.add_summary(summary, cur_step)
print(
"step {}:{}/{}: loss={}, iou={}, f1={}, acc={}".format(
e + 1, cur_step, step_num, loss_batch, iou, f1,
accuracy))
#output trainig input image
if (cur_step) % 10 == 0:
val_imgs = x[:1, :, :, :]
val_img_labels = img_classes[:1, :, :]
val_img_labels_gt = label_mapper[:1, :, :]
imgs_inferred = utils.draw_labels_batch(
val_imgs, val_img_labels, source.label_colors)
imgs_gt = utils.draw_labels_batch(
val_imgs, val_img_labels_gt, source.label_colors)
val_imgs = utils.convert_rgb_batch(val_imgs)
all_imgs = np.concatenate(
[val_imgs, imgs_gt, imgs_inferred], axis=0)
summary = sess.run(
train_img_summary_op,
feed_dict={validation_imgs: all_imgs})
net.train_writer.add_summary(summary, cur_step)
#monitor inference on valiaton data
if (cur_step) % 10 == 0:
val_generator = source.valid_generator(args.batch_size)
#jut try out one batch
x, y = next(val_generator)
feed = {
net.image_input: x,
net.labels: y,
net.keep_prob: 1.0,
net.is_training: False
}
sess.run(net.reset_iou_op)
summary, _, loss_batch, label_mapper, img_classes, f1, accuracy = sess.run(
[
net.merged, net.update_iou_op, net.loss,
net.label_mapper, net.classes, net.f1,
net.accuracy
],
feed_dict=feed)
net.val_writer.add_summary(summary, cur_step)
iou, summary = sess.run(
[net.metric_iou__op, net.merged_update])
net.val_writer.add_summary(summary, cur_step)
print(
"#####validation: step {}/{}: loss={}, iou={}, f1={}, acc={}#####"
.format(cur_step, step_num, loss_batch, iou, f1,
accuracy))
val_imgs = x[:1, :, :, :]
val_img_labels = img_classes[:1, :, :]
val_img_labels_gt = label_mapper[:1, :, :]
imgs_inferred = utils.draw_labels_batch(
val_imgs, val_img_labels, source.label_colors)
imgs_gt = utils.draw_labels_batch(
val_imgs, val_img_labels_gt, source.label_colors)
val_imgs = utils.convert_rgb_batch(val_imgs)
all_imgs = np.concatenate(
[val_imgs, imgs_gt, imgs_inferred], axis=0)
summary = sess.run(
validation_img_summary_op,
feed_dict={validation_imgs: all_imgs})
net.val_writer.add_summary(summary, cur_step)
if (e + 1) % args.checkpoint_interval == 0:
checkpoint = '{}/e{}.ckpt'.format(args.name, e + 1)
saver.save(sess, checkpoint)
print('Checkpoint saved:', checkpoint)
return
def load_fcnvgg(session, num_classes):
return FCNVGG(session, num_classes)
print('[i] # validation samples: ', source.num_validation)
print('[i] # classes: ', source.num_classes)
print('[i] Image size: ', source.image_size)
train_generator = source.train_generator
valid_generator = source.valid_generator
label_colors = source.label_colors
except (ImportError, AttributeError, RuntimeError) as e:
print('[!] Unable to load data source:', str(e))
sys.exit(1)
#-------------------------------------------------------------------------------
# Create the network
#-------------------------------------------------------------------------------
with tf.Session() as sess:
print('[i] Creating the model...')
net = FCNVGG(sess)
net.build_from_vgg(args.vgg_dir, source.num_classes, progress_hook='tqdm')
labels = tf.placeholder(tf.float32,
shape=[None, None, None, source.num_classes])
optimizer, loss = net.get_optimizer(labels)
summary_writer = tf.summary.FileWriter(args.tensorboard_dir, sess.graph)
saver = tf.train.Saver(max_to_keep=10)
label_mapper = tf.argmax(labels, axis=3)
n_train_batches = int(math.ceil(source.num_training/args.batch_size))
initialize_uninitialized_variables(sess)
print('[i] Training...')