def fine_tune_grid_exp(height_mode,
layers_to_keep_num,
learning_rate,
decay_step,
decay_rate,
epochs,
model_name):
# set gpu
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = flags.GPU
# environment settings
np.random.seed(flags.random_seed)
tf.set_random_seed(flags.random_seed)
# get weight
tf.reset_default_graph()
if height_mode == 'subtract':
kernel = sis_utils.get_unet_first_layer_weight(flags.pre_trained_model, 1)
elif height_mode == 'all':
kernel = sis_utils.get_unet_first_layer_weight(flags.pre_trained_model, 2)
else:
kernel = sis_utils.get_unet_first_layer_weight(flags.pre_trained_model, 3)
tf.reset_default_graph()
# data prepare step
Data = rsrClassData(flags.rsr_data_dir)
(collect_files_train, meta_train) = Data.getCollectionByName(flags.train_data_dir)
pe_train = patch_extractor.PatchExtractorUrbanMapper(flags.rsr_data_dir,
collect_files_train, patch_size=flags.input_size,
tile_dim=meta_train['dim_image'][:2],
appendix=flags.train_patch_appendix)
train_data_dir = pe_train.extract(flags.patch_dir)
(collect_files_valid, meta_valid) = Data.getCollectionByName(flags.valid_data_dir)
pe_valid = patch_extractor.PatchExtractorUrbanMapper(flags.rsr_data_dir,
collect_files_valid, patch_size=flags.input_size,
tile_dim=meta_train['dim_image'][:2],
appendix=flags.valid_patch_appendix)
valid_data_dir = pe_valid.extract(flags.patch_dir)
# image reader
coord = tf.train.Coordinator()
# load reader
reader_train = image_reader.ImageLabelReaderHeight(train_data_dir, flags.input_size, coord,
city_list=flags.city_name, tile_list=flags.train_tile_names,
ds_name='urban_mapper', data_aug=flags.data_aug,
height_mode=flags.height_mode)
reader_valid = image_reader.ImageLabelReaderHeight(valid_data_dir, flags.input_size, coord,
city_list=flags.city_name, tile_list=flags.valid_tile_names,
ds_name='urban_mapper', data_aug=flags.data_aug,
height_mode=flags.height_mode)
reader_train_iter = reader_train.image_height_label_iterator(flags.batch_size)
reader_valid_iter = reader_valid.image_height_label_iterator(flags.batch_size)
# define place holder
if height_mode == 'all':
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 5], name='X')
elif height_mode == 'subtract':
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 4], name='X')
else:
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 6], name='X')
y = tf.placeholder(tf.int32, shape=[None, flags.input_size[0], flags.input_size[1], 1], name='y')
mode = tf.placeholder(tf.bool, name='mode')
# initialize model
model = unet.UnetModel_Height_Appendix({'X':X, 'Y':y}, trainable=mode, model_name=model_name, input_size=flags.input_size)
model.create_graph('X', flags.num_classes)
model.load_weights(flags.pre_trained_model, layers_to_keep_num, kernel)
model.make_loss('Y')
model.make_learning_rate(learning_rate,
tf.cast(flags.n_train/flags.batch_size * decay_step, tf.int32), decay_rate)
model.make_update_ops('X', 'Y')
model.make_optimizer(model.learning_rate)
# set ckdir
model.make_ckdir(flags.ckdir)
# make summary
model.make_summary()
# set up graph and initialize
config = tf.ConfigProto()
# run training
start_time = time.time()
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=1)
if os.path.exists(model.ckdir) and tf.train.get_checkpoint_state(model.ckdir):
latest_check_point = tf.train.latest_checkpoint(model.ckdir)
saver.restore(sess, latest_check_point)
print('Model Loaded {}'.format(model.ckdir))
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
try:
train_summary_writer = tf.summary.FileWriter(model.ckdir, sess.graph)
model.train('X', 'Y', epochs, flags.n_train, flags.batch_size, sess, train_summary_writer,
train_iterator=reader_train_iter, valid_iterator=reader_valid_iter, image_summary=sis_utils.image_summary)
finally:
coord.request_stop()
coord.join(threads)
saver.save(sess, '{}/model.ckpt'.format(model.ckdir), global_step=model.global_step)
duration = time.time() - start_time
print('duration {:.2f} hours'.format(duration/60/60))
def test_authentic_unet_height_fmap(rsr_data_dir,
test_data_dir,
input_size,
model_name,
num_classes,
ckdir,
city,
batch_size,
ds_name='inria',
GPU='0',
random_seed=1234,
height_mode='subtract'):
import re
import scipy.misc
import tensorflow as tf
from network import unet
from dataReader import image_reader
from rsrClassData import rsrClassData
def name_has_city(city_list, name):
for city in city_list:
if city in name:
return True
return False
city = city.split(',')
# set gpu
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = GPU
tf.reset_default_graph()
# environment settings
np.random.seed(random_seed)
tf.set_random_seed(random_seed)
# data prepare step
Data = rsrClassData(rsr_data_dir)
(collect_files_test, meta_test) = Data.getCollectionByName(test_data_dir)
# image reader
coord = tf.train.Coordinator()
# define place holder
if height_mode == 'all':
X = tf.placeholder(tf.float32,
shape=[None, input_size[0], input_size[1], 5],
name='X')
elif height_mode == 'subtract':
X = tf.placeholder(tf.float32,
shape=[None, input_size[0], input_size[1], 7],
name='X')
else:
X = tf.placeholder(tf.float32,
shape=[None, input_size[0], input_size[1], 6],
name='X')
y = tf.placeholder(tf.int32,
shape=[None, input_size[0], input_size[1], 1],
name='y')
mode = tf.placeholder(tf.bool, name='mode')
# initialize model
#model = unet.UnetModel_Origin({'X': X, 'Y': y}, trainable=mode, model_name=model_name, input_size=input_size)
model = unet.UnetModel_Height_Appendix({
'X': X,
'Y': y
},
trainable=mode,
model_name=model_name,
input_size=input_size)
model.create_graph('X', num_classes)
model.make_update_ops('X', 'Y')
# set ckdir
model.make_ckdir(ckdir)
# set up graph and initialize
config = tf.ConfigProto()
result_dict = {}
# run training
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=1)
if os.path.exists(model.ckdir) and tf.train.get_checkpoint_state(
model.ckdir):
latest_check_point = tf.train.latest_checkpoint(model.ckdir)
saver.restore(sess, latest_check_point)
print('loaded model from {}'.format(latest_check_point))
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
for (image_name, dsm_name, dtm_name, f_name,
label_name) in collect_files_test:
#if city in image_name:
if name_has_city(city, image_name):
if ds_name == 'inria':
city_name = re.findall('[a-z\-]*(?=[0-9]+\.)',
image_name)[0]
tile_id = re.findall('[0-9]+(?=\.tif)', image_name)[0]
else:
city_name = os.path.basename(image_name)[:3]
tile_id = os.path.basename(image_name)[9:12]
iter_file_list = [
os.path.join(rsr_data_dir, image_name),
os.path.join(rsr_data_dir, dsm_name),
os.path.join(rsr_data_dir, dtm_name),
os.path.join(rsr_data_dir, f_name)
]
# load reader
iterator_test = image_reader.image_height_fmap_label_iterator(
iter_file_list,
batch_size=batch_size,
tile_dim=meta_test['dim_image'][:2],
patch_size=input_size,
overlap=184,
padding=92,
height_mode=height_mode)
# run
result = model.test('X', sess, iterator_test)
pred_label_img = get_output_label(
result, (meta_test['dim_image'][0] + 184,
meta_test['dim_image'][1] + 184),
input_size,
meta_test['colormap'],
overlap=184,
output_image_dim=meta_test['dim_image'],
output_patch_size=(input_size[0] - 184,
input_size[1] - 184))
#pred_label_img[np.where(pred_label_img==2)] = 1
# evaluate
truth_label_img = scipy.misc.imread(
os.path.join(rsr_data_dir, label_name))
#truth_label_img[np.where(truth_label_img == 2)] = 1
#truth_label_img[np.where(truth_label_img == 1)] = 255
iou = iou_metric(truth_label_img, pred_label_img)
result_dict['{}{}'.format(city_name, tile_id)] = iou
coord.request_stop()
coord.join(threads)
return result_dict
def main(flags):
# set gpu
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = flags.GPU
# environment settings
np.random.seed(flags.random_seed)
tf.set_random_seed(flags.random_seed)
# data prepare step
Data = rsrClassData(flags.rsr_data_dir)
(collect_files_train, meta_train) = Data.getCollectionByName(flags.train_data_dir)
pe_train = patch_extractor.PatchExtractorInria(flags.rsr_data_dir,
collect_files_train, patch_size=flags.input_size,
tile_dim=meta_train['dim_image'][:2],
appendix=flags.train_patch_appendix,
overlap=184)
train_data_dir = pe_train.extract(flags.patch_dir, pad=184)
(collect_files_valid, meta_valid) = Data.getCollectionByName(flags.valid_data_dir)
pe_valid = patch_extractor.PatchExtractorInria(flags.rsr_data_dir,
collect_files_valid, patch_size=flags.input_size,
tile_dim=meta_valid['dim_image'][:2],
appendix=flags.valid_patch_appendix,
overlap=184)
valid_data_dir = pe_valid.extract(flags.patch_dir, pad=184)
# image reader
coord = tf.train.Coordinator()
# load reader
with tf.name_scope('image_loader'):
reader_train = image_reader.ImageLabelReader(train_data_dir, flags.input_size, coord,
city_list=flags.city_name, tile_list=flags.train_tile_names,
data_aug=flags.data_aug)
reader_valid = image_reader.ImageLabelReader(valid_data_dir, flags.input_size, coord,
city_list=flags.city_name, tile_list=flags.valid_tile_names,
data_aug=flags.data_aug)
X_batch_op, y_batch_op = reader_train.dequeue(flags.batch_size)
X_batch_op_valid, y_batch_op_valid = reader_valid.dequeue(flags.batch_size)
reader_train_op = [X_batch_op, y_batch_op]
reader_valid_op = [X_batch_op_valid, y_batch_op_valid]
# define place holder
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 3], name='X')
y = tf.placeholder(tf.int32, shape=[None, flags.input_size[0], flags.input_size[1], 1], name='y')
mode = tf.placeholder(tf.bool, name='mode')
# initialize model
flags.model_name = '{}_EP-{}_DS-{}_LR-{}'.format(flags.model_name, flags.epochs, flags.decay_step, flags.learning_rate)
model = unet.UnetModel_Height_Appendix({'X':X, 'Y':y}, trainable=mode, model_name=flags.model_name, input_size=flags.input_size)
model.create_graph('X', flags.num_classes, start_filter_num=40)
model.make_loss('Y')
model.make_learning_rate(flags.learning_rate,
tf.cast(flags.n_train/flags.batch_size * flags.decay_step, tf.int32), flags.decay_rate)
model.make_update_ops('X', 'Y')
model.make_optimizer(model.learning_rate)
# set ckdir
model.make_ckdir(flags.ckdir)
# make summary
model.make_summary()
# set up graph and initialize
config = tf.ConfigProto()
# run training
start_time = time.time()
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=1)
if os.path.exists(model.ckdir) and tf.train.get_checkpoint_state(model.ckdir):
latest_check_point = tf.train.latest_checkpoint(model.ckdir)
saver.restore(sess, latest_check_point)
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
try:
train_summary_writer = tf.summary.FileWriter(model.ckdir, sess.graph)
model.train('X', 'Y', flags.epochs, flags.n_train, flags.batch_size, sess, train_summary_writer,
train_reader=reader_train_op, valid_reader=reader_valid_op, image_summary=utils.image_summary)
finally:
coord.request_stop()
coord.join(threads)
saver.save(sess, '{}/model.ckpt'.format(model.ckdir), global_step=model.global_step)
duration = time.time() - start_time
print('duration {:.2f} hours'.format(duration/60/60))
def test_and_save(flags, model_name, save_dir):
# set gpu
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = flags.GPU
# environment settings
np.random.seed(flags.random_seed)
tf.set_random_seed(flags.random_seed)
# data prepare step
Data = rsrClassData(flags.rsr_data_dir)
(collect_files_test, meta_test) = Data.getCollectionByName(flags.test_data_dir)
# image reader
coord = tf.train.Coordinator()
# define place holder
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 3], name='X')
y = tf.placeholder(tf.int32, shape=[None, flags.input_size[0], flags.input_size[1], 1], name='y')
mode = tf.placeholder(tf.bool, name='mode')
# initialize model
if 'appendix' in model_name:
model = unet.UnetModel_Height_Appendix({'X':X, 'Y':y}, trainable=mode, model_name=model_name, input_size=flags.input_size)
elif 'Res' in model_name:
model = unet.ResUnetModel_Crop({'X': X, 'Y': y}, trainable=mode, model_name=model_name,
input_size=flags.input_size)
else:
model = unet.UnetModel_Origin({'X':X, 'Y':y}, trainable=mode, model_name=model_name, input_size=flags.input_size)
if 'large' in model_name:
model.create_graph('X', flags.num_classes, start_filter_num=40)
else:
model.create_graph('X', flags.num_classes)
model.make_update_ops('X', 'Y')
# set ckdir
model.make_ckdir(flags.ckdir)
# set up graph and initialize
config = tf.ConfigProto()
# make fold if not exists
save_path = os.path.join(save_dir, 'temp_save', model_name)
if not os.path.exists(save_path):
os.makedirs(save_path)
else:
return save_path
# run training
start_time = time.time()
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=1)
if os.path.exists(model.ckdir) and tf.train.get_checkpoint_state(model.ckdir):
latest_check_point = tf.train.latest_checkpoint(model.ckdir)
saver.restore(sess, latest_check_point)
print('loaded {}'.format(latest_check_point))
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
try:
for (image_name, label_name) in collect_files_test:
c_names = flags.city_name.split(',')
for c_name in c_names:
if c_name in image_name:
city_name = re.findall('[a-z\-]*(?=[0-9]+\.)', image_name)[0]
tile_id = re.findall('[0-9]+(?=\.tif)', image_name)[0]
print('Scoring {}_{} using {}...'.format(city_name, tile_id, model_name))
# load reader
iterator_test = image_reader.image_label_iterator(
os.path.join(flags.rsr_data_dir, image_name),
batch_size=flags.batch_size,
tile_dim=meta_test['dim_image'][:2],
patch_size=flags.input_size,
overlap=184, padding=92,
image_mean=IMG_MEAN)
# run
result = model.test('X', sess, iterator_test, soft_pred=True)
pred_label_img = sis_utils.get_output_label(result,
(meta_test['dim_image'][0]+184, meta_test['dim_image'][1]+184),
flags.input_size,
meta_test['colormap'], overlap=184,
output_image_dim=meta_test['dim_image'],
output_patch_size=(flags.input_size[0]-184, flags.input_size[1]-184),
make_map=False, soft_pred=True)
file_name = os.path.join(save_path, '{}_{}.npy'.format(city_name, tile_id))
np.save(file_name, pred_label_img)
finally:
coord.request_stop()
coord.join(threads)
duration = time.time() - start_time
print('duration {:.2f} minutes'.format(duration/60))
return save_path
def test(flags, model_name, save_dir):
# set gpu
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = flags.GPU
# environment settings
np.random.seed(flags.random_seed)
tf.set_random_seed(flags.random_seed)
# data prepare step
Data = rsrClassData(flags.rsr_data_dir)
(collect_files_test, meta_test) = Data.getCollectionByName(flags.test_data_dir)
# image reader
coord = tf.train.Coordinator()
# define place holder
X = tf.placeholder(tf.float32, shape=[None, flags.input_size[0], flags.input_size[1], 3], name='X')
y = tf.placeholder(tf.int32, shape=[None, flags.input_size[0], flags.input_size[1], 1], name='y')
mode = tf.placeholder(tf.bool, name='mode')
# initialize model
if 'appendix' in model_name:
model = unet.UnetModel_Height_Appendix({'X':X, 'Y':y}, trainable=mode, model_name=model_name, input_size=flags.input_size)
else:
model = unet.UnetModel_Origin({'X':X, 'Y':y}, trainable=mode, model_name=model_name, input_size=flags.input_size)
if 'large' in model_name:
model.create_graph('X', flags.num_classes, start_filter_num=40)
else:
model.create_graph('X', flags.num_classes)
model.make_update_ops('X', 'Y')
# set ckdir
model.make_ckdir(flags.ckdir)
# set up graph and initialize
config = tf.ConfigProto()
# run training
start_time = time.time()
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=1)
if os.path.exists(model.ckdir) and tf.train.get_checkpoint_state(model.ckdir):
latest_check_point = tf.train.latest_checkpoint(model.ckdir)
saver.restore(sess, latest_check_point)
print('loaded {}'.format(latest_check_point))
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
try:
iou_record = {}
for (image_name, label_name) in collect_files_test:
c_names = flags.city_name.split(',')
for c_name in c_names:
if c_name in image_name:
city_name = re.findall('[a-z\-]*(?=[0-9]+\.)', image_name)[0]
tile_id = re.findall('[0-9]+(?=\.tif)', image_name)[0]
# load reader
iterator_test = image_reader.image_label_iterator(
os.path.join(flags.rsr_data_dir, image_name),
batch_size=flags.batch_size,
tile_dim=meta_test['dim_image'][:2],
patch_size=flags.input_size,
overlap=184, padding=92,
image_mean=IMG_MEAN)
# run
result = model.test('X', sess, iterator_test)
pred_label_img = sis_utils.get_output_label(result,
(meta_test['dim_image'][0]+184, meta_test['dim_image'][1]+184),
flags.input_size,
meta_test['colormap'], overlap=184,
output_image_dim=meta_test['dim_image'],
output_patch_size=(flags.input_size[0]-184, flags.input_size[1]-184),
make_map=False)
# evaluate
truth_label_img = scipy.misc.imread(os.path.join(flags.rsr_data_dir, label_name))
iou = sis_utils.iou_metric(truth_label_img, pred_label_img * 255)
'''ax1 = plt.subplot(121)
ax1.imshow(truth_label_img)
ax2 = plt.subplot(122, sharex=ax1, sharey=ax1)
ax2.imshow(pred_label_img)
plt.show()'''
iou_record[image_name] = iou
print('{}_{}: iou={:.2f}'.format(city_name, tile_id, iou*100))
finally:
coord.request_stop()
coord.join(threads)
duration = time.time() - start_time
print('duration {:.2f} minutes'.format(duration/60))
np.save(os.path.join(save_dir, '{}.npy').format(model.model_name), iou_record)
iou_mean = []
for _, val in iou_record.items():
iou_mean.append(val)
print(np.mean(iou_mean))
return np.mean(iou_mean)