def test_RDLR():
# set options
opt = Option()
opt.root_dir = root + '/dataset/test3000'
opt.checkpoints_dir = root + '/checkpoints/RDLR'
root_result = root + '/dataset/result3000'
opt.gpu_ids = [0]
opt.batch_size = 4
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.is_train = False
# load data
root_dir_train = opt.root_dir
dataset_train = RDLRDataLoader(root_dir=root_dir_train,
train=False,
coarse=opt.coarse)
data_loader_test = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = RDLRModel()
model.initialize(opt)
model.load_networks(-1)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch, 0)
model.forward()
fake_R = model.fake_street_R.detach().cpu()
fake_L = model.fake_street_L.detach().cpu()
fake_sate_D = model.fake_sate_D.detach().cpu()
fake_sate_L = model.fake_sate_L.detach().cpu()
n, c, h, w = fake_R.size()
for i in range(0, n):
rgb = fake_R[i, :, :, :] * 0.5 + 0.5
label = fake_L[i, :, :, :] * 0.5 + 0.5
sate_depth = fake_sate_D[i, :, :, :] * 0.5 + 0.5
sate_label = fake_sate_L[i, :, :, :] * 0.5 + 0.5
img_id = data_batch['img_id'][i]
# save image
path_depth = root_result + '/' + img_id + '_pred_depth_w_mask.png'
path_sate_label = root_result + '/' + img_id + '_pred_label_w_mask.png'
path_rgb = root_result + '/' + img_id + '_pred_rgb_w_mask.png'
path_label = root_result + '/' + img_id + '_pred_sem_w_mask.png'
torchvision.utils.save_image(sate_depth.float(), path_depth)
torchvision.utils.save_image(sate_label.float(), path_sate_label)
torchvision.utils.save_image(rgb.float(), path_rgb)
torchvision.utils.save_image(label.float(), path_label)
def train_DLR():
# set options
opt = Option()
opt.root_dir = root + '/dataset/DLR/'
opt.checkpoints_dir = root + '/checkpoints/DLR'
opt.gpu_ids = [0]
opt.batch_size = 16
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.learning_rate = 1e-3 # learning rage
# load data
root_dir_train = opt.root_dir + '/train'
dataset_train = DLRDataLoader(root_dir=root_dir_train,
train=True,
coarse=opt.coarse)
data_loader_train = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
print(opt)
# load model
model = DLRModel()
model.initialize(opt)
model.load_networks(-1)
# do training
for epoch in range(opt.epoch_count, opt.niter + opt.niter_decay + 1):
file = open(opt.root_dir + '/logs.txt', 'a')
for idx_batch, data_batch in enumerate(data_loader_train):
print(idx_batch)
model.set_input(data_batch, epoch)
model.optimize_parameters()
print('epoch: ' + str(epoch)
#+ ', train loss_G1_Loss: ' + str(model.loss_G_L1.data)
+ ', train loss_G2_Loss: ' + str(model.loss_G_L2.data) +
', train loss_GAN_Loss: ' + str(model.loss_G_GAN.data) +
', train loss_D_Loss: ' + str(model.loss_D.data))
file.write('epoch: ' + str(epoch) + ', train loss_G_Loss: ' +
str(model.loss_G.data) + ', train loss_D_Loss: ' +
str(model.loss_D.data) + '\n')
file.close()
# save
if epoch % 5 == 0:
model.save_networks(epoch)
def test_DLL():
# set options
opt = Option()
opt.root_dir = root + '/dataset'
opt.checkpoints_dir = root + '/checkpoints/DLL'
opt.gpu_ids = [0]
opt.batch_size = 8
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.is_train = False
# load data
root_dir_train = opt.root_dir + '/test3000'
dataset_train = DLLDataLoader(root_dir=root_dir_train,
train=False,
coarse=opt.coarse)
data_loader_test = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = DLLModel()
model.initialize(opt)
model.load_networks(30)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch)
model.forward()
fake_S = model.fake_S.detach().cpu()
n, c, h, w = fake_S.size()
for i in range(0, n):
label = fake_S[i, :, :, :] * 0.5 + 0.5
label = label.numpy()
label_rgb = np.zeros([256, 256, 3]).astype(np.uint8)
label_rgb[:, :, 2] = label[0, :, :] * 255
label_rgb[:, :, 1] = label[1, :, :] * 255
label_rgb[:, :, 0] = label[2, :, :] * 255
label_rgb = cv2.resize(label_rgb, (512, 256))
img_id = data_batch['img_id'][i]
# save image
path_label = root_dir_train + '/' + img_id + '_pred_sem_dll.png'
#torchvision.utils.save_image(label_rgb, path_label)
cv2.imwrite(path_label, label_rgb)
def test_R2D():
# set options
opt = Option()
opt.root_dir = root + '/dataset/R2D'
opt.checkpoints_dir = root + '/checkpoints/R2D'
opt.gpu_ids = [0]
opt.batch_size = 16
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
# load data
root_dir_test = opt.root_dir + '/test'
dataset_test = R2DDataLoader(root_dir=root_dir_test,
train=True,
coarse=opt.coarse)
data_loader_test = DataLoader(dataset_test,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = R2DModel()
model.initialize(opt)
model.load_networks(-1)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch)
model.forward()
fake_L = model.fake_L.detach().cpu()
fake_D = model.fake_D.detach().cpu()
n, c, h, w = fake_L.size()
for i in range(0, n):
label = fake_L[i, :, :, :] * 0.5 + 0.5
depth = fake_D[i, :, :, :] * 0.5 + 0.5
img_id = data_batch['img_id'][i]
# save image
path_depth = 'F:/' + img_id + '_pred_depth.png'
path_label = 'F:/' + img_id + '_label.png'
#torchvision.utils.save_image(depth.float(), path_depth)
torchvision.utils.save_image(label.float(), path_label)
torchvision.utils.save_image(depth.float(), path_depth)
def test_D2L():
# set options
opt = Option()
opt.root_dir = root + '/dataset/test'
opt.checkpoints_dir = root + '/checkpoints/D2L'
opt.result_dir = opt.root_dir
opt.gpu_ids = [0]
opt.batch_size = 16
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.is_train = False
opt.fine_tune_sidewalk = False
# load data
root_dir_train = opt.root_dir
dataset_train = D2LDataLoader(root_dir=root_dir_train,
train=opt.is_train,
coarse=opt.coarse,
fine_tune_sidewalk=opt.fine_tune_sidewalk)
data_loader_test = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = D2LModel()
model.initialize(opt)
model.load_networks(50)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch, 0)
model.forward()
fake_S = model.fake_S.detach().cpu()
n, c, h, w = fake_S.size()
for i in range(0, n):
sem = fake_S[i, :, :, :] * 0.5 + 0.5
img_id = data_batch['img_id'][i]
# save image
path_sem = root_dir_train + '/' + img_id + '_pred_sem_wo_mask.png'
#torchvision.utils.save_image(depth.float(), path_depth)
torchvision.utils.save_image(sem.float(), path_sem)
def test_L2R():
# set options
opt = Option()
opt.root_dir = root + '/dataset/'
opt.checkpoints_dir = root + '/checkpoints/L2R'
opt.gpu_ids = [0]
opt.batch_size = 16
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.is_train = False
# load data
root_dir_train = opt.root_dir + '/test3000'
dataset_train = L2RDataLoader(root_dir=root_dir_train,
train=False,
coarse=opt.coarse)
data_loader_test = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = L2RModel()
model.initialize(opt)
model.load_networks(-1)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch)
model.forward()
fake_R = model.fake_R.detach().cpu()
n, c, h, w = fake_R.size()
for i in range(0, n):
rgb = fake_R[i, :, :, :] * 0.5 + 0.5
img_id = data_batch['img_id'][i]
# save image
path_rgb = root_dir_train + '/' + img_id + '_pred_rgb_dll.png'
#torchvision.utils.save_image(depth.float(), path_depth)
torchvision.utils.save_image(rgb.float(), path_rgb)
def train_R2D():
# set options
opt = Option()
opt.root_dir = 'D:/permanent/aligned_2k/train_R2D'
opt.checkpoints_dir = 'C:/Users/lu.2037/Downloads/ICCV2019/checkpoints/R2D'
opt.gpu_ids = [0]
opt.batch_size = 16
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
# load data
root_dir_train = opt.root_dir
dataset_train = R2DDataLoader(root_dir=root_dir_train, train=True, coarse=opt.coarse)
data_loader_train = DataLoader(dataset_train,batch_size=opt.batch_size,
shuffle=opt.shuffle, num_workers=opt.num_workers, pin_memory=opt.pin_memory)
# load model
model = R2DModel()
model.initialize(opt)
model.load_networks(-1)
# do training
for epoch in range(opt.epoch_count, opt.niter + opt.niter_decay + 1):
file = open(opt.root_dir+'/logs.txt', 'a')
for idx_batch, data_batch in enumerate(data_loader_train):
print(idx_batch)
model.set_input(data_batch)
model.optimize_parameters()
print('epoch: ' + str(epoch) + ', train loss_G_Loss1: ' + str(model.loss_G_Loss1.data) +
', train loss_G_Loss2: ' + str(model.loss_G_Loss2.data) )
file.write('epoch: ' + str(epoch) + ', train loss_G_Loss1: ' + str(model.loss_G_Loss1.data) +
', train loss_G_Loss2: ' + str(model.loss_G_Loss2.data) + '\n')
file.close()
# save
if epoch%5 ==0:
model.save_networks(epoch)
from os.path import isfile, join
import re
from model import R2DModel, D2LModel, L2RModel, DLRModel, RDLRModel, DLLModel, L2RAllModel, L2RNoiseModel
from utils import Option
from dataset import R2DDataLoader, D2LDataLoader, L2RDataLoader, DLRDataLoader, RDLRDataLoader, DLLDataLoader, L2RAllDataLoader
from geo_process_layer import depth2voxel, voxel2pano
import socket
opt = Option()
host_name = socket.gethostname()
if host_name == 'cnb-d102-04a':
root = '/local/zoli/xiaohu_iccv2019'
noise_b = False
opt.batch_size = 4
opt.use_multiple_G = True
elif host_name == 'cvg-desktop-17-ubuntu':
root = '/home/zoli/xiaohu_iccv2019'
noise_b = True
opt.batch_size = 1
opt.use_multiple_G = False
else:
raise ValueError('Root Error!')
def test_L2RAll():
# set options
opt.root_dir = root + '/dataset/L2R_Zuoyue'
opt.checkpoints_dir = root + '/checkpoints/L2R_Zuoyue_Noise'
opt.gpu_ids = [0]
def test_RDLR():
t = '5'
# set options
opt = Option()
opt.root_dir = 'D:/permanent/aligned_2k/test_augment/test_' + t
opt.checkpoints_dir = 'C:/Users/lu.2037/Downloads/ICCV2019/checkpoints/RDLR'
root_result = 'D:/permanent/aligned_2k/test_augment/test_' + t
opt.gpu_ids = [0]
opt.batch_size = 4
opt.coarse = False
opt.pool_size = 0
opt.no_lsgan = True
opt.is_train = False
# load data
root_dir_train = opt.root_dir
dataset_train = RDLRDataLoader(root_dir=root_dir_train,
train=False,
coarse=opt.coarse)
data_loader_test = DataLoader(dataset_train,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
# load model
model = RDLRModel()
model.initialize(opt)
model.load_networks(-1)
# do testung
for idx_batch, data_batch in enumerate(data_loader_test):
print(idx_batch)
model.set_input(data_batch, 0)
model.forward()
fake_R = model.fake_street_R.detach().cpu()
fake_L = model.fake_street_L.detach().cpu()
fake_sate_D = model.fake_sate_D.detach().cpu()
fake_sate_L = model.fake_sate_L.detach().cpu()
fake_proj_dis = model.proj_D.detach().cpu()
n, c, h, w = fake_R.size()
for i in range(0, n):
rgb = fake_R[i, :, :, :] * 0.5 + 0.5
label = fake_L[i, :, :, :] * 0.5 + 0.5
sate_depth = fake_sate_D[i, :, :, :] * 0.5 + 0.5
sate_label = fake_sate_L[i, :, :, :] * 0.5 + 0.5
proj_depth = fake_proj_dis[i, :, :, :] * 0.5 + 0.5
img_id = data_batch['img_id'][i]
# save image
tt = "_0" + t
path_depth = root_result + '/' + img_id + '_pred_depth' + tt + '.png'
path_sate_label = root_result + '/' + img_id + '_pred_label' + tt + '.png'
path_rgb = root_result + '/' + img_id + '_pred_rgb' + tt + '.png'
path_label = root_result + '/' + img_id + '_pred_sem' + tt + '.png'
path_proj_dis = root_result + '/' + img_id + '_proj_dis' + tt + '.png'
torchvision.utils.save_image(sate_depth.float(), path_depth)
# torchvision.utils.save_image(sate_label.float(), path_sate_label)
torchvision.utils.save_image(rgb.float(), path_rgb)
torchvision.utils.save_image(label.float(), path_label)
torchvision.utils.save_image(proj_depth.float(), path_proj_dis)
from skimage import io, transform, color
from skimage.transform import resize
import time
from model import R2DModel, D2LModel, L2RModel, DLRModel, RDLRModel, DLLModel, L2RAllModel, L2RNoiseModel
from utils import Option
from dataset import R2DDataLoader, D2LDataLoader, L2RDataLoader, DLRDataLoader, RDLRDataLoader, DLLDataLoader, L2RAllDataLoader
import socket
opt = Option()
host_name = socket.gethostname()
if host_name == 'cnb-d102-04a':
root = '/local/zoli/xiaohu_iccv2019'
noise_b = False
opt.batch_size = 16
opt.use_multiple_G = True
elif host_name == 'cvg-desktop-17-ubuntu':
root = '/home/zoli/xiaohu_iccv2019'
noise_b = True
opt.batch_size = 8
opt.use_multiple_G = False
else:
raise ValueError('Root Error!')
def train_L2R_Zuoyue():
# set options
opt.root_dir = root + '/dataset/L2R_Zuoyue'
opt.checkpoints_dir = root + '/checkpoints/L2R_Zuoyue_Noise'
opt.gpu_ids = [0]
