-
Notifications
You must be signed in to change notification settings - Fork 28
/
run_single_shape.py
163 lines (146 loc) · 8.1 KB
/
run_single_shape.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
import numpy as np
import os, sys
import cv2
import torch
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from core.dataset import LoaderSingle
from core.inv_optimizer import optimize_single_view
from core.evaluation import *
from core.utils.render_utils import *
from core.utils.decoder_utils import load_decoder
from core.visualize.vis_utils import *
from core.visualize import Visualizer
from core.sdfrenderer import SDFRenderer
import pickle
def init_info():
# TODO
# set this mesh_data_dir to the path to your NormalizationParameters and SurfaceSamples
mesh_data_dir = os.path.expanduser('data')
# mesh_data_dir = os.path.expanduser('~/data')
basedir = os.path.join(os.path.dirname(os.path.abspath(__file__)))
data_dir = os.path.join(basedir, 'data')
model_dir = os.path.join(basedir, 'deepsdf')
experiment_directory = os.path.join(model_dir, 'experiments/sofas')
split_file = os.path.join(model_dir, 'examples/splits/sv2_sofas_test.json')
synthetic_data_dir = os.path.join(data_dir, 'demo_singleview_syn')
return mesh_data_dir, experiment_directory, split_file, synthetic_data_dir
def train(args):
# initialize output_dir
if not os.path.exists(args.vis_folder):
os.makedirs(args.vis_folder)
#########################################################################
# load data
#########################################################################
mesh_data_dir, experiment_directory, split_file, synthetic_data_dir = init_info()
upper_loader = LoaderSingle(synthetic_data_dir, mesh_data_dir, experiment_directory, split_file)
shape_md5, image_data, mesh_data, camera, depth = upper_loader[6]
img, _, normal, _ = image_data
gt_samples, norm_params = mesh_data
points_gt = np.array(gt_samples.vertices)
points_gt = (points_gt + norm_params['offset']) * norm_params['scale']
gt_pack = {}
gt_pack['depth'] = torch.from_numpy(depth).cuda()
gt_pack['normal'] = torch.from_numpy(normal).cuda()
gt_pack['silhouette'] = torch.from_numpy(depth < 1e5).type(torch.uint8).cuda()
# # visualize gt
# cv2.imwrite('img.png', img)
# visualize_depth('test0.png', depth)
# with open('camera.pkl', 'wb') as f:
# pickle.dump(camera, f)
#########################################################################
# initialize tensor
#########################################################################
decoder = load_decoder(experiment_directory, args.checkpoint)
decoder = decoder.module.cuda()
evaluator = Evaluator(decoder)
latent_size = 256
std_ = 0.1
rand_tensor = torch.ones(1, latent_size).normal_(mean=0, std=std_)
if args.use_random_init:
latent_tensor = rand_tensor
else:
latent_code_dir = os.path.join(synthetic_data_dir, 'latent_codes', '{0}.pth'.format(shape_md5))
latent_code = torch.load(latent_code_dir)
latent_tensor = latent_code[0].detach().cpu()
latent_size = latent_tensor.shape[-1]
if not args.use_gt_code:
latent_tensor = latent_tensor + rand_tensor
latent_tensor = latent_tensor.float().cuda()
if (not args.profile) and (not args.no_pretest):
points_tmp = evaluator.latent_vec_to_points(latent_tensor, fname=os.path.join(args.vis_folder, 'initial.ply'), silent=True)
if points_tmp is None:
print('The current latent code does not correspond to a valid shape.')
dist = None
else:
dist = evaluator.compute_chamfer_distance(points_gt, points_tmp)
print('STD: {0:.3f}'.format(std_))
print('CHAMFER DISTANCE: {0:.3f}'.format(dist * 1000))
latent_tensor.requires_grad = True
optimizer_latent = torch.optim.Adam([latent_tensor], lr=args.lr)
#########################################################################
# optimization
#########################################################################
weight_dict = {}
weight_dict['w_depth'] = 10.0
weight_dict['w_normal'] = 5.0
weight_dict['w_mask_gt'] = 1.0
weight_dict['w_mask_out'] = 1.0
weight_dict['w_l2reg'] = 1.0
img_h, img_w = img.shape[0], img.shape[1]
img_hw = (img_h, img_w)
print('Image size: {0}.'. format(img_hw))
if args.visualize:
visualizer = Visualizer(img_hw)
visualizer.add_chamfer(dist)
else:
visualizer = None
# initialize renderer
if args.use_multiscale:
sdf_renderer = SDFRenderer(decoder, camera.intrinsic, img_hw=img_hw, march_step=100, buffer_size=1, threshold=args.threshold, ray_marching_ratio=args.ratio, use_depth2normal=args.use_depth2normal)
sdf_renderer_1_2 = SDFRenderer(decoder, downsize_camera_intrinsic(camera.intrinsic, 2), march_step=100, buffer_size=3, threshold=args.threshold, use_depth2normal=args.use_depth2normal)
sdf_renderer_1_4 = SDFRenderer(decoder, downsize_camera_intrinsic(camera.intrinsic, 4), march_step=100, buffer_size=5, threshold=args.threshold, use_depth2normal=args.use_depth2normal)
renderer_list = [sdf_renderer, sdf_renderer_1_2, sdf_renderer_1_4]
else:
sdf_renderer = SDFRenderer(decoder, camera.intrinsic, img_hw=img_hw, march_step=100, buffer_size=args.buffer_size, threshold=args.threshold, ray_marching_ratio=args.ratio, use_depth2normal=args.use_depth2normal)
renderer_list = [sdf_renderer]
extrinsic = torch.from_numpy(camera.extrinsic).float().cuda()
if args.oracle:
num_iters = 1
else:
num_iters = args.num_iters
# optimization start
latent_tensor, optimizer_latent = optimize_single_view(renderer_list, evaluator, optimizer_latent, latent_tensor, extrinsic, gt_pack, weight_dict, optimizer_type="shape", num_iters=num_iters, points_gt=points_gt, test_step=args.test_step, profile=args.profile, visualizer=visualizer, ray_marching_type=args.method, vis_folder=args.vis_folder)
# Main
if __name__ == '__main__':
import argparse
arg_parser = argparse.ArgumentParser(
description="Use differentiable renderer to optimize shapes from 2D observations."
)
arg_parser.add_argument("--checkpoint", "-c", dest="checkpoint", default="2000",
help='The checkpoint weights to use. This can be a number indicated an epoch or "latest" '
+ "for the latest weights (this is the default)",
)
# test settings
arg_parser.add_argument('--gpu', '-g', default='0', help='gpu id.')
arg_parser.add_argument('--test_step', '-t', type=int, default=50, help='test step.')
arg_parser.add_argument('--lr', type=float, default=1e-3, help='learning rate.')
arg_parser.add_argument('--num_iters', type=int, default=200, help='number of iterations.')
arg_parser.add_argument('--profile', action='store_true', help='renderer profiling.')
arg_parser.add_argument('--visualize', action='store_true', help='visualization flag.')
arg_parser.add_argument('--vis_folder', type=str, default='vis/demo_singleview_shape', help='folder for visualization.')
arg_parser.add_argument('--oracle', action='store_true', help='oracle rendering feedforward')
arg_parser.add_argument('--no_pretest', action='store_true', help='do not test initialization performance. just to speed up')
arg_parser.add_argument('--use_gt_code', action='store_true', help='use groundtruth shape code')
arg_parser.add_argument('--use_random_init', action='store_true', help='use random initialization')
# renderer settings
arg_parser.add_argument('--ratio', type=float, default=1.5, help='test step.')
arg_parser.add_argument('--method', type=str, default='pyramid_recursive', help='ray marching implementation.')
arg_parser.add_argument('--threshold', type=float, default=5e-5, help='threshold')
arg_parser.add_argument('--buffer_size', type=int, default=3, help='buffer size')
arg_parser.add_argument('--use_depth2normal', action='store_true', help='use normal converted from depth')
arg_parser.add_argument('--use_multiscale', action='store_true', help='use multiscale optimization')
args = arg_parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
if args.oracle:
args.no_pretest, args.visualize = True, True
train(args)