-
Notifications
You must be signed in to change notification settings - Fork 3
/
reconstruct.py
148 lines (124 loc) · 6.68 KB
/
reconstruct.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
import os
import sys
import importlib
from datetime import datetime
import plyfile
import numpy as np
import torch
from torch.optim import Adam
from torch.optim.lr_scheduler import StepLR
from torch.utils.tensorboard import SummaryWriter
from skimage.measure import marching_cubes
from network import Network, CodeCloud
from dataset import ShapeNet
class LossHelper:
def __init__(self, num_shapes, num_iters):
self.num_shapes = num_shapes
self.num_iters = num_iters
self.clear()
def clear(self):
self.loss_dict = None
def accumulate(self, loss_dict, shape_idx, iter_idx):
if self.loss_dict is None:
self.loss_dict = {}
for k in loss_dict.keys():
self.loss_dict[k] = np.zeros((self.num_shapes, self.num_iters))
for k, v in loss_dict.items():
self.loss_dict[k][shape_idx][iter_idx] = v.item()
def write_log(self, writer):
if self.loss_dict is None:
return
for k in self.loss_dict.keys():
self.loss_dict[k] = np.mean(self.loss_dict[k], axis=0)
for k, v in self.loss_dict.items():
for i in range(self.num_iters):
writer.add_scalar(k, v[i], global_step=i)
def optimize_latent_codes(network, dataset, shape_idx, config, loss_helper):
network.code_cloud = CodeCloud(config['network']['code_cloud'], 1).cuda()
network.code_cloud.train()
network.decoder.eval()
optimizer = Adam(network.code_cloud.parameters(), lr=config['reconstruct']['init_lr'])
lr_scheduler = StepLR(optimizer, step_size=config['reconstruct']['lr_decay_step'], gamma=config['reconstruct']['lr_decay_rate'])
for iter_idx in range(config['reconstruct']['num_iteration']):
query_points, gt_sd, _ = dataset[shape_idx]
query_points = torch.tensor(query_points).float().unsqueeze(0).cuda()
gt_sd = torch.tensor(gt_sd).float().unsqueeze(0).cuda()
indices = torch.zeros(1, dtype=torch.long).cuda()
optimizer.zero_grad()
pred_sd = network(indices, query_points)
loss_dict = network.loss(gt_sd)
loss_helper.accumulate(loss_dict, shape_idx, iter_idx)
loss_dict['total_loss'].backward()
optimizer.step()
lr_scheduler.step()
print('Optimizing latent code, progress: %d/%d...' % (iter_idx, config['reconstruct']['num_iteration']), end='\r')
print('Optimizing latent code, progress: %d/%d. Done.' % (config['reconstruct']['num_iteration'], config['reconstruct']['num_iteration']))
def latent_codes_to_mesh(network, config):
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), 'Reconstructing mesh(marching cubes resolution=%d).'%config['reconstruct']['marching_cubes_resolution'])
N = config['reconstruct']['marching_cubes_resolution']
voxel_size = 1.1 / (N - 1)
voxel_origin = [-0.55, -0.55, -0.55]
overall_index = torch.arange(0, N**3, 1, dtype=torch.long)
query_points = torch.zeros(N**3, 3)
query_points[:, 2] = overall_index % N
query_points[:, 1] = (overall_index / N).long() % N
query_points[:, 0] = ((overall_index / N) / N).long() % N
query_points[:, 0] = (query_points[:, 0] * voxel_size) + voxel_origin[2]
query_points[:, 1] = (query_points[:, 1] * voxel_size) + voxel_origin[1]
query_points[:, 2] = (query_points[:, 2] * voxel_size) + voxel_origin[0]
query_points_sdf = np.zeros(N**3)
network.eval()
query_points.requires_grad = False
batch_indices = torch.zeros(1, dtype=torch.long).cuda()
batch_size = 1024
num_query_points = N ** 3
query_point_index = 0
while query_point_index < num_query_points:
batch_query_points = query_points[query_point_index : min(query_point_index+batch_size, num_query_points)].unsqueeze(0).cuda()
pred_sd = network(batch_indices, batch_query_points).squeeze(0)
query_points_sdf[query_point_index : min(query_point_index+batch_size, num_query_points)] = pred_sd.cpu().detach().numpy()
query_point_index += batch_size
sdf_values = query_points_sdf.reshape(N, N, N)
vertices, faces, _, _ = marching_cubes(sdf_values, level=0.0, spacing=[voxel_size]*3)
vertices[:, 0] += voxel_origin[0]
vertices[:, 1] += voxel_origin[1]
vertices[:, 2] += voxel_origin[2]
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), 'Reconstruction done.')
return vertices, faces
def reconstruct(config):
dataset = ShapeNet(config['recon_dataset'])
loss_helper = LossHelper(len(dataset), config['reconstruct']['num_iteration'])
writer = SummaryWriter(config['experiment']['recon_log_dir'])
network = Network(config['network'], 1).cuda()
ckpt_decoder = torch.load(config['reconstruct']['ckpt_decoder'])
network.decoder.load_state_dict(ckpt_decoder['decoder'])
for shape_idx in range(len(dataset)):
print('****** %s ******\ntime: %s\ndata index: %d/%d' % (config['experiment']['name'], datetime.now().strftime('%Y-%m-%d %H:%M:%S'), shape_idx, len(dataset)))
optimize_latent_codes(network, dataset, shape_idx, config, loss_helper)
data_id = dataset.split[shape_idx]
save_path = os.path.join(config['experiment']['recon_latent_codes_dir'], data_id+'.pth')
save_path_dir = os.path.dirname(save_path)
os.makedirs(save_path_dir, exist_ok=True)
torch.save({'latent_codes': network.code_cloud.state_dict()}, save_path)
print('Optimized latent code is saved to ' + save_path)
vertices, faces = latent_codes_to_mesh(network, config)
vertices_tuple = np.zeros((vertices.shape[0],), dtype=[('x', 'f4'), ('y', 'f4'), ('z', 'f4')])
for i in range(vertices.shape[0]):
vertices_tuple[i] = tuple(vertices[i, :])
faces_tuple = []
for i in range(faces.shape[0]):
faces_tuple.append(((faces[i, :].tolist(),)))
faces_tuple = np.array(faces_tuple, dtype=[('vertex_indices', 'i4', (3,))])
ply_data = plyfile.PlyData([plyfile.PlyElement.describe(vertices_tuple, 'vertex'), plyfile.PlyElement.describe(faces_tuple, 'face')])
save_path = os.path.join(config['experiment']['recon_meshes_dir'], data_id+'.ply')
save_path_dir = os.path.dirname(save_path)
os.makedirs(save_path_dir, exist_ok=True)
ply_data.write(save_path)
print('Reconstructed mesh is saved to ' + save_path)
loss_helper.write_log(writer)
writer.close()
if __name__ == '__main__':
config = importlib.import_module(sys.argv[1]).config
os.makedirs(config['experiment']['recon_latent_codes_dir'])
os.makedirs(config['experiment']['recon_meshes_dir'])
reconstruct(config)