-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmain.py
168 lines (136 loc) · 6.08 KB
/
main.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
163
164
165
166
167
168
import os
import numpy as np
import torch
import torch.nn as nn
import torch.optim
from torch.utils.tensorboard import SummaryWriter
from models import SDFModule
from render import Renderer
from utils import *
def gauss_kernel(size=5, device=torch.device('cuda:0'), channels=3):
kernel = torch.tensor([[1., 4., 6., 4., 1],
[4., 16., 24., 16., 4.],
[6., 24., 36., 24., 6.],
[4., 16., 24., 16., 4.],
[1., 4., 6., 4., 1.]])
kernel /= 256.
kernel = kernel.repeat(channels, 1, 1, 1)
kernel = kernel.to(device)
return kernel
def downsample(x):
return x[:, :, ::2, ::2]
def conv_gauss(img, kernel):
img = torch.nn.functional.pad(img, (2, 2, 2, 2), mode='reflect')
out = torch.nn.functional.conv2d(img, kernel.cuda(), groups=img.shape[1])
return out
def img_loss(imgs, target_imgs, multi_scale=True):
loss = 0
kernel = gauss_kernel()
count = 0
for i in range(imgs.shape[0]):
count += 1
loss = loss + (imgs[i] - target_imgs[i]).square().mean()
if multi_scale:
current_est = imgs[i].permute(2, 0, 1).unsqueeze(0)
current_gt = target_imgs[i].permute(2, 0, 1).unsqueeze(0)
for j in range(4):
filtered_est = conv_gauss(current_est, kernel)
filtered_gt = conv_gauss(current_gt, kernel)
down_est = downsample(filtered_est)
down_gt = downsample(filtered_gt)
current_est = down_est
current_gt = down_gt
loss = loss + (current_est - current_gt).square().mean() / (j+1)
loss = loss / count
return loss
def main(config):
model_cfg = Namespace(dim=3, out_dim=1,
hidden_size=512,
n_blocks=4, z_dim=1,
const=60.)
module = SDFModule(cfg=model_cfg, f=config.init_ckpt).cuda()
logger = SummaryWriter(log_dir=config.expdir, flush_secs=5)
with torch.no_grad():
R = Renderer(config.num_views, config.res, fname=config.mesh,
scale=config.scale)
target_imgs = R.target_imgs
logger.add_image('target', target_imgs[-1].permute(2, 0, 1).clamp(0, 1))
optimizer = torch.optim.Adam(list(module.parameters()), lr=config.lr,
weight_decay=config.weight_decay)
gt_sdf = torch.zeros(config.max_v, 1).cuda()
F = torch.zeros(config.max_v, 1).cuda()
vertices = torch.zeros((config.max_v, 3)).cuda()
normals = torch.zeros((config.max_v, 3)).cuda()
faces = torch.empty((config.max_v, 3), dtype=torch.int32).cuda()
vertices.requires_grad_()
for e in range(config.epochs):
laplace_lam = config.max_laplace_lam
if e >= config.fine_e:
laplace_lam = config.min_laplace_lam
mesh_res = config.mesh_res_limit + np.random.randint(low=-3, high=3)
else:
laplace_lam = config.max_laplace_lam
mesh_res = config.mesh_res_base + np.random.randint(low=-3, high=3)
with torch.no_grad():
vertices_np, faces_np = module.get_zero_points(mesh_res=mesh_res)
v = vertices_np.shape[0]
f = faces_np.shape[0]
vertices.data[:v] = torch.from_numpy(vertices_np)
faces.data[:f] = torch.from_numpy(np.ascontiguousarray(faces_np))
vertices.grad = None
edges = compute_edges(vertices[:v], faces[:f])
L = laplacian_simple(vertices[:v], edges.long())
laplacian_loss = torch.trace(((L @ vertices[:v]).T @ vertices[:v]))
face_normals = compute_face_normals(vertices[:v], faces[:f])
vertex_normals = compute_vertex_normals(vertices[:v], faces[:f], face_normals)
imgs = R.render(vertices[:v], faces[:f], vertex_normals)
loss = img_loss(imgs, target_imgs, multi_scale=True)
loss = loss + laplace_lam * laplacian_loss
loss.backward()
logger.add_scalar('loss', loss.item(), global_step=e)
with torch.no_grad():
dE_dx = vertices.grad[:v].detach()
idx = 0
while idx < v:
optimizer.zero_grad()
min_i = idx
max_i = min(min_i + config.batch_size, v)
vertices_subset = vertices[min_i:max_i]
vertices_subset.requires_grad_()
pred_sdf = module.forward(vertices_subset.unsqueeze(0)).squeeze(0)
normals[min_i:max_i] = gradient(pred_sdf, vertices_subset).detach()
F[min_i:max_i] = torch.nan_to_num(torch.sum(normals[min_i:max_i] *\
dE_dx[min_i:max_i], dim=-1, keepdim=True))
gt_sdf[min_i:max_i] = (pred_sdf + config.eps * F[min_i:max_i]).detach()
idx += config.batch_size
n_batches = v // config.batch_size
for j in range(config.iters):
optimizer.zero_grad()
idx = 0
while idx < v:
min_i = idx
max_i = min(min_i + config.batch_size, v)
vertices_subset = vertices[min_i:max_i].detach()
pred_sdf = module.forward(vertices_subset.unsqueeze(0)).squeeze(0)
loss = (gt_sdf[min_i:max_i] - pred_sdf).abs().mean() / n_batches
loss.backward()
idx += config.batch_size
optimizer.step()
if e % 1 == 0:
print(f'Iter: {e}')
if e % config.img_log_freq == 0:
logger.add_image('est', imgs[-1].permute(2, 0, 1).clamp(0, 1), global_step=e)
if e % config.mesh_log_freq == 0:
with torch.no_grad():
mse = (imgs-target_imgs).square().mean()
psnr = -10.0*torch.log10(mse)
logger.add_scalar('psnr', psnr, global_step=e)
mesh = trimesh.Trimesh(vertices_np, faces_np)
cd = compute_trimesh_chamfer(R.mesh, mesh)
logger.add_scalar('cd', cd, global_step=e)
mesh.export(f'{config.expdir}/mesh_{e:04d}.ply')
if e % config.ckpt_log_freq == 0:
torch.save(module.state_dict(), f'{config.expdir}/iter_{e:04d}.ckpt')
if __name__ == '__main__':
config = parse_config(create_dir=False)
main(config)