-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathsfs_optim.py
215 lines (173 loc) · 9.03 KB
/
sfs_optim.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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import os
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:256'
from os.path import join
from glob import glob
from tqdm import tqdm
import argparse
from pyhocon import ConfigFactory
import numpy as np
import cv2
import trimesh
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Adam, RMSprop, SGD
import nvdiffrast.torch as dr
from models.utils import get_normals, get_radiance, get_matrix, meshcleaning, laplacian_smoothing, mynormalize
from models.sap import PSR2Mesh, DPSR, grid_interp, point_rasterize, sap_transform, sap_generate, gen_inputs
from get_data import get_dtu_data, get_nhr_data
def main(conf, scan_id):
conf = ConfigFactory.parse_file(conf)
data_path = conf.get_string('data_path')
data_type = conf.get_string('data_type')
num = conf.get_int('num')
w = conf.get_int('w')
h = conf.get_int('h')
resolution = (h, w)
sfs_weight = conf.get_float('sfs_weight')
lap_weight = conf.get_float('lap_weight')
albedeo_weight = conf.get_float('albedo_weight')
mask_weight = conf.get_float('mask_weight')
edge_weight = conf.get_float('edge_weight')
degree = conf.get_int('degree')
batch = conf.get_int('batch')
input_mesh_dire = conf.get_string('input_mesh_dire')
out_mesh_dire = conf.get_string('out_mesh_dire')
lr = conf.get_float('lr')
albedo_lr = conf.get_float('albedo_lr')
mesh = trimesh.load(join(input_mesh_dire, '%d.obj'%scan_id), process=False, maintain_order=True)
vertices = torch.from_numpy(mesh.vertices.astype(np.float32)).cuda()
faces = torch.from_numpy(mesh.faces.astype(np.int32)).cuda()
glctx = dr.RasterizeGLContext()
if data_type == 'nhr':
imgs, grayimgs, masks, w2cs, projs = get_nhr_data(data_path, scan_id, num, (w,h))
num = imgs.shape[0]
# compute sphere harmonic coefficient as initialization
with torch.no_grad():
valid_normals = []
valid_grayimgs = []
for k in range(0,num,batch):
n = min(num, k+batch) - k
w2c = w2cs[k:k+batch]
proj = projs[k:k+batch]
mask = masks[k:k+batch]
grayimg = grayimgs[k:k+batch]
vertsw = torch.cat([vertices, torch.ones_like(vertices[:,0:1])], axis=1).unsqueeze(0).expand(n,-1,-1)
rot_verts = torch.einsum('ijk,ikl->ijl', vertsw, w2c)
proj_verts = torch.einsum('ijk,ikl->ijl', rot_verts, proj)
normals = get_normals(vertsw[:,:,:3], faces.long())
rast_out, _ = dr.rasterize(glctx, proj_verts, faces, resolution=resolution)
feat, _ = dr.interpolate(normals, rast_out, faces)
pred_normals = feat.contiguous()
pred_normals = dr.antialias(pred_normals, rast_out, proj_verts, faces)
pred_normals = F.normalize(pred_normals,p=2,dim=3)
valid_idx = (mask > 0) & (rast_out[:,:,:,3] > 0)
valid_normals.append(pred_normals[valid_idx].detach().cpu().numpy())
valid_grayimgs.append(grayimg[valid_idx].detach().cpu().numpy())
valid_normals = np.concatenate(valid_normals, axis=0)
valid_grayimgs = np.concatenate(valid_grayimgs, axis=0)
matrix = get_matrix(valid_normals, degree)
sh_coeff = np.linalg.lstsq(matrix, valid_grayimgs, rcond=None)[0]
sh_coeff = torch.from_numpy(sh_coeff.astype(np.float32)).cuda()
sh_coeff.requires_grad_(True)
albedo = torch.ones_like(vertices).unsqueeze(0) * 0.01
albedo.requires_grad_(True)
optimizer = Adam([{'params': albedo, 'lr': 0.02}, {'params': sh_coeff, 'lr': 0.001}])
pbar = tqdm(range(100))
for i in pbar:
perm = torch.randperm(num).cuda()
for k in range(0, num, batch):
n = min(num, k+batch) - k
w2c = w2cs[perm[k:k+batch]]
proj = projs[perm[k:k+batch]]
img = imgs[perm[k:k+batch]]
mask = masks[perm[k:k+batch]]
vertsw = torch.cat([vertices, torch.ones_like(vertices[:,0:1])], axis=1).unsqueeze(0).expand(n,-1,-1)
rot_verts = torch.einsum('ijk,ikl->ijl', vertsw, w2c)
proj_verts = torch.einsum('ijk,ikl->ijl', rot_verts, proj)
normals = get_normals(vertsw[:,:,:3], faces.long())
rast_out, _ = dr.rasterize(glctx, proj_verts, faces, resolution=resolution)
feat = torch.cat([normals, albedo.expand(n,-1,-1)], dim=2)
feat, _ = dr.interpolate(feat, rast_out, faces)
pred_normals = feat[:,:,:,:3].contiguous()
rast_albedo = feat[:,:,:,3:6].contiguous()
pred_normals = dr.antialias(pred_normals, rast_out, proj_verts, faces)
pred_normals = F.normalize(pred_normals,p=2,dim=3)
rast_albedo = dr.antialias(rast_albedo, rast_out, proj_verts, faces)
valid_idx = (mask > 0) & (rast_out[:,:,:,3] > 0)
valid_normals = pred_normals[valid_idx]
valid_albedo = rast_albedo[valid_idx]
valid_img = img[valid_idx]
radiance = get_radiance(sh_coeff, valid_normals, degree).unsqueeze(-1)
pred_img = radiance * valid_albedo
sfs_loss = sfs_weight * F.l1_loss(pred_img, valid_img)
albedo_loss = albedeo_weight * laplacian_smoothing(albedo.squeeze(0), faces.long(), method="uniform")
loss = sfs_loss + albedo_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
des = 'sfs:%.4f'%sfs_loss.item() + ' albedo:%.4f'%albedo_loss.item()
pbar.set_description(des)
vertices.requires_grad_(True)
verts_optimizer = Adam([{'params': vertices, 'lr': lr}, {'params': sh_coeff, 'lr': 0.001}])
albedo_optimizer = Adam([{'params': albedo, 'lr': albedo_lr}])
pbar = tqdm(range(100))
for i in pbar:
perm = torch.randperm(num).cuda()
for k in range(0, num, batch):
n = min(num, k+batch) - k
w2c = w2cs[perm[k:k+batch]]
proj = projs[perm[k:k+batch]]
img = imgs[perm[k:k+batch]]
mask = masks[perm[k:k+batch]]
vertsw = torch.cat([vertices, torch.ones_like(vertices[:,0:1])], axis=1).unsqueeze(0).expand(n,-1,-1)
rot_verts = torch.einsum('ijk,ikl->ijl', vertsw, w2c)
proj_verts = torch.einsum('ijk,ikl->ijl', rot_verts, proj)
normals = get_normals(vertsw[:,:,:3], faces.long())
rast_out, _ = dr.rasterize(glctx, proj_verts, faces, resolution=resolution)
feat = torch.cat([normals, albedo.expand(n,-1,-1), torch.ones_like(vertsw[:,:,:1])], dim=2)
feat, _ = dr.interpolate(feat, rast_out, faces)
pred_normals = feat[:,:,:,:3].contiguous()
rast_albedo = feat[:,:,:,3:6].contiguous()
pred_mask = feat[:,:,:,6:7].contiguous()
pred_normals = F.normalize(pred_normals,p=2,dim=3)
pred_mask = dr.antialias(pred_mask, rast_out, proj_verts, faces).squeeze(-1)
valid_idx = (mask > 0) & (rast_out[:,:,:,3] > 0)
valid_normals = pred_normals[valid_idx]
valid_albedo = rast_albedo[valid_idx]
valid_img = img[valid_idx]
radiance = get_radiance(sh_coeff, valid_normals, degree).unsqueeze(-1)
pred_img = radiance * valid_albedo
tmp_img = torch.zeros_like(img)
tmp_img[valid_idx] = pred_img
tmp_img = dr.antialias(tmp_img, rast_out, proj_verts, faces)
sfs_loss = sfs_weight * F.l1_loss(tmp_img[valid_idx], valid_img)
lap_loss = lap_weight * laplacian_smoothing(vertices, faces.long(), method="uniform")
albedo_loss = albedeo_weight * laplacian_smoothing(albedo.squeeze(0), faces.long(), method="uniform")
mask_loss = mask_weight * F.mse_loss(pred_mask, mask)
a = vertices[faces[:, 0].long()]
b = vertices[faces[:, 1].long()]
c = vertices[faces[:, 2].long()]
edge_loss = edge_weight * torch.cat([((a - b) ** 2).sum(1), ((c - b) ** 2).sum(1), ((a - c) ** 2).sum(1)]).mean()
loss = sfs_loss + lap_loss + albedo_loss + mask_loss + edge_loss
verts_optimizer.zero_grad()
albedo_optimizer.zero_grad()
# optimizer.zero_grad()
loss.backward()
# optimizer.step()
verts_optimizer.step()
albedo_optimizer.step()
des = 'sfs:%.4f'%sfs_loss.item() + ' lap:%.4f'%lap_loss.item() + ' albedo:%.4f'%albedo_loss.item() +\
' mask:%.4f'%mask_loss.item() + ' edge:%.4f'%edge_loss.item()
pbar.set_description(des)
torch.save({'sh_coeff': sh_coeff, 'albedo': albedo}, join(out_mesh_dire, '%d.pt'%scan_id))
np_verts = vertices.squeeze().detach().cpu().numpy()
np_faces = faces.squeeze().detach().cpu().numpy()
mesh = trimesh.Trimesh(np_verts, np_faces)
mesh.export(join(out_mesh_dire, '%d.obj'%scan_id))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--conf', type=str, default='confs/thu_sfs.conf')
parser.add_argument('--scan_id', type=int, default=0)
args = parser.parse_args()
main(args.conf, args.scan_id)