-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathshapenet_train.py
161 lines (124 loc) · 6.17 KB
/
shapenet_train.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
import argparse
import json
import copy
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
from datasets.shapenet import build_shapenet
from models.nerf import build_nerf
from models.rendering import get_rays_shapenet, sample_points, volume_render
def inner_loop(model, optim, imgs, poses, hwf, bound, num_samples, raybatch_size, inner_steps):
"""
train the inner model for a specified number of iterations
"""
pixels = imgs.reshape(-1, 3)
rays_o, rays_d = get_rays_shapenet(hwf, poses)
rays_o, rays_d = rays_o.reshape(-1, 3), rays_d.reshape(-1, 3)
num_rays = rays_d.shape[0]
for step in range(inner_steps):
indices = torch.randint(num_rays, size=[raybatch_size])
raybatch_o, raybatch_d = rays_o[indices], rays_d[indices]
pixelbatch = pixels[indices]
t_vals, xyz = sample_points(raybatch_o, raybatch_d, bound[0], bound[1],
num_samples, perturb=True)
optim.zero_grad()
rgbs, sigmas = model(xyz)
colors = volume_render(rgbs, sigmas, t_vals, white_bkgd=True)
loss = F.mse_loss(colors, pixelbatch)
loss.backward()
optim.step()
def train_meta(args, meta_model, meta_optim, data_loader, device):
"""
train the meta_model for one epoch using reptile meta learning
https://arxiv.org/abs/1803.02999
"""
for imgs, poses, hwf, bound in data_loader:
imgs, poses, hwf, bound = imgs.to(device), poses.to(device), hwf.to(device), bound.to(device)
imgs, poses, hwf, bound = imgs.squeeze(), poses.squeeze(), hwf.squeeze(), bound.squeeze()
meta_optim.zero_grad()
inner_model = copy.deepcopy(meta_model)
inner_optim = torch.optim.SGD(inner_model.parameters(), args.inner_lr)
inner_loop(inner_model, inner_optim, imgs, poses,
hwf, bound, args.num_samples,
args.train_batchsize, args.inner_steps)
with torch.no_grad():
for meta_param, inner_param in zip(meta_model.parameters(), inner_model.parameters()):
meta_param.grad = meta_param - inner_param
meta_optim.step()
def report_result(model, imgs, poses, hwf, bound, num_samples, raybatch_size):
"""
report view-synthesis result on heldout views
"""
ray_origins, ray_directions = get_rays_shapenet(hwf, poses)
view_psnrs = []
for img, rays_o, rays_d in zip(imgs, ray_origins, ray_directions):
rays_o, rays_d = rays_o.reshape(-1, 3), rays_d.reshape(-1, 3)
t_vals, xyz = sample_points(rays_o, rays_d, bound[0], bound[1],
num_samples, perturb=False)
synth = []
num_rays = rays_d.shape[0]
with torch.no_grad():
for i in range(0, num_rays, raybatch_size):
rgbs_batch, sigmas_batch = model(xyz[i:i+raybatch_size])
color_batch = volume_render(rgbs_batch, sigmas_batch,
t_vals[i:i+raybatch_size],
white_bkgd=True)
synth.append(color_batch)
synth = torch.cat(synth, dim=0).reshape_as(img)
error = F.mse_loss(img, synth)
psnr = -10*torch.log10(error)
view_psnrs.append(psnr)
scene_psnr = torch.stack(view_psnrs).mean()
return scene_psnr
def val_meta(args, model, val_loader, device):
"""
validate the meta trained model for few-shot view synthesis
"""
meta_trained_state = model.state_dict()
val_model = copy.deepcopy(model)
val_psnrs = []
for imgs, poses, hwf, bound in val_loader:
imgs, poses, hwf, bound = imgs.to(device), poses.to(device), hwf.to(device), bound.to(device)
imgs, poses, hwf, bound = imgs.squeeze(), poses.squeeze(), hwf.squeeze(), bound.squeeze()
tto_imgs, test_imgs = torch.split(imgs, [args.tto_views, args.test_views], dim=0)
tto_poses, test_poses = torch.split(poses, [args.tto_views, args.test_views], dim=0)
val_model.load_state_dict(meta_trained_state)
val_optim = torch.optim.SGD(val_model.parameters(), args.tto_lr)
inner_loop(val_model, val_optim, tto_imgs, tto_poses, hwf,
bound, args.num_samples, args.tto_batchsize, args.tto_steps)
scene_psnr = report_result(val_model, test_imgs, test_poses, hwf, bound,
args.num_samples, args.test_batchsize)
val_psnrs.append(scene_psnr)
val_psnr = torch.stack(val_psnrs).mean()
return val_psnr
def main():
parser = argparse.ArgumentParser(description='shapenet few-shot view synthesis')
parser.add_argument('--config', type=str, required=True,
help='config file for the shape class (cars, chairs or lamps)')
args = parser.parse_args()
with open(args.config) as config:
info = json.load(config)
for key, value in info.items():
args.__dict__[key] = value
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_set = build_shapenet(image_set="train", dataset_root=args.dataset_root,
splits_path=args.splits_path, num_views=args.train_views)
train_loader = DataLoader(train_set, batch_size=1, shuffle=True)
val_set = build_shapenet(image_set="val", dataset_root=args.dataset_root,
splits_path=args.splits_path,
num_views=args.tto_views+args.test_views)
val_loader = DataLoader(val_set, batch_size=1, shuffle=False)
meta_model = build_nerf(args)
meta_model.to(device)
meta_optim = torch.optim.Adam(meta_model.parameters(), lr=args.meta_lr)
for epoch in range(1, args.meta_epochs+1):
train_meta(args, meta_model, meta_optim, train_loader, device)
val_psnr = val_meta(args, meta_model, val_loader, device)
print(f"Epoch: {epoch}, val psnr: {val_psnr:0.3f}")
torch.save({
'epoch': epoch,
'meta_model_state_dict': meta_model.state_dict(),
'meta_optim_state_dict': meta_optim.state_dict(),
}, f'meta_epoch{epoch}.pth')
if __name__ == '__main__':
main()