-
Notifications
You must be signed in to change notification settings - Fork 11
/
evaluation.py
186 lines (138 loc) · 7.54 KB
/
evaluation.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
import os
n_thread = 1
os.environ["MKL_NUM_THREADS"] = f"{n_thread}"
os.environ["NUMEXPR_NUM_THREADS"] = f"{n_thread}"
os.environ["OMP_NUM_THREADS"] = f"4"
os.environ["VECLIB_MAXIMUM_THREADS"] = f"{n_thread}"
os.environ["OPENBLAS_NUM_THREADS"] = f"{n_thread}"
from omegaconf import OmegaConf
import os, torch, math, imageio, cv2
import torch.nn as nn
import numpy as np
from tqdm import tqdm
import sys,json
from lightning.system import system
from torch.utils.data import DataLoader
import pytorch_lightning as L
from dataLoader import dataset_dict
from pytorch_msssim import ssim
from tools.gen_video_path import uni_video_path,uni_mesh_path
import lpips
import torch.nn.functional as F
from tools.depth import acc_threshold,abs_error
@torch.no_grad()
def main(cfg):
torch.set_float32_matmul_precision('medium')
# data loader
dataset = dataset_dict[cfg.infer.dataset.dataset_name]
loader = DataLoader(dataset(cfg.infer.dataset),
batch_size=cfg.infer.dataset.batch_size,
num_workers=cfg.infer.dataset.num_workers,
shuffle=False,
pin_memory=False)
loader_iter = iter(loader)
device = 'cuda'
my_system = system.load_from_checkpoint(cfg.infer.ckpt_path, cfg=cfg, map_location=device)
# metrics
lpips_vgg_fun = lpips.LPIPS(net='vgg').to(device)
lpips_alex_fun = lpips.LPIPS(net='alex').to(device)
names, depth_accs = [], []
psnrs,ssims, lpips_vggs, lpips_alexs = [],[],[],[]
os.makedirs(cfg.infer.save_folder, exist_ok=True)
for i in tqdm(range(len(loader))):#len(loader)
sample = next(loader_iter)
sample = {key: tensor.to(device) if torch.is_tensor(tensor) else tensor for key, tensor in sample.items()}
my_system.net.eval()
return_buffer = cfg.infer.video_frames > 0 or cfg.infer.save_mesh
output = my_system.net(sample, with_fine=True, return_buffer=return_buffer)
name = sample['meta']['scene'][0].split('.')[0]
images = output['image_fine'][0]
img_gt = sample['tar_rgb'][0].permute(1,0,2,3).reshape(images.shape)
alpha = output['acc_map'][0][...,None]
normal_white = ((output['rend_normal_fine'][0]*alpha+1-alpha) + 1)/2
n_view = cfg.n_views
if i<100:
cv2.imwrite(os.path.join(cfg.infer.save_folder, name + '.jpg'), torch.cat((img_gt,images,normal_white),dim=0).detach().cpu().numpy()[...,::-1]*255)
if cfg.infer.eval_novel_view_only:
width = sample['meta']['tar_w']
images = images.permute(2,0,1)[None][...,width*n_view:]
img_gt = img_gt.permute(2,0,1)[None][...,width*n_view:]
else:
images = images.permute(2,0,1)[None]
img_gt = img_gt.permute(2,0,1)[None]
if images.shape[-1] > 0:
color_loss_all = (images-img_gt)**2
psnr = -10. * torch.log(color_loss_all.mean()) / torch.log(torch.tensor([10.]).to(device))
ssim_val = ssim(images, img_gt, data_range=1.0, size_average=False)
lpips_vgg = lpips_vgg_fun(img_gt*2-1,images*2-1)
lpips_alex = lpips_alex_fun(img_gt*2-1,images*2-1)
psnrs.append(psnr.item())
ssims.append(ssim_val.item())
lpips_vggs.append(lpips_vgg.item())
lpips_alexs.append(lpips_alex.item())
if len(cfg.infer.eval_depth):
B,N,H,W = sample['tar_msk'].shape
mask = sample['tar_msk'].permute(0,2,1,3).reshape(B,H,N*W)
mask = mask.cpu().detach().bool().numpy()
depth_gt = sample['tar_dep'].permute(0,2,1,3).reshape(B,H,N*W)
depth_gt = depth_gt.cpu().detach().numpy()
depth_pred = output['depth_fine'].cpu().squeeze(-1).detach().numpy()
depth_acc = []
errors = abs_error(depth_pred, depth_gt, mask).mean().item()
depth_acc.append(errors)
for threshold in cfg.infer.eval_depth:
depth_acc.append(acc_threshold(depth_pred, depth_gt, mask, threshold=threshold).mean())
depth_accs.append(depth_acc)
names.append(name)
fov = [sample['fovx'],sample['fovy']]
if cfg.infer.video_frames > 0:
cams = uni_video_path(cfg.infer.video_frames, cfg.infer.dataset, sample, fov=fov)
gs_params = output['render_pkg'][1] # fine ouputs
_centers, _shs, _opacity, _scaling, _rotation, mask = gs_params
imgs,normal_blks,normal_whites = [],[],[]
for cam in cams:
cam.to_device(device)
rays = cam.get_rays().to(device)
output_img = my_system.net.gs_render.render_img(cam, rays, _centers, _shs, _opacity[mask], _scaling[mask], _rotation[mask], device)
img, normal = output_img['image'], output_img['rend_normal']
img = np.round(img.cpu().detach().numpy()*255).astype('uint8')
alpha = output_img['acc_map'][...,None]
normal_white = np.round((((normal*alpha+1-alpha) + 1)/2).cpu().detach().numpy()*255).astype('uint8')
imgs.append(img)
normal_whites.append(normal_white)
imageio.mimwrite(f'{cfg.infer.save_folder}/{name}.mp4', imgs, fps=30, quality=10)
imageio.mimwrite(f'{cfg.infer.save_folder}/{name}_nrm.mp4', normal_whites, fps=30, quality=10)
if cfg.infer.save_mesh:
from tools.meshExtractor import MeshExtractor
aabb = cfg.infer.aabb
gs_params = output['render_pkg'][1] # fine ouputs
meshExtractor = MeshExtractor(gs_params, my_system.net.gs_render, aabb=aabb)
meshExtractor.extract(f'{cfg.infer.save_folder}/{name}.obj', cfg.infer.dataset, sample=sample,fov=fov)
if cfg.infer.mesh_video_frames > 0:
from tools.meshRender import render_mesh
cams = uni_video_path(cfg.infer.video_frames, cfg.infer.dataset, sample=sample, fov=fov)
mesh_imgs = render_mesh(cams, f'{cfg.infer.save_folder}/{name}.obj')[...,:3]
imageio.mimwrite(f'{cfg.infer.save_folder}/{name}_mesh.mp4', mesh_imgs, fps=30, quality=10)
del sample
if len(cfg.infer.eval_depth):
mean_depth_acc = np.mean( np.stack(depth_accs),axis=0).tolist()
else:
mean_depth_acc = 0.0
if len(psnrs) and cfg.infer.metric_path is not None:
print(f'evaluation score, psnr: {np.mean(psnrs)} ssim: {np.mean(ssims)}, lpips_vgg:{np.mean(lpips_vggs)}, lpips_alex: {np.mean(lpips_alexs)}, depth_acc:{mean_depth_acc}')
scores = {'name':names, 'psnr':psnrs, 'ssim':ssims, \
'lpips_vgg':lpips_vggs,'lpips_alex':lpips_alexs, \
'depth_acc': depth_accs}
scores.update({'psnr_mean':np.mean(psnrs), 'ssim_mean':np.mean(ssims),
'lpips_vgg_mean':np.mean(lpips_vggs),'lpips_alex_mean':np.mean(lpips_alexs),
'depth_acc': mean_depth_acc})
os.makedirs(os.path.dirname(cfg.infer.metric_path), exist_ok=True)
with open(cfg.infer.metric_path, 'w') as f:
json.dump(scores, f, indent=4)
if __name__ == '__main__':
base_conf = OmegaConf.load('configs/base.yaml')
path_config = sys.argv[1]
cli_conf = OmegaConf.from_cli()
second_conf = OmegaConf.load(path_config)
cfg = OmegaConf.merge(base_conf, second_conf, cli_conf)
main(cfg)