forked from cvlab-stonybrook/PaperEdge
-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
110 lines (98 loc) · 4.03 KB
/
eval.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
import argparse
import json
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import cv2
cv2.setNumThreads(0)
cv2.ocl.setUseOpenCL(False)
from visdom import Visdom
import matplotlib.pyplot as plt
vis = Visdom(port=10086)
import os
from pathlib import Path
# parse the experiment configuration
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True, type=str, help="evaluation configuration files.")
exp_config = parser.parse_args().config
with open('configs/' + exp_config, 'r') as fid:
args = json.load(fid)
from data.benchmark import DocUNet130
d130 = DocUNet130(root_dir=args['docunet_benchmark_img_dir'])
from torch.utils.data import DataLoader
dloader = DataLoader(d130, batch_size=64, num_workers=2, pin_memory=True)
from networks.paperedge import GlobalWarper, LocalWarper, WarperUtil
netG = GlobalWarper().to('cuda')
# path to the Enet checkpoint
netG.load_state_dict(torch.load(args['Enet_ckpt'])['G'])
# netG.load_state_dict(torch.load('G_w_checkpoint_13820.pt')['G'])
netG.eval()
netL = LocalWarper().to('cuda')
# path to the Tnet checkpoint
netL.load_state_dict(torch.load('Tnet_ckpt')['L'])
# netL.load_state_dict(torch.load('L_w_checkpoint_27640.pt')['L'])
netL.eval()
warpUtil = WarperUtil(64).to('cuda')
with torch.no_grad():
res = []
res0 = []
inx = []
gs = []
ls = []
for x in dloader:
inx.append(x)
x = x.to('cuda')
# global
d = netG(x)
d = warpUtil.global_post_warp(d, 64)
gs.append(d)
d = F.interpolate(d, size=256, mode='bilinear', align_corners=True)
y0 = F.grid_sample(x, d.permute(0, 2, 3, 1), align_corners=True)
# y0 = x.clone()
# change `nit` to a number > 1 to test iteratively applying Tnet to refine the unwarping result.
# seems to improve the result slightly. Needs more investigation. Not reported in the paper.
nit = 1
y = y0.clone()
for it in range(nit):
d = netL(y)
if it > 0:
d = F.grid_sample(pd, d.permute(0, 2, 3, 1), padding_mode='reflection', align_corners=True)
pd = d
d = F.interpolate(d, size=256, mode='bilinear', align_corners=True)
y = F.grid_sample(y0, d.permute(0, 2, 3, 1), align_corners=True)
# d = netL(y)
d = F.interpolate(d, size=256, mode='bilinear', align_corners=True)
d = d.clamp(-1.0, 1.0)
y = F.grid_sample(y0, d.permute(0, 2, 3, 1), align_corners=True)
res.append(y)
res0.append(y0)
ls.append(d)
res0 = torch.cat(res0, dim=0)
res = torch.cat(res, dim=0)
inx = torch.cat(inx, dim=0)
gs = torch.cat(gs, dim=0)
ls = torch.cat(ls, dim=0)
vis.images(inx.clamp(1e-3, 1 - 1e-3).detach().cpu().numpy(), env='benchmark', win='inx', opts={'caption': 'inx'})
vis.images(res.clamp(1e-3, 1 - 1e-3).detach().cpu().numpy(), env='benchmark', win='res', opts={'caption': 'res'})
vis.images(res0.clamp(1e-3, 1 - 1e-3).detach().cpu().numpy(), env='benchmark', win='res0', opts={'caption': 'res0'})
# replace the following two lines with the benchmark image path and target result path
img_dir = args['docunet_benchmark_img_dir']
dst_dir = args['dst_dir']
Path(dst_dir).mkdir(parents=True, exist_ok=True)
ct = 0
for k in range(1, 66):
print(k)
for m in range(1, 3):
im = cv2.imread(os.path.join(img_dir, f'{k}_{m} copy.png')).astype(np.float32) / 255.0
im = torch.from_numpy(np.transpose(im, (2, 0, 1)))
im = im.to('cuda').unsqueeze(0)
d = gs[ct : ct + 1, ...]
d = F.interpolate(d, (im.size(2), im.size(3)), mode='bilinear', align_corners=True)
y = F.grid_sample(im, d.permute(0, 2, 3, 1), align_corners=True).detach()
d = ls[ct : ct + 1, ...]
d = F.interpolate(d, (im.size(2), im.size(3)), mode='bilinear', align_corners=True)
y = F.grid_sample(y, d.permute(0, 2, 3, 1), align_corners=True).detach()
y = y.squeeze().permute(1, 2, 0).cpu().numpy()
cv2.imwrite(f'{dst_dir}/{k}_{m}.png', y*255.)
ct += 1