-
Notifications
You must be signed in to change notification settings - Fork 7
/
syn_stage1.py
137 lines (123 loc) · 5.55 KB
/
syn_stage1.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
# coding=utf-8
import pdb
import torch
import torch.nn.functional as F
import torchvision
from tqdm import tqdm
from PIL import Image
import numpy as np
from datasets import VOC, Saliency
from datasets import palette as palette_voc
from evaluate_seg import evaluate_iou
from evaluate_sal import fm_and_mae
import json
import os
from jls_fcn import JLSFCN
from logger import Logger
import pydensecrf.densecrf as dcrf
from pydensecrf.utils import unary_from_softmax, unary_from_labels
from multiprocessing import Pool
image_size = 256
batch_size = 8
c_output = 21
path_save_checkpoints = "stage1.pth"
path_save_train_voc = "voc_train_pred_debug"
path_save_train_voc_prob = "voc_train_pred_debug_prob"
path_save_train_voc_crf = "voc_train_pred_debug_softcrf"
voc_save_gt_dir = "voc_train_pred_debug_gt"
voc_save_img_dir = "voc_train_pred_debug_img"
if not os.path.exists(path_save_train_voc): os.mkdir(path_save_train_voc)
if not os.path.exists(path_save_train_voc_prob): os.mkdir(path_save_train_voc_prob)
if not os.path.exists(path_save_train_voc_crf): os.mkdir(path_save_train_voc_crf)
if not os.path.exists(voc_save_gt_dir): os.mkdir(voc_save_gt_dir)
if not os.path.exists(voc_save_img_dir): os.mkdir(voc_save_img_dir)
net = JLSFCN(c_output).cuda()
net.load_state_dict(torch.load(path_save_checkpoints))
mean = torch.Tensor([0.485, 0.456, 0.406])[None, ..., None, None].cuda()
std = torch.Tensor([0.229, 0.224, 0.225])[None, ..., None, None].cuda()
voc_train_img_dir = '/home/zeng/data/datasets/segmentation/VOCdevkit/VOC2012/JPEGImages'
voc_train_gt_dir = '/home/zeng/data/datasets/segmentation/VOCdevkit/VOC2012/SegmentationClassAug'
voc_train_split = '/home/zeng/data/datasets/segmentation/VOCdevkit/VOC2012/ImageSets/Segmentation/trainaug.txt'
voc_loader = torch.utils.data.DataLoader(
VOC(voc_train_img_dir, voc_train_gt_dir, voc_train_split,
crop=0.9, flip=True, rotate=10, size=image_size, training=False, tproc=True),
batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
def val_voc():
net.eval()
with torch.no_grad():
for t in range(10):
for it, (img, gt, batch_name, WW, HH) in tqdm(enumerate(voc_loader), desc='train'):
gt_cls = gt[:, None, ...] == torch.arange(c_output)[None, ..., None, None]
gt_cls = (gt_cls.sum(3, keepdim=True).sum(2, keepdim=True)>0).float().cuda()
raw_img = img
img = (img.cuda()-mean)/std
batch_seg, _, seg32x = net(img)
batch_seg[:, 1:] *= gt_cls[:, 1:]
batch_prob = F.softmax(batch_seg, 1)
_, batch_seg = batch_seg.detach().max(1)
for n, name in enumerate(batch_name):
msk = batch_seg[n]
prob = batch_prob[n]
prob = prob.detach().cpu().numpy()
np.save('{}/{}_{}.npy'.format(path_save_train_voc_prob, name, t), prob)
_img = raw_img[n]*255
_img = _img.numpy().astype(np.uint8)
_img = _img.transpose((1,2,0))
_img = Image.fromarray(_img)
_img.save('{}/{}_{}.jpg'.format(voc_save_img_dir, name, t))
_gt = gt[n]
_gt = _gt.numpy()
_gt = _gt.astype(np.uint8)
_gt = Image.fromarray(_gt.astype(np.uint8))
_gt = _gt.convert('P')
_gt.putpalette(palette_voc)
_gt.save('{}/{}_{}.png'.format(voc_save_gt_dir, name, t), 'PNG')
msk = msk.detach().cpu().numpy()
msk = Image.fromarray(msk.astype(np.uint8))
msk = msk.convert('P')
msk.putpalette(palette_voc)
msk.save('{}/{}_{}.png'.format(path_save_train_voc, name, t), 'PNG')
miou = evaluate_iou(path_save_train_voc, voc_save_gt_dir, c_output)
net.train()
return miou
def one_dcrf(name):
_name = ".".join(name.split(".")[:-1])
img = Image.open(os.path.join(voc_save_img_dir, _name+".jpg")).convert("RGB")
w,h = img.size
img = np.array(img)
prob = np.load(os.path.join(path_save_train_voc_prob, name))
U = unary_from_softmax(prob)
#msk = Image.open(os.path.join(path_save_train_voc, _name+".png")).convert("P")
#msk = np.array(msk)
#msk = msk+1
#U = unary_from_labels(msk, c_output, gt_prob=0.7)
d = dcrf.DenseCRF2D(w, h, c_output)
d.setUnaryEnergy(U)
# This adds the color-independent term, features are the locations only.
d.addPairwiseGaussian(sxy=(3, 3), compat=3, kernel=dcrf.DIAG_KERNEL,
normalization=dcrf.NORMALIZE_SYMMETRIC)
# This adds the color-dependent term, i.e. features are (x,y,r,g,b).
d.addPairwiseBilateral(sxy=(80, 80), srgb=(13, 13, 13), rgbim=img,
compat=10,
kernel=dcrf.DIAG_KERNEL,
normalization=dcrf.NORMALIZE_SYMMETRIC)
Q = d.inference(5)
msk = np.argmax(Q, axis=0).reshape((h,w))
msk = msk
msk = Image.fromarray(msk.astype(np.uint8))
msk = msk.convert('P')
msk.putpalette(palette_voc)
msk.save('{}/{}.png'.format(path_save_train_voc_crf, _name), 'PNG')
print(name)
def proc_dcrf():
names = os.listdir(path_save_train_voc_prob)
pool = Pool(4)
pool.map(one_dcrf, names)
#for name in tqdm(names):
# one_dcrf(name)
miou = evaluate_iou(path_save_train_voc_crf, voc_save_gt_dir, c_output)
return miou
if __name__ == "__main__":
miou = val_voc()
miou = proc_dcrf()
# print(miou)