-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_test_RAGRNet.py
187 lines (161 loc) · 8.39 KB
/
train_test_RAGRNet.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
import torch
import torch.nn.functional as F
from torch.autograd import Variable
import imageio
import numpy as np
import os, argparse
from datetime import datetime
from skimage import img_as_ubyte
from model.RAGRNet import RAGRNet
from utils.data import get_loader
from utils.data import test_dataset
from utils.func import label_edge_prediction, AvgMeter, clip_gradient, adjust_lr
import time
import pytorch_iou
import pytorch_fm
torch.cuda.set_device(0)
CE = torch.nn.BCEWithLogitsLoss()
IOU = pytorch_iou.IOU(size_average=True)
floss = pytorch_fm.FLoss()
size_rates = [0.75, 1, 1.25]
def _get_adaptive_threshold(matrix, max_value = 1):
"""
Return an adaptive threshold, which is equal to twice the mean of ``matrix``.
:param matrix: a data array
:param max_value: the upper limit of the threshold
:return: min(2 * matrix.mean(), max_value)
"""
return min(2 * matrix.mean(), max_value)
def cal_adaptive_fm(pred, gt):
"""
Calculate the adaptive F-measure.
:return: adaptive_fm
"""
# ``np.count_nonzero`` is faster and better
beta = 0.3
adaptive_threshold = _get_adaptive_threshold(pred, max_value=1)
binary_predcition = (pred >= adaptive_threshold).astype(np.float32)
area_intersection = np.count_nonzero(binary_predcition * gt)
if area_intersection == 0:
adaptive_fm = 0
else:
pre = area_intersection * 1.0 / np.count_nonzero(binary_predcition)
rec = area_intersection * 1.0 / np.count_nonzero(gt)
adaptive_fm = (1 + beta) * pre * rec / (beta * pre + rec)
return adaptive_fm
def run(train_i):
best_adp_fm = 0
best_mae = 1
best_epoch = 0
for epoch in range(1, opt.epoch):
adjust_lr(optimizer, opt.lr, epoch, opt.decay_rate, opt.decay_epoch)
model.train()
loss_record1, loss_record2, loss_record3, loss_record4, loss_record5, loss_record6 = AvgMeter(), AvgMeter(), AvgMeter(), AvgMeter(), AvgMeter(), AvgMeter()
for i, pack in enumerate(train_loader, start=1):
for rate in size_rates:
optimizer.zero_grad()
images, gts = pack
images = Variable(images).cuda()
gts = Variable(gts).cuda()
# edge prediction
edges = label_edge_prediction(gts)
# multi-scale training samples
trainsize = int(round(opt.trainsize * rate / 32) * 32)
if rate != 1:
images = F.interpolate(images, size=(trainsize, trainsize), mode='bilinear', align_corners=True)
gts = F.interpolate(gts, size=(trainsize, trainsize), mode='bilinear', align_corners=True)
edges = F.interpolate(edges, size=(trainsize, trainsize), mode='bilinear', align_corners=True)
s1, s1_sig, edg1, s2, s2_sig, edg2, s3, s3_sig, edg3, s4, s4_sig, edg4, edgesal, graphsal, graphsal_sig, g1sal, g1sal_sig = model(images)
loss1 = CE(s1, gts) + IOU(s1_sig, gts) + floss(s1_sig, gts) + CE(edg1, edges)
loss2 = CE(s2, gts) + IOU(s2_sig, gts) + floss(s2_sig, gts) + CE(edg2, edges)
loss3 = CE(s3, gts) + IOU(s3_sig, gts) + floss(s3_sig, gts) + CE(edg3, edges)
loss4 = CE(s4, gts) + IOU(s4_sig, gts) + floss(s4_sig, gts) + CE(edg4, edges)
loss5 = CE(graphsal, gts) + IOU(graphsal_sig, gts) + floss(graphsal_sig, gts) + CE(edgesal, edges)
loss6 = CE(g1sal, gts) + IOU(g1sal_sig, gts) + floss(g1sal_sig, gts)
loss = loss1 + loss2 + loss3 + loss4 + loss5 + loss6
loss.backward()
clip_gradient(optimizer, opt.clip)
optimizer.step()
if rate == 1:
loss_record1.update(loss1.data, opt.batchsize)
loss_record2.update(loss2.data, opt.batchsize)
loss_record3.update(loss3.data, opt.batchsize)
loss_record4.update(loss4.data, opt.batchsize)
loss_record5.update(loss5.data, opt.batchsize)
loss_record6.update(loss6.data, opt.batchsize)
if i % 100 == 0 or i == total_step:
print(
'{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], Learning Rate: {}, Loss: {:.4f}, Loss1: {:.4f}, Loss2: {:.4f}, Lossgraph: {:.4f}'.
format(datetime.now(), epoch, opt.epoch, i, total_step,
opt.lr * opt.decay_rate ** (epoch // opt.decay_epoch), loss.data, loss1.data,
loss2.data, loss5.data))
save_path = 'models/Your_Files/'
if not os.path.exists(save_path):
os.makedirs(save_path)
res_save_path = save_path + 'salmap/'
if not os.path.exists(res_save_path):
os.makedirs(res_save_path)
# test
with torch.no_grad():
model.eval()
time_sum = 0
adaptive_fms = 0.0
mae = 0.0
for i in range(test_loader.size):
image, gt, name = test_loader.load_data()
gt = np.asarray(gt, np.float32)
gt /= (gt.max() + 1e-8)
image = image.cuda()
time_start = time.time()
res, s1_sig, e1, s2, s2_sig, e2, s3, s3_sig, e3, s4, s4_sig, e4, edgesal, graphsal, graphsal_sig, g1sal, g1sal_sig = model(image)
time_end = time.time()
time_sum = time_sum + (time_end - time_start)
res = F.upsample(res, size=gt.shape, mode='bilinear', align_corners=False)
res = res.sigmoid().data.cpu().numpy().squeeze()
res = (res - res.min()) / (res.max() - res.min() + 1e-8)
adaptive_fm = cal_adaptive_fm(pred=res, gt=gt)
# adaptive_fm = 0
adaptive_fms += adaptive_fm
mae += np.sum(np.abs(gt - res)) / (gt.shape[0] * gt.shape[1])
imageio.imsave(res_save_path + name, img_as_ubyte(res))
print('FPS {:.5f}'.format(test_loader.size / time_sum))
torch.save(model.state_dict(), save_path + 'SEI_ORSSD.pth' + '.%d' % epoch)
adp_fm = adaptive_fms / test_loader.size
mae_mean = mae / test_loader.size
if mae_mean < best_mae:
best_adp_fm = adp_fm
best_mae = mae_mean
best_epoch = epoch
print('Epoch [{:03d}], best_adp_fm {:.4f}, best_mae {:.4f}'.format(epoch, best_adp_fm, best_mae))
print('Current_epoch [{:03d}], adp_fm {:.4f}, mae {:.4f}'.format(epoch, adp_fm, mae_mean))
print('Best_epoch [{:03d}], best_adp_fm {:.4f}, best_mae {:.4f}'.format(best_epoch, best_adp_fm, best_mae))
print("Let's go!")
for train_i in range(0, 1):
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int, default=50, help='epoch number')
parser.add_argument('--lr', type=float, default=5e-5, help='learning rate')
parser.add_argument('--batchsize', type=int, default=8, help='training batch size')
parser.add_argument('--trainsize', type=int, default=256, help='training dataset size')
parser.add_argument('--clip', type=float, default=0.5, help='gradient clipping margin')
parser.add_argument('--decay_rate', type=float, default=0.1, help='decay rate of learning rate')
parser.add_argument('--decay_epoch', type=int, default=30, help='every n epochs decay learning rate')
opt = parser.parse_args()
# build train_models
model = RAGRNet()
model.cuda()
params = model.parameters()
optimizer = torch.optim.Adam(params, opt.lr)
image_root = '/Your_File/SOD/train_dataset/EORSSD/Images/'
gt_root = '/Your_File/SOD/train_dataset/EORSSD/GT/'
train_loader = get_loader(image_root, gt_root, batchsize=opt.batchsize, trainsize=opt.trainsize)
total_step = len(train_loader)
# build test_models
test_dataset_path = '/Your_File/SOD/test_dataset/'
# test_datasets = 'ORSSD'
test_datasets = 'EORSSD'
test_image_root = test_dataset_path + test_datasets + '/Images/'
test_gt_root = test_dataset_path + test_datasets + '/GT/'
test_loader = test_dataset(test_image_root, test_gt_root, opt.trainsize)
print('Statr {}-th training!!!'.format(train_i))
print('Learning Rate: {}'.format(opt.lr))
run(train_i)