-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdetector.py
203 lines (177 loc) · 8.27 KB
/
detector.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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
#coding=utf-8
import os
import math
import tqdm
import torch
import numpy as np
from torch import nn
import torchvision as tv
from sklearn import metrics
import torchvision.utils as vutils
from torch.nn import functional as F
from utils import visualization
from dataset import augmentations
from utils.losses import CountLoss
from utils.lr_scheduler import LR_Scheduler
class Detector(object):
def __init__(self, net, train_loader=None, test_loader=None, batch_size=None,
optimizer='adam', lr=1e-3, patience=5, interval=1, num_classes=1, cov=1,
checkpoint_dir='saved_models', checkpoint_name='', devices=[0], log_size=(96, 96)):
self.train_loader = train_loader
self.test_loader = test_loader
self.lr = lr
self.batch_size = batch_size
self.patience = patience
self.interval = interval
self.checkpoint_dir = checkpoint_dir
self.checkpoint_name = checkpoint_name
self.scale = cov * math.pi * 2
self.num_classes = num_classes
self.log_size = log_size
if not os.path.exists(checkpoint_dir):
os.mkdir(checkpoint_dir)
self.net_single = net
self.criterion = CountLoss(self.scale)
if len(devices) == 0:
self.device = torch.device('cpu')
elif len(devices) == 1:
self.device = torch.device('cuda')
self.net = self.net_single.to(self.device)
self.criterion = self.criterion.to(self.device)
else:
self.device = torch.device('cuda')
# torch.distributed.init_process_group(backend='nccl', init_method='env://')
# self.net = nn.parallel.DistributedDataParallel(self.net_single)
self.net = nn.DataParallel(self.net_single, device_ids=range(len(devices))).to(self.device)
self.criterion = nn.DataParallel(self.criterion, device_ids=range(len(devices))).to(self.device)
train_params = [{'params': self.net_single.get_1x_lr_params(), 'lr': lr},
{'params': self.net_single.get_10x_lr_params(), 'lr': lr * 10}]
if optimizer == 'sgd':
self.opt = torch.optim.SGD(
train_params, lr=lr, weight_decay=5e-4, momentum=0.9)
elif optimizer == 'adam':
self.opt = torch.optim.Adam(
train_params, lr=lr, weight_decay=5e-4)
else:
raise Exception('Optimizer {} Not Exists'.format(optimizer))
def reset_grad(self):
self.opt.zero_grad()
def train(self, max_epoch, writer=None, epoch_size=100):
max_step = epoch_size * max_epoch
scheduler = LR_Scheduler('poly', self.lr, max_epoch, epoch_size)
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(self.opt, max_epoch * epoch_size)
torch.cuda.manual_seed(1)
best_score = 0
step = 0
for epoch in tqdm.tqdm(range(max_epoch), total=max_epoch):
torch.cuda.empty_cache()
self.net.train()
for batch_idx, data in enumerate(self.train_loader):
img = data['image'].to(self.device)
hm = data['heatmap'].to(self.device)
mask = data['mask'].to(self.device)
num = data['num'].to(self.device)
scheduler(self.opt, batch_idx, epoch, best_score)
self.reset_grad()
pred_hm, pred_mask = self.net(img)
rate = math.exp(-step / (max_step / 10))
loss = self.get_loss((pred_hm, pred_mask), (hm, mask, num), rate, backward=False)
loss.backward()
self.opt.step()
if writer:
writer.add_scalar(
'rate', rate, global_step=step)
writer.add_scalar(
'loss', loss.data, global_step=step)
writer.add_scalar(
'lr', self.opt.param_groups[0]['lr'], global_step=step)
step += 1
# scheduler.step(step)
if epoch % self.interval == 0:
torch.cuda.empty_cache()
acc, imgs, pred_hms, gt_hms, pred_masks, gt_masks = self.test()
if writer:
writer.add_scalar(
'Acc', acc, global_step=epoch)
score = acc
pred_hms = self.draw_heatmap(imgs, pred_hms)
writer.add_image('Pred HM', pred_hms, epoch)
gt_hms = self.draw_heatmap(imgs, gt_hms)
writer.add_image('GT HM', gt_hms, epoch)
pred_masks = self.draw_mask(imgs, pred_masks)
writer.add_image('Pred Mask', pred_masks, epoch)
gt_masks = self.draw_mask(imgs, gt_masks, is_gt=True)
writer.add_image('GT Mask', gt_masks, epoch)
if best_score <= score + 0.01:
best_score = score
self.save_model(self.checkpoint_dir)
def test(self):
self.net.eval()
with torch.no_grad():
imgs = []
gt_hms = []
pred_hms = []
gt_masks = []
pred_masks = []
acc = 0
count = 0
for batch_idx, data in enumerate(self.test_loader):
img = data['image'].to(self.device)
hm = data['heatmap']
mask = data['mask']
num = data['num']
pred_hm, pred_mask = self.net(img)
pred_hm = pred_hm.detach().cpu()
pred_mask = pred_mask.detach().cpu()
pred_num = torch.round(pred_hm.sum(-1).sum(-1) / self.scale)
acc += (pred_num.type(torch.int64) == num.type(torch.int64)).type(torch.float32).mean()
img = img.cpu()
imgs.append(img)
pred_hms.append(pred_hm)
gt_hms.append(hm)
pred_masks.append(pred_mask)
gt_masks.append(mask)
count += img.shape[0]
if count >= 40:
break
gt_hms = torch.cat(gt_hms)
pred_hms = torch.cat(pred_hms)
gt_masks = torch.cat(gt_masks)
pred_masks = torch.cat(pred_masks)
imgs = torch.cat(imgs)
acc /= batch_idx + 1
return acc, imgs[: 40], pred_hms[: 40], gt_hms[: 40], pred_masks[: 40], gt_masks[: 40]
def draw_heatmap(self, img, hm, size=None):
if size is None:
size = self.log_size
img = F.interpolate(img, size, mode='bilinear', align_corners=True)
img = vutils.make_grid(img).numpy()
rgb = visualization.heatmap_to_rgb(hm, self.num_classes, size)
result = np.clip((rgb + img) / 2, 0, 1)
return result
def draw_mask(self, img, mask, size=None, is_gt=False):
if size is None:
size = self.log_size
img = F.interpolate(img, size, mode='bilinear', align_corners=True)
img = vutils.make_grid(img).numpy()
rgb = visualization.mask_to_rgb(mask, self.num_classes, size, is_gt=is_gt)
result = np.clip((rgb + img) / 2, 0, 1)
return result
def save_model(self, checkpoint_dir, comment=None):
if comment is None:
torch.save(self.net_single, '{}/best_model_{}.pt'.format(checkpoint_dir, self.checkpoint_name))
else:
torch.save(self.net_single, '{}/best_model_{}_{}.pt'.format(checkpoint_dir, self.checkpoint_name, comment))
def load_model(self, model_path):
self.net_single.load_state_dict(torch.load(model_path).state_dict())
def predict(self, img):
x = torch.from_numpy(img).type(torch.float32).permute(0, 3, 1, 2).to(self.device) / 255
self.net.eval()
with torch.no_grad():
pred_hm, pred_mask = self.net(x)
pred_hm = pred_hm.detach().cpu().numpy()
pred_mask = pred_mask.detach().cpu().numpy()
return pred_hm, pred_mask
def get_loss(self, pred, target, rate, backward=False):
loss = self.criterion(pred, target, rate, backward)
return loss.mean()