-
Notifications
You must be signed in to change notification settings - Fork 31
/
Copy pathtrainer.py
159 lines (125 loc) · 5.61 KB
/
trainer.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
import os
import os.path as osp
import time
import torch
import datetime
import torch.nn as nn
from torchvision.utils import save_image
import numpy as np
import torch.nn.functional as F
from verifier import Verifier
from networks import get_model
from utils import *
from criterion import *
from lr_scheduler import WarmupPolyLR
from torch.utils.tensorboard import SummaryWriter
__BA__ = ["CE2P", "FaceParseNet18", "FaceParseNet34", "FaceParseNet50", "FaceParseNet"]
class Trainer(object):
"""Training pipline"""
def __init__(self, data_loader, config, val_loader):
# Data loader
self.data_loader = data_loader
self.verifier = Verifier(val_loader, config)
self.writer = SummaryWriter('runs/training')
# Model hyper-parameters
self.imsize = config.imsize
self.parallel = config.parallel
self.arch = config.arch
self.epochs = config.epochs
self.batch_size = config.batch_size
self.num_workers = config.num_workers
self.total_iters = self.epochs * len(self.data_loader)
self.classes = config.classes
self.g_lr = config.g_lr
self.momentum = config.momentum
self.weight_decay = config.weight_decay
self.pretrained_model = config.pretrained_model # int type
self.indicator = False if self.pretrained_model > 0 else True # 是否是断点状态？
self.img_path = config.img_path
self.label_path = config.label_path
self.model_save_path = config.model_save_path
self.sample_path = config.sample_path
self.sample_step = config.sample_step
self.tb_step = config.tb_step
# Path
self.sample_path = osp.join(config.sample_path, self.arch)
self.model_save_path = osp.join(
config.model_save_path, self.arch)
self.build_model()
# Start with trained model
if self.pretrained_model:
self.load_pretrained_model()
self.lr_scheduler = WarmupPolyLR(
self.g_optimizer, max_iters=self.total_iters, power=0.9, warmup_factor=1.0 / 3, warmup_iters=500,
warmup_method='linear')
def train(self):
if self.pretrained_model:
start = self.pretrained_model + 1
else:
start = 0
criterion = CriterionAll()
criterion.cuda()
best_miou = 0
# Data iterator
for epoch in range(start, self.epochs):
self.G.train()
for i_iter, batch in enumerate(self.data_loader):
i_iter += len(self.data_loader) * epoch
# lr = adjust_learning_rate(self.g_lr,
# self.g_optimizer, i_iter, self.total_iters)
imgs, labels, edges = batch
size = labels.size()
imgs = imgs.cuda()
labels = labels.cuda()
if self.arch in __BA__:
edges = edges.cuda()
preds = self.G(imgs)
c_loss = criterion(preds, [labels, edges])
labels_predict = preds[0][-1]
else:
labels = labels.cuda()
# oneHot_size = (size[0], self.classes, size[1], size[2])
# labels_real = torch.cuda.FloatTensor(
# torch.Size(oneHot_size)).zero_()
# labels_real = labels_real.scatter_(
# 1, labels.data.long().cuda(), 1.0)
labels_predict = self.G(imgs)
c_loss = cross_entropy2d(
labels_predict, labels.long(), reduction='mean')
self.reset_grad()
c_loss.backward()
# 备注：这里为了渐简便没有对优化器进行参数断点记录！！！
self.g_optimizer.step()
self.lr_scheduler.step(epoch=None)
# scalr info on tensorboard
if (i_iter + 1) % self.tb_step == 0:
self.writer.add_scalar(
'cross_entrophy_loss', c_loss.data, i_iter)
self.writer.add_scalar(
'learning_rate', self.g_optimizer.param_groups[0]['lr'], i_iter)
# # Sample images
# if (i_iter + 1) % self.sample_step == 0:
# labels_sample = generate_label(
# labels_predict, self.imsize)
# save_image(denorm(labels_sample.data),
# osp.join(self.sample_path, '{}_predict.png'.format(i_iter + 1)))
print('iter={} of {} completed, loss={}'.format(
i_iter, self.total_iters, c_loss.data))
miou = self.verifier.validation(self.G)
if miou > best_miou:
best_miou = miou
torch.save(self.G.state_dict(), osp.join(
self.model_save_path, '{}_{}_G.pth'.format(str(epoch), str(round(best_miou, 4)))))
def build_model(self):
self.G = get_model(self.arch, pretrained=self.indicator).cuda()
if self.parallel:
self.G = nn.DataParallel(self.G)
# Loss and optimizer
self.g_optimizer = torch.optim.SGD(filter(
lambda p: p.requires_grad, self.G.parameters()), self.g_lr, self.momentum, self.weight_decay)
def load_pretrained_model(self):
self.G.load_state_dict(torch.load(osp.join(
self.model_save_path, '{}_G.pth'.format(self.pretrained_model))))
print('loaded trained models (step: {})..!'.format(self.pretrained_model))
def reset_grad(self):
self.g_optimizer.zero_grad()