-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathfinetune_temporal_distribute.py
executable file
·188 lines (157 loc) · 7.7 KB
/
finetune_temporal_distribute.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
from torch.optim import SGD,Adam
import torch
from Data.dataloader import get_video_dataset
from analysis.evaluate import AutoEvaluate
from analysis.onlinetest import AutoTest
import torch.nn.functional as F
from utils.utils import get_Logger_and_SummaryWriter
import os
from utils.Distribute.engine import Engine
from config import config
from Models.mobilenetv3temporal_PCSA import Fastnet
import torch.nn as nn
from utils.SalEval import SalEval
from torchnet.meter import AverageValueMeter
import numpy as np
class CrossEntropyLoss(nn.Module):
def __init__(self):
super(CrossEntropyLoss, self).__init__()
def forward(self, *inputs):
pred, target = tuple(inputs)
total_loss = F.binary_cross_entropy(pred, target.float())
return total_loss
class TrainSchedule(object):
def __init__(self, batches_per_epoch):
self.cur_epoch = 0
self.total_epoch = config.finetune_epoches
self.cur_batches = 0
self.batches_per_epoch = batches_per_epoch
def update(self):
self.cur_batches += 1
if not self.cur_batches < self.batches_per_epoch:
self.cur_batches = 0
self.cur_epoch += 1
def state_dict(self):
state_dict = {"cur_batches": self.cur_batches,
"cur_epoch": self.cur_epoch,
"total_epoch": self.total_epoch,
"batches_per_epoch": self.batches_per_epoch}
return state_dict
def load_state_dict(self, state_dict):
self.cur_batches = state_dict["cur_batches"]
self.cur_epoch = state_dict["cur_epoch"]
self.total_epoch = state_dict["total_epoch"]
self.batches_per_epoch = state_dict["batches_per_epoch"]
class Train(object):
def __init__(self):
self.logger, self.writer, self.tag_dir = get_Logger_and_SummaryWriter()
self.engine=Engine(self.logger)
self.device = torch.device("cuda")
self.network=Fastnet()
self.load_backbone(self.logger)
self.network=self.network.cuda()
self.network = self.engine.data_parallel(self.network)
self.criterion = CrossEntropyLoss().to(self.device)
base_params = [params for name, params in self.network.named_parameters() if ("temporal_high" in name)]
finetune_params = [params for name, params in self.network.named_parameters() if ("temporal_high" not in name)]
self.optim = Adam([
{'params': base_params, 'lr': config.base_lr,'weight_decay':1e-4, 'name': "base_params"},
{'params': finetune_params, 'lr': config.finetune_lr,'weight_decay':1e-4, 'name': 'finetune_params'}])
self.train_dataset, self.train_multiscale_dataset, statistics = get_video_dataset()
self.train_multiscale_loader = []
self.train_multiscale_smapler = []
for dst in self.train_multiscale_dataset:
ld, sp = self.engine.get_train_loader(dst,config.video_batchsize)
self.train_multiscale_loader.append(ld)
self.train_multiscale_smapler.append(sp)
self.train_loader,self.train_sampler=self.engine.get_train_loader(self.train_dataset,config.video_batchsize)
batches_per_epoch = 0
batches_per_epoch += len(self.train_loader)
for loader in self.train_multiscale_loader:
batches_per_epoch += len(loader)
self.sche = TrainSchedule(batches_per_epoch)
if self.engine.local_rank==0:
self.logger.info(config)
self.logger.info(self.network)
total_paramters = sum([np.prod(p.size()) for p in self.network.parameters()])
self.logger.info('Total network parameters: ' + str(total_paramters))
def save_checkpoint(self):
os.makedirs(os.path.join(self.tag_dir, "epoch_%d_batch_%d" % (
self.sche.cur_epoch, self.sche.cur_batches)), exist_ok=True)
save_root = os.path.join(self.tag_dir, "epoch_%d_batch_%d" % (
self.sche.cur_epoch, self.sche.cur_batches))
torch.save(self.state_dict(), os.path.join(save_root, "checkpoint.pth"))
def adjust_learning_rate(self):
if config.lr_mode == 'poly':
cur_iter = self.sche.batches_per_epoch * self.sche.cur_epoch + self.sche.cur_batches
max_iter = self.sche.batches_per_epoch * self.sche.total_epoch
base_lr = config.base_lr * (1 - cur_iter * 1.0 / max_iter) ** 0.9
finetune_lr = config.finetune_lr * (1 - cur_iter * 1.0 / max_iter) ** 0.9
for param_group in self.optim.param_groups:
if param_group["name"] == "base_params":
param_group['lr'] = base_lr
if param_group["name"] == "finetune_params":
param_group['lr'] = finetune_lr
return base_lr, finetune_lr
def train_per_loader(self, trainloader):
self.network.train()
loss_meter = AverageValueMeter()
for idx, (img, label) in enumerate(trainloader):
baselr,finetunelr = self.adjust_learning_rate()
img = img.to(self.device)
label = label.to(self.device)
if len(label.shape) == 5:
label = label.view(-1, *(label.shape[2:]))
output = self.network(img)
loss = self.criterion(output, label)
self.optim.zero_grad()
loss.backward()
self.optim.step()
loss_meter.add(float(loss))
if self.engine.local_rank == 0:
if self.sche.cur_batches % config.log_inteval == 0:
self.logger.info("%s-epoch:%d/%d batch:%d/%d loss:%.4f base_lr:%e finetune_lr:%e" % (
self.tag_dir.split("/")[-1], self.sche.cur_epoch, self.sche.total_epoch, self.sche.cur_batches,
self.sche.batches_per_epoch, loss_meter.value()[0], baselr,finetunelr))
self.sche.update()
return loss_meter.value()[0]
def train_per_epoch(self):
for idx,loader in enumerate(self.train_multiscale_loader):
self.train_per_loader(loader)
loss_train= self.train_per_loader(self.train_loader)
if self.engine.local_rank == 0:
self.logger.info("train_img_loss:%.4f" % (loss_train))
def train(self):
while self.sche.cur_epoch < self.sche.total_epoch:
self.train_sampler.set_epoch(self.sche.cur_epoch)
for sp in self.train_multiscale_smapler:
sp.set_epoch(self.sche.cur_epoch)
self.train_per_epoch()
if self.engine.local_rank == 0:
self.save_checkpoint()
def state_dict(self):
if config.parallel is True:
state_dict = {"net": self.network.module.state_dict(),
'optimizer': self.optim.state_dict(),
'sche': self.sche.state_dict()}
else:
state_dict = {"net": self.network.state_dict(),
'optimizer': self.optim.state_dict(),
'sche': self.sche.state_dict()}
return state_dict
def load_state_dict(self, state_dict):
if config.parallel is True:
#self.sche.load_state_dict(state_dict["sche"])
#self.optim.load_state_dict(state_dict["optimizer"])
self.network.module.load_state_dict(state_dict["net"])
else:
self.sche.load_state_dict(state_dict["sche"])
self.optim.load_state_dict(state_dict["optimizer"])
self.network.load_state_dict(state_dict["net"])
def load_backbone(self,logger):
assert config.pretrain_state_dict is not None,"error"
self.network.load_backbone(torch.load(config.pretrain_state_dict,map_location=torch.device('cpu'))["net"],logger)
if __name__ == "__main__":
torch.backends.cudnn.benchmark = True
trainer = Train()
trainer.train()