-
Notifications
You must be signed in to change notification settings - Fork 10
/
base_trainer.py
234 lines (196 loc) · 8.52 KB
/
base_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
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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import os
import time
import yaml
import torch
import datetime
from torch.utils.tensorboard import SummaryWriter
import torch.utils.data
import numpy as np
import glob
import shutil
from utils.net_util import to_cuda
def worker_init_fn(worker_id): # set numpy's random seed
seed = torch.initial_seed()
seed = seed % (2 ** 32)
np.random.seed(seed + worker_id)
class BaseTrainer:
def __init__(self, opt):
self.opt = opt
self.dataset = None
self.network = None
self.net_dict = {}
self.optm_dict = {}
self.update_keys = None
self.lr_schedule_dict = {}
self.iter_idx = 0
self.loss_weight = self.opt['train']['loss_weight']
@staticmethod
def load_pretrained(path, dict_):
data = torch.load(path)
for k in dict_:
if k in data:
print('# Loading %s...' % k)
dict_[k].load_state_dict(data[k])
else:
print('# %s not found!' % k)
def load_ckpt(self, path, load_optm = True):
self.load_pretrained(path + '/net.pt', self.net_dict)
if load_optm:
if os.path.exists(path + '/optm.pt'):
self.load_pretrained(path + '/optm.pt', self.optm_dict)
else:
print('# Optimizer not found!')
@staticmethod
def save_trained(path, dict_):
data = {}
for k in dict_:
data[k] = dict_[k].state_dict()
torch.save(data, path)
def save_ckpt(self, path, save_optm = True):
self.save_trained(path + '/net.pt', self.net_dict)
if save_optm:
self.save_trained(path + '/optm.pt', self.optm_dict)
def zero_grad(self):
if self.update_keys is None:
update_keys = self.optm_dict.keys()
else:
update_keys = self.update_keys
for k in update_keys:
self.optm_dict[k].zero_grad()
def step(self):
if self.update_keys is None:
update_keys = self.optm_dict.keys()
else:
update_keys = self.update_keys
for k in update_keys:
self.optm_dict[k].step()
def update_lr(self, iter_idx):
lr_dict = {}
if self.update_keys is None:
update_keys = self.optm_dict.keys()
else:
update_keys = self.update_keys
for k in update_keys:
lr = self.lr_schedule_dict[k].get_learning_rate(iter_idx)
for param_group in self.optm_dict[k].param_groups:
param_group['lr'] = lr
lr_dict[k] = lr
return lr_dict
def set_dataset(self, dataset):
self.dataset = dataset
def set_network(self, network):
self.network = network
def set_net_dict(self, net_dict):
self.net_dict = net_dict
def set_optm_dict(self, optm_dict):
self.optm_dict = optm_dict
def set_update_keys(self, update_keys):
self.update_keys = update_keys
def set_lr_schedule_dict(self, lr_schedule_dict):
self.lr_schedule_dict = lr_schedule_dict
def set_train(self, flag = True):
if flag:
for k, net in self.net_dict.items():
if k in self.update_keys:
net.train()
else:
net.eval()
else:
for k, net in self.net_dict.items():
net.eval()
def train(self):
# log
os.makedirs(self.opt['train']['net_ckpt_dir'], exist_ok = True)
log_dir = self.opt['train']['net_ckpt_dir'] + '/' + datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S')
os.makedirs(log_dir, exist_ok = True)
writer = SummaryWriter(log_dir)
yaml.dump(self.opt, open(log_dir + '/config_bk.yaml', 'w'), sort_keys = False)
self.set_train()
self.dataset.training = True
batch_size = self.opt['train'].get('batch_size', 1)
num_workers = self.opt['train'].get('num_workers', 0)
dataloader = torch.utils.data.DataLoader(self.dataset,
batch_size = batch_size,
shuffle = True,
num_workers = num_workers,
worker_init_fn = worker_init_fn,
drop_last = True)
batch_num = len(self.dataset) // batch_size
if self.opt['train'].get('save_init_ckpt', False) and self.opt['train'].get('start_epoch', 0) == 0:
init_folder = self.opt['train']['net_ckpt_dir'] + '/init_ckpt'
if not os.path.exists(init_folder) or self.opt['train']['start_epoch'] == 0:
os.makedirs(init_folder, exist_ok = True)
self.save_ckpt(init_folder, False)
else:
print('# Init checkpoint has been saved!')
if self.opt['train']['prev_ckpt'] is not None:
self.load_ckpt(self.opt['train']['prev_ckpt'])
start_epoch = self.opt['train'].get('start_epoch', 0)
end_epoch = self.opt['train'].get('end_epoch', 999)
for epoch_idx in range(start_epoch, end_epoch):
self.update_config_before_epoch(epoch_idx)
epoch_losses = dict()
for batch_idx, items in enumerate(dataloader):
time0 = time.time()
iter_idx = batch_idx + batch_num * epoch_idx
self.iter_idx = iter_idx
lr_dict = self.update_lr(iter_idx)
items = to_cuda(items)
self.zero_grad()
loss, batch_losses = self.forward_one_pass(items)
loss.backward()
self.step()
# record batch loss
log_info = 'epoch %d, batch %d, ' % (epoch_idx, batch_idx)
log_info += 'lr: '
for k in lr_dict.keys():
log_info += '%s %e, ' % (k, lr_dict[k])
for key in batch_losses.keys():
log_info = log_info + ('%s: %f, ' % (key, batch_losses[key]))
writer.add_scalar('%s/Batch' % key, batch_losses[key], iter_idx)
if key in epoch_losses:
epoch_losses[key] += batch_losses[key]
else:
epoch_losses[key] = batch_losses[key]
print(log_info)
with open(os.path.join(log_dir, 'loss.txt'), 'a') as fp:
# record loss weight
if batch_idx == 0:
loss_weights_info = ''
for k in self.opt['train']['loss_weight'].keys():
loss_weights_info += '%s: %f, ' % (k, self.opt['train']['loss_weight'][k])
fp.write('# Loss weights: \n' + loss_weights_info + '\n')
fp.write(log_info + '\n')
if iter_idx % self.opt['train']['ckpt_interval']['batch'] == 0 and iter_idx != 0:
for folder in glob.glob(self.opt['train']['net_ckpt_dir'] + '/batch_*'):
shutil.rmtree(folder)
model_folder = self.opt['train']['net_ckpt_dir'] + '/batch_%d' % iter_idx
os.makedirs(model_folder, exist_ok = True)
self.save_ckpt(model_folder, save_optm = False)
if iter_idx % self.opt['train']['eval_interval'] == 0 and iter_idx != 0:
# if True:
self.mini_test()
time1 = time.time()
print('One iteration costs %f secs' % (time1 - time0))
""" EPOCH """
# record epoch loss
for key in epoch_losses.keys():
epoch_losses[key] /= batch_num
writer.add_scalar('%s/Epoch' % key, epoch_losses[key], epoch_idx)
if epoch_idx % self.opt['train']['ckpt_interval']['epoch'] == 0:
model_folder = self.opt['train']['net_ckpt_dir'] + '/epoch_%d' % epoch_idx
os.makedirs(model_folder, exist_ok = True)
self.save_ckpt(model_folder)
if batch_num > 50:
latest_folder = self.opt['train']['net_ckpt_dir'] + '/epoch_latest'
os.makedirs(latest_folder, exist_ok = True)
self.save_ckpt(latest_folder)
writer.close()
@torch.no_grad()
def mini_test(self):
""" Test during training """
pass
def forward_one_pass(self, items):
raise NotImplementedError('"forward_one_pass" method is not implemented!')
def update_config_before_epoch(self, epoch_idx):
pass