-
Notifications
You must be signed in to change notification settings - Fork 35
/
trainer.py
221 lines (180 loc) · 7.91 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
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
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from torch.autograd import Variable
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR
import torch.nn as nn
import math
import os
import time
import numpy as np
from utils import AverageMeter, angular_error
from model import gaze_network
import shutil
class Trainer(object):
def __init__(self, config, data_loader):
"""
Construct a new Trainer instance.
Args
----
- config: object containing command line arguments.
- data_loader: data iterator
"""
self.config = config
# data params
if config.is_train:
self.train_loader = data_loader
self.num_train = len(self.train_loader.dataset)
else:
self.test_loader = data_loader
self.num_test = len(self.test_loader.dataset)
self.batch_size = config.batch_size
# training params
self.epochs = config.epochs # the total epoch to train
self.start_epoch = 0
self.lr = config.init_lr
self.lr_patience = config.lr_patience
self.lr_decay_factor = config.lr_decay_factor
# misc params
self.use_gpu = config.use_gpu
self.ckpt_dir = config.ckpt_dir # output dir
self.print_freq = config.print_freq
self.train_iter = 0
self.pre_trained_model_path = config.pre_trained_model_path
if self.use_gpu and torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
# configure tensorboard logging
log_dir = './logs/' + os.path.basename(os.getcwd())
if os.path.exists(log_dir) and os.path.isdir(log_dir):
shutil.rmtree(log_dir)
self.writer = SummaryWriter(log_dir=log_dir)
# build model
self.model = gaze_network()
if self.use_gpu:
self.model.cuda()
print('[*] Number of model parameters: {:,}'.format(
sum([p.data.nelement() for p in self.model.parameters()])))
# initialize optimizer and scheduler
self.optimizer = optim.Adam(
self.model.parameters(), lr=self.lr)
self.scheduler = StepLR(
self.optimizer, step_size=self.lr_patience, gamma=self.lr_decay_factor)
def train(self):
print("\n[*] Train on {} samples".format(self.num_train))
# train for each epoch
for epoch in range(self.start_epoch, self.epochs):
print(
'\nEpoch: {}/{} - base LR: {:.6f}'.format(
epoch + 1, self.epochs, self.lr)
)
for param_group in self.optimizer.param_groups:
print('Learning rate: ', param_group['lr'])
# train for 1 epoch
print('Now go to training')
self.model.train()
train_acc, loss_gaze = \
self.train_one_epoch(epoch, self.train_loader)
# save the model for each epoch
add_file_name = 'epoch_' + str(epoch)
self.save_checkpoint(
{'epoch': epoch + 1,
'model_state': self.model.state_dict(),
'optim_state': self.optimizer.state_dict(),
'scheule_state': self.scheduler.state_dict()
}, add=add_file_name
)
self.scheduler.step() # update learning rate
self.writer.close()
def train_one_epoch(self, epoch, data_loader, is_train=True):
"""
Train the model for 1 epoch of the training set.
"""
batch_time = AverageMeter()
errors = AverageMeter()
losses_gaze = AverageMeter()
tic = time.time()
for i, (input_img, target) in enumerate(data_loader):
input_var = torch.autograd.Variable(input_img.float().cuda())
target_var = torch.autograd.Variable(target.float().cuda())
# train gaze net
pred_gaze= self.model(input_var)
gaze_error_batch = np.mean(angular_error(pred_gaze.cpu().data.numpy(), target_var.cpu().data.numpy()))
errors.update(gaze_error_batch.item(), input_var.size()[0])
loss_gaze = F.l1_loss(pred_gaze, target_var)
self.optimizer.zero_grad()
loss_gaze.backward()
self.optimizer.step()
losses_gaze.update(loss_gaze.item(), input_var.size()[0])
if i % self.print_freq == 0:
self.writer.add_scalar('Loss/gaze', losses_gaze.avg, self.train_iter)
# report information
if i % self.print_freq == 0 and i is not 0:
print('--------------------------------------------------------------------')
msg = "train error: {:.3f} - loss_gaze: {:.5f}"
print(msg.format(errors.avg, losses_gaze.avg))
# measure elapsed time
print('iteration ', self.train_iter)
toc = time.time()
batch_time.update(toc - tic)
# print('Current batch running time is ', np.round(batch_time.avg / 60.0), ' mins')
tic = time.time()
# estimate the finish time
est_time = (self.epochs - epoch) * (self.num_train / self.batch_size) * batch_time.avg / 60.0
print('Estimated training time left: ', np.round(est_time), ' mins')
self.writer.add_scalar('Error/train', errors.avg, self.train_iter)
errors.reset()
losses_gaze.reset()
self.train_iter = self.train_iter + 1
toc = time.time()
batch_time.update(toc-tic)
print('running time is ', batch_time.avg)
return errors.avg, losses_gaze.avg
def test(self):
"""
Test the pre-treained model on the whole test set. Note there is no label released to public, you can
only save the predicted results. You then need to submit the test resutls to our evaluation website to
get the final gaze estimation error.
"""
print('We are now doing the final test')
self.model.eval()
self.load_checkpoint(is_strict=False, input_file_path=self.pre_trained_model_path)
pred_gaze_all = np.zeros((self.num_test, 2))
mean_error = []
save_index = 0
print('Testing on ', self.num_test, ' samples')
for i, (input_img) in enumerate(self.test_loader):
input_var = torch.autograd.Variable(input_img.float().cuda())
pred_gaze = self.model(input_var)
pred_gaze_all[save_index:save_index+self.batch_size, :] = pred_gaze.cpu().data.numpy()
save_index += input_var.size(0)
if save_index != self.num_test:
print('the test samples save_index ', save_index, ' is not equal to the whole test set ', self.num_test)
print('Tested on : ', pred_gaze_all.shape[0], ' samples')
np.savetxt('test_results.txt', pred_gaze_all, delimiter=',')
def save_checkpoint(self, state, add=None):
"""
Save a copy of the model
"""
if add is not None:
filename = add + '_ckpt.pth.tar'
else:
filename ='ckpt.pth.tar'
ckpt_path = os.path.join(self.ckpt_dir, filename)
torch.save(state, ckpt_path)
print('save file to: ', filename)
def load_checkpoint(self, input_file_path='./ckpt/ckpt.pth.tar', is_strict=True):
"""
Load the copy of a model.
"""
print('load the pre-trained model: ', input_file_path)
ckpt = torch.load(input_file_path)
# load variables from checkpoint
self.model.load_state_dict(ckpt['model_state'], strict=is_strict)
self.optimizer.load_state_dict(ckpt['optim_state'])
self.scheduler.load_state_dict(ckpt['scheule_state'])
self.start_epoch = ckpt['epoch'] - 1
print(
"[*] Loaded {} checkpoint @ epoch {}".format(
input_file_path, ckpt['epoch'])
)