-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
201 lines (153 loc) · 6.72 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
import os
import sys
import os.path as osp
import cv2
import datetime
import numpy as np
from functools import partial
from omegaconf import OmegaConf
from tqdm import tqdm
from rich.progress import track
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.optim.lr_scheduler import CyclicLR
import torchvision
from torch.utils.tensorboard import SummaryWriter
from typing import Optional
from torchsummary import summary
from losses.stereo_loss import IterationLoss, StereoL1Loss
from losses.gaze_loss import GazeLoss
from utils.helper import AverageMeter, recover_image
from utils.math import rotation_matrix_2d, pitchyaw_to_vector, vector_to_pitchyaw, angular_error
class Trainer(nn.Module):
def __init__(self,
config,
model,
metrics,
train_loader,
test_loader,):
super().__init__()
self.config = config
self.train_loader = train_loader
self.test_loader = test_loader
self.model = model
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if config.ckpt_resume is not None:
ckpt = torch.load( config.ckpt_resume )
self.model.load_state_dict(ckpt, strict=True)
print('load from ckpt: ', config.ckpt_resume )
self.model.to(self.device)
summary(model)
self.metrics = metrics
self.optimizer = optim.Adam(self.model.parameters(), lr=0, weight_decay=1e-6)
num_step_per_epoch = len(train_loader.dataset) // config.batch_size
step_size_up = int(num_step_per_epoch // 2)
step_size_down = num_step_per_epoch - step_size_up
self.scheduler = CyclicLR(self.optimizer, base_lr=1e-6, max_lr=1e-3,
step_size_up=step_size_up,
step_size_down=step_size_down,
mode='triangular2', cycle_momentum=False)
self.start_epoch = 0
self.epochs = 15
self.train_iter = 0
self.output_dir = config.output_dir
os.makedirs(self.output_dir, exist_ok=True)
OmegaConf.save(config, osp.join(self.output_dir, 'config.yaml'))
self.ckpt_dir = osp.join(self.output_dir, 'ckpt')
os.makedirs(self.ckpt_dir, exist_ok=True)
self.image_dir = osp.join(self.output_dir, 'image')
os.makedirs(self.image_dir, exist_ok=True)
self.tensorboard_dir = osp.join(self.output_dir, 'tensorboard')
os.makedirs(self.tensorboard_dir, exist_ok=True)
self.writer = SummaryWriter(log_dir=self.tensorboard_dir)
self.print_freq = config.print_freq
def train(self):
error = self.test(-1)
for epoch in range(self.start_epoch, self.epochs):
self.train_one_epoch(epoch)
error = self.test(epoch)
if (epoch + 1) % self.config.save_epoch == 0:
add_file_name = 'epoch_' + str(epoch+1).zfill(2) + '_error=' + str(round(error, 2))
self.save_checkpoint(
state=self.model.module.state_dict() if isinstance(self.model, torch.nn.DataParallel) else self.model.state_dict(),
add=add_file_name
)
def prepare_dual_input(self, batch):
img_0 = batch['img_0'].float().to(self.device)
gt_gaze = batch['gt_gaze'].float().to(self.device)
head_pose_0 = batch['head_pose_0'].float().to(self.device)
img_1 = batch['img_1'].float().to(self.device)
gt_gaze_1 = batch['gt_gaze_1'].float().to(self.device)
head_pose_1 = batch['head_pose_1'].float().to(self.device)
rot_0 = rotation_matrix_2d(head_pose_0) ## from canonical to head_0
rot_1 = rotation_matrix_2d(head_pose_1) ## from canonical to head_1
data = {"img_0": img_0, "rot_0": rot_0, "gt_gaze": gt_gaze,
"img_1": img_1, "rot_1": rot_1, "gt_gaze_1": gt_gaze_1}
data.update({"idx_0": batch['idx_0'], "idx_1": batch['idx_1']})
return data
def train_one_epoch(self, epoch):
print(f'Epoch: {epoch + 1} / {self.epochs}')
self.model.train()
for i, data in enumerate(track(self.train_loader, description='Training', transient=True)):
data = self.prepare_dual_input(data)
data = self.model(data)
loss_gaze = self.metrics(data)
pred_gaze = data["pred_gaze"]
gaze_var = data["gt_gaze"]
error_gaze = np.mean(angular_error(pred_gaze.cpu().data.numpy(), gaze_var.cpu().data.numpy()))
if self.train_iter!=0 and self.train_iter % self.print_freq == 0:
print('train on iter: ', self.train_iter)
print('loss_gaze: ', loss_gaze.item())
print('error_gaze: ', error_gaze.item())
self.writer.add_scalar( 'train/loss_gaze', loss_gaze.item(), self.train_iter)
self.writer.add_scalar( 'train/error_gaze', error_gaze.item(), self.train_iter)
samples_to_show = min(8, gaze_var.size(0))
self.writer.add_image( 'train/images_0', torchvision.utils.make_grid(data["img_0"][:samples_to_show], nrow=(samples_to_show//2), normalize=True), self.train_iter)
self.writer.add_image( 'train/images_1', torchvision.utils.make_grid(data["img_1"][:samples_to_show], nrow=(samples_to_show//2), normalize=True), self.train_iter)
self.optimizer.zero_grad()
loss_gaze.backward()
self.optimizer.step()
self.train_iter += 1
self.scheduler.step()
def save_checkpoint(self, state, add=None):
"""
Save a copy of the model
"""
if add is not None:
filename = add + '.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: ', ckpt_path)
def test(self, epoch):
self.model.eval()
pred_gaze_all = np.zeros((len(self.test_loader.dataset), 2))
gt_gaze_all = np.zeros((len(self.test_loader.dataset), 2))
save_index = 0
for i, data in enumerate(track(self.test_loader, description='Testing', transient=True)):
with torch.no_grad():
data = self.prepare_dual_input(data)
data = self.model(data)
pred_gaze = data["pred_gaze"]
gaze_var = data["gt_gaze"]
input_var = data["img_0"]
batch_size = input_var.size(0)
if i != 0 and i % self.print_freq == 0:
samples_to_show = min(8, batch_size)
self.writer.add_image( 'test/images_0', torchvision.utils.make_grid(data["img_0"][:samples_to_show], nrow=(samples_to_show//2), normalize=True), i)
self.writer.add_image( 'test/images_1', torchvision.utils.make_grid(data["img_1"][:samples_to_show], nrow=(samples_to_show//2), normalize=True), i)
pred_gaze_all[save_index:save_index + batch_size, :] = pred_gaze.cpu().data.numpy()
gt_gaze_all[save_index:save_index + batch_size, :] = gaze_var.cpu().data.numpy()
save_index += input_var.size(0)
if save_index != len(self.test_loader.dataset):
print('the test samples save_index ', save_index, ' is not equal to the whole test set ', len(self.test_loader.dataset))
avg_error_gaze = np.mean(angular_error(pred_gaze_all, gt_gaze_all))
msg = 'test on epoch {}, error: {}\n'.format(epoch + 1, avg_error_gaze)
print( msg )
self.writer.add_scalar('test/epoch_error_gaze', avg_error_gaze, epoch)
with open(osp.join(self.output_dir, 'test_results.txt'), 'a') as f:
f.write(msg)
return avg_error_gaze