/
train_tasknet.py
237 lines (192 loc) · 8.81 KB
/
train_tasknet.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
235
236
237
import argparse
import copy
import numpy as np
import json
import os
import torch
from utils import RunningStatistics, adjust_learning_rate
import losses
from nets import GazeHeadResNet
from dataset import HDFDataset
import torch.nn as nn
from torch.utils.data import DataLoader
import logging
from argparse import Namespace
import warnings
warnings.filterwarnings('ignore')
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
parser = argparse.ArgumentParser(description='Train a gaze/headpose estimation model.')
parser.add_argument('--config_json', type=str, help='Path to config in JSON format')
args = parser.parse_args()
#####################################################
# load configurations
assert os.path.isfile(args.config_json)
logging.info('Loading ' + args.config_json)
config = json.load(open(args.config_json))
config = Namespace(**config)
if not os.path.exists(config.save_path):
os.makedirs(config.save_path)
config.lr = config.batch_size*config.base_learning_rate
#####################################################
# save configurations
target_dir = config.save_path + '/configs'
if not os.path.isdir(target_dir):
os.makedirs(target_dir)
fpath = os.path.relpath(target_dir + '/params.json')
with open(fpath, 'w') as f:
json.dump(vars(config), f, indent=4)
logging.info('Written %s' % fpath)
#####################################################
# load datasets
with open('./gazecapture_split.json', 'r') as f:
all_gc_prefixes = json.load(f)
train_prefixes = all_gc_prefixes['train']
train_dataset = HDFDataset(hdf_file_path=config.gazecapture_file,
key_data='',
prefixes=train_prefixes)
val_prefixes = all_gc_prefixes['val']
val_dataset = HDFDataset(hdf_file_path=config.gazecapture_file,
key_data='',
prefixes=val_prefixes)
test_prefixes = all_gc_prefixes['test']
test_gc_dataset = HDFDataset(hdf_file_path=config.gazecapture_file,
key_data='',
prefixes=test_prefixes)
mpii_dataset = HDFDataset(hdf_file_path=config.mpiigaze_file, key_data='', prefixes=None)
train_dataloader = DataLoader(train_dataset,
batch_size=int(config.batch_size),
shuffle=True,
drop_last=True,
num_workers=config.num_data_loaders,
pin_memory=True,
)
val_dataloader = DataLoader(val_dataset,
batch_size=int(config.batch_size),
shuffle=False,
drop_last=False,
num_workers=config.num_data_loaders,
pin_memory=True,
)
test_gc_dataloader = DataLoader(test_gc_dataset,
batch_size=int(config.batch_size),
shuffle=False,
drop_last=False,
num_workers=config.num_data_loaders,
pin_memory=True,
)
test_mpii_dataloader = DataLoader(mpii_dataset,
batch_size=int(config.batch_size),
shuffle=False,
drop_last=False,
num_workers=config.num_data_loaders,
pin_memory=True,
)
# logging data stats.
logging.info('')
logging.info("Train datset size: %s" % len(train_dataset))
logging.info("Val dataset size: %s" % (len(val_dataset)))
#####################################################
# create network
network = GazeHeadResNet(norm_layer='instance').to(device)
# Transfer on the GPU before constructing and optimizer
if torch.cuda.device_count() >= 1:
logging.info('Using %d GPUs!' % torch.cuda.device_count())
network = nn.DataParallel(network)
net_optimizer = torch.optim.Adam(network.parameters(), lr=config.lr, weight_decay=config.l2_reg)
#####################################################
# single training step
def execute_training_step(current_step):
global train_data_iterator
try:
train_input = next(train_data_iterator)
except StopIteration:
torch.cuda.empty_cache()
global train_dataloader
train_data_iterator = iter(train_dataloader)
train_input = next(train_data_iterator)
network.train()
gaze_hat, head_hat = network(train_input['_image'].to(device))
g_loss = losses.gaze_angular_loss(y=train_input['_gaze'].to(device), y_hat=gaze_hat)
h_loss = losses.gaze_angular_loss(y=train_input['_head'].to(device), y_hat=head_hat)
loss = g_loss + h_loss
net_optimizer.zero_grad()
loss.backward()
net_optimizer.step()
running_losses.add('train_gaze_loss', g_loss.detach().cpu().numpy())
running_losses.add('train_head_loss', h_loss.detach().cpu().numpy())
#####################################################
# single val/test step
def execute_test(test_data):
test_losses = RunningStatistics()
with torch.no_grad():
network.eval()
for idx, data_dict in enumerate(test_data):
gaze_hat, head_hat = network(data_dict['_image'].to(device))
gaze_loss = losses.gaze_angular_loss(data_dict['_gaze'].to(device), gaze_hat)
head_loss = losses.gaze_angular_loss(data_dict['_head'].to(device), head_hat)
test_losses.add('val_gaze_loss', gaze_loss.detach().cpu().numpy())
test_losses.add('val_head_loss', head_loss.detach().cpu().numpy())
test_loss_means = test_losses.means()
logging.info('Test Losses at [%7d]: %s' %
(current_step, ', '.join(['%s: %.6f' % v for v in test_loss_means.items()])))
return np.sum([v for k, v in test_loss_means.items()])
#####################################################
logging.info('Training')
running_losses = RunningStatistics()
train_data_iterator = iter(train_dataloader)
val_best_acc = float('inf')
best_model = copy.deepcopy(network)
# main training loop
for current_step in range(0, config.num_training_steps):
# lr decay
if (current_step % config.decay_steps == 0):
lr = adjust_learning_rate([net_optimizer], config.decay, int(current_step /config.decay_steps), config.lr)
# Testing loop: every specified iterations compute the test statistics
if current_step % config.print_freq_test == 0 and current_step != 0:
network.eval()
torch.cuda.empty_cache()
# test
val_loss = execute_test(val_dataloader)
if val_loss < val_best_acc:
val_best_acc = val_loss
torch.save(network.state_dict(), os.path.join(config.save_path, str(current_step) + '.pth.tar'))
best_model = copy.deepcopy(network)
torch.cuda.empty_cache()
# Training step
execute_training_step(current_step)
# Print training loss
if current_step != 0 and (current_step % config.print_freq_train == 0):
running_loss_means = running_losses.means()
logging.info('Losses at [%7d]: %s' %
(current_step,
', '.join(['%s: %.5f' % v
for v in running_loss_means.items()])))
running_losses.reset()
logging.info('Finished Training')
###################################
# TESTING
## on MPIIGaze
test_losses = RunningStatistics()
with torch.no_grad():
best_model.eval()
for idx, data_dict in enumerate(test_mpii_dataloader):
gaze_hat, head_hat = best_model(data_dict['_image'].to(device))
gaze_loss = losses.gaze_angular_loss(data_dict['_gaze'].to(device), gaze_hat)
head_loss = losses.gaze_angular_loss(data_dict['_head'].to(device), head_hat)
test_losses.add('val_gaze_loss', gaze_loss.detach().cpu().numpy())
test_losses.add('val_head_loss', head_loss.detach().cpu().numpy())
test_loss_means = test_losses.means()
logging.info('MPIIGaze Test Losses: %s' % ', '.join(['%s: %.6f' % v for v in test_loss_means.items()]))
## on GC Test
test_losses = RunningStatistics()
with torch.no_grad():
best_model.eval()
for idx, data_dict in enumerate(test_gc_dataloader):
gaze_hat, head_hat = best_model(data_dict['_image'].to(device))
gaze_loss = losses.gaze_angular_loss(data_dict['_gaze'].to(device), gaze_hat)
head_loss = losses.gaze_angular_loss(data_dict['_head'].to(device), head_hat)
test_losses.add('val_gaze_loss', gaze_loss.detach().cpu().numpy())
test_losses.add('val_head_loss', head_loss.detach().cpu().numpy())
test_loss_means = test_losses.means()
logging.info('GC Test Losses: %s' % ', '.join(['%s: %.6f' % v for v in test_loss_means.items()]))