-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain.py
207 lines (163 loc) · 6.58 KB
/
main.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
from __future__ import print_function, division
import os
import random
import ctypes
import setproctitle
import time
import numpy as np
import torch
import torch.multiprocessing as mp
from tensorboardX import SummaryWriter
from utils import command_parser
from utils.class_finder import model_class, agent_class, optimizer_class
from utils.model_util import ScalarMeanTracker
from utils.data_utils import check_data, loading_scene_list
from main_eval import main_eval
from full_eval import full_eval
from runners import a3c_train, a3c_val
os.environ["OMP_NUM_THREADS"] = "1"
def main():
setproctitle.setproctitle("Train/Test Manager")
args = command_parser.parse_arguments()
print('Training started from: {}'.format(
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
)
args.learned_loss = False
args.num_steps = 50
target = a3c_val if args.eval else a3c_train
scenes = loading_scene_list(args)
create_shared_model = model_class(args.model)
init_agent = agent_class(args.agent_type)
optimizer_type = optimizer_class(args.optimizer)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
random.seed(args.seed)
if args.eval:
args.test_or_val = 'test'
main_eval(args, create_shared_model, init_agent)
return
start_time = time.time()
local_start_time_str = time.strftime(
'%Y_%m_%d_%H_%M_%S', time.localtime(start_time)
)
tb_log_dir = args.log_dir + '/' + args.title + '_' + args.phase + '_' + local_start_time_str
log_writer = SummaryWriter(log_dir=tb_log_dir)
if args.gpu_ids == -1:
args.gpu_ids = [-1]
else:
torch.cuda.manual_seed(args.seed)
mp.set_start_method("spawn" , force = True)
shared_model = create_shared_model(args)
train_total_ep = 0
n_frames = 0
if args.pretrained_trans is not None:
saved_state = torch.load(
args.pretrained_trans, map_location=lambda storage, loc: storage
)
model_dict = shared_model.state_dict()
pretrained_dict = {k: v for k, v in saved_state['model'].items() if
(k in model_dict and v.shape == model_dict[k].shape)}
model_dict.update(pretrained_dict)
shared_model.load_state_dict(model_dict)
if args.continue_training is not None:
orgin_state = shared_model.state_dict()
saved_state = torch.load(
args.continue_training, map_location=lambda storage, loc: storage
)
orgin_state.update(saved_state)
shared_model.load_state_dict(orgin_state)
train_total_ep = int(args.continue_training.split('_')[-7])
n_frames = int(args.continue_training.split('_')[-8])
if args.fine_tuning is not None:
saved_state = torch.load(
args.fine_tuning, map_location=lambda storage, loc: storage
)
model_dict = shared_model.state_dict()
pretrained_dict = {k: v for k, v in saved_state.items() if (k in model_dict and v.shape == model_dict[k].shape)}
model_dict.update(pretrained_dict)
shared_model.load_state_dict(model_dict)
if args.update_meta_network:
for layer, parameters in shared_model.named_parameters():
if not layer.startswith('meta'):
parameters.requires_grad = False
shared_model.share_memory()
if args.pretrained_trans is not None:
optimizer = optimizer_type(
[
{'params': [v for k, v in shared_model.named_parameters() if
v.requires_grad and (k in pretrained_dict)],
'lr': args.pretrained_low_lr},
{'params': [v for k, v in shared_model.named_parameters() if
v.requires_grad and (k not in pretrained_dict)],
'lr': args.lr},
]
)
else:
optimizer = optimizer_type(
[v for k, v in shared_model.named_parameters() if v.requires_grad], lr=args.lr
)
optimizer = optimizer_type(
[v for k, v in shared_model.named_parameters() if v.requires_grad], lr=args.lr
)
optimizer.share_memory()
print(shared_model)
processes = []
end_flag = mp.Value(ctypes.c_bool, False)
train_res_queue = mp.Queue()
for rank in range(0, args.workers):
p = mp.Process(
target=target,
args=(
rank,
args,
create_shared_model,
shared_model,
init_agent,
optimizer,
train_res_queue,
end_flag,
scenes,
),
)
p.start()
setproctitle.setproctitle('python')
processes.append(p)
time.sleep(0.1)
print("Train agents created.")
train_thin = args.train_thin
train_scalars = ScalarMeanTracker()
try:
while train_total_ep < args.max_ep:
train_result = train_res_queue.get()
train_scalars.add_scalars(train_result)
train_total_ep += 1
n_frames += train_result['ep_length']
if (train_total_ep % train_thin) == 0:
log_writer.add_scalar('n_frames', n_frames, train_total_ep)
tracked_means = train_scalars.pop_and_reset()
for k in tracked_means:
log_writer.add_scalar(
k + '/train', tracked_means[k], train_total_ep
)
if (train_total_ep % args.ep_save_freq) == 0:
print('{}: {}'.format(train_total_ep, n_frames))
if not os.path.exists(args.save_model_dir):
os.makedirs(args.save_model_dir)
state_to_save = shared_model.state_dict()
save_path = os.path.join(
args.save_model_dir,
'{0}_{1}_{2}_{3}.dat'.format(
args.title, n_frames, train_total_ep, local_start_time_str
),
)
torch.save(state_to_save, save_path)
finally:
log_writer.close()
end_flag.value = True
for p in processes:
time.sleep(0.1)
p.join()
if args.test_after_train:
full_eval()
if __name__ == "__main__":
main()