-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_imitation.py
165 lines (150 loc) · 6.69 KB
/
train_imitation.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
import os
import glob
import argparse
import torch
from envs import MultimodalEnvs
from envs.env_norm import make_env
from gail.buffer import SerializedBuffer
from gail.algo import ALGOS
from gail.trainer_info import Trainer
from gail.utils import set_seed
def run(args):
# Environment for training
env = make_env(MultimodalEnvs[args.env_id](num_modes=args.num_modes))
# Environment for evaluation
env_test = make_env(MultimodalEnvs[args.env_id](num_modes=args.num_modes))
# Set random seed
set_seed(args.seed)
env.seed(args.seed)
env_test.seed(args.seed)
# Load dataset
buffer_lb = args.buffer_dir + args.env_id + '_' + str(args.num_modes) + '_modes_lb/*'
buffer_ulb = args.buffer_dir + args.env_id + '_' + str(args.num_modes) + '_modes_ulb/*'
buffer_dir_lb = sorted(glob.glob(buffer_lb)) # sort by name
buffer_dir_ulb = sorted(glob.glob(buffer_ulb)) # sort by name
# Labeled dataset for training
buffer_exp_lb = SerializedBuffer(
path=buffer_dir_lb,
device=torch.device("cuda" if args.cuda else "cpu"),
num_modes=args.num_modes
)
# Unlabeled and imbalanced dataset for training
buffer_exp_ulb = SerializedBuffer(
path=buffer_dir_ulb,
device=torch.device("cuda" if args.cuda else "cpu"),
num_modes=args.num_modes,
im_ratio=args.im_ratio
)
# Labeled dataset for evaluation
buffer_exp_ulb_eval = SerializedBuffer(
path=buffer_dir_ulb,
device=torch.device("cuda" if args.cuda else "cpu"),
num_modes=args.num_modes
)
# Initialize Ess-InfoGAIL algorithm
algo = ALGOS[args.algo](
buffer_exp_lb=buffer_exp_lb,
buffer_exp_ulb=buffer_exp_ulb,
buffer_exp_ulb_eval=buffer_exp_ulb_eval,
state_shape=tuple([env.observation_space.shape[0] + args.num_modes + 1]),
action_shape=env.action_space.shape,
device=torch.device("cuda" if args.cuda else "cpu"),
seed=args.seed,
rollout_length=args.rollout_length,
num_steps=args.num_steps,
batch_size=args.batch_size,
epoch_ppo=args.epoch_ppo,
epoch_disc=args.epoch_disc,
surrogate_loss_coef=args.surrogate_loss_coef,
disc_grad_penalty=args.disc_grad_penalty,
value_loss_coef=args.value_loss_coef,
disc_coef=args.disc_coef,
us_coef=args.us_coef,
ss_coef=args.ss_coef,
reward_i_coef=args.reward_i_coef,
reward_us_coef=args.reward_us_coef,
reward_ss_coef=args.reward_ss_coef,
reward_t_coef=args.reward_t_coef,
info_max_coef1=args.info_max_coef1,
info_max_coef2=args.info_max_coef2,
info_max_coef3=args.info_max_coef3,
dim_c=args.num_modes,
lr_actor=args.lr_actor,
lr_critic=args.lr_critic,
lr_prior=args.lr_prior,
lr_disc=args.lr_disc,
lr_q=args.lr_q,
auto_lr=args.auto_lr,
epoch_prior=args.epoch_prior,
use_obs_norm=args.use_obs_norm,
obs_horizon=args.obs_horizon,
obs_his_steps=args.obs_his_steps,
begin_weight=args.begin_weight
)
# Path to load model
if args.model_dir:
algo.load_models(args.model_dir)
# Path to save log
log_dir = os.path.join(
'logs', args.env_id, args.algo, f'{args.idx}')
classifier_dir = args.classifier_dir + '{}_{}_modes_classifier/'.format(args.env_id, args.num_modes) + 'model.pth'
# Initialize a trainer
trainer = Trainer(
env=env,
env_test=env_test,
algo=algo,
log_dir=log_dir,
num_steps=args.num_steps,
eval_interval=args.eval_interval,
idx=args.idx,
rend_env=args.rend_env,
classifier_dir=classifier_dir
)
# Start training
trainer.train()
if __name__ == '__main__':
p = argparse.ArgumentParser()
p.add_argument('--idx', type=int, default=-1, help='Training index')
p.add_argument('--seed', type=int, default=9, help='Set a random seed')
p.add_argument('--env_id', type=str, default='Reacher-v4',
help='Environment ID: 2D-Trajectory, Reacher-v4, Pusher-v4, Walker2d-v4, Humanoid-v4')
p.add_argument('--num_modes', type=int, default=6, help='Number of behavioral modes')
p.add_argument('--rend_env', type=bool, default=False, help='Whether to render the environment')
p.add_argument('--buffer_dir', type=str, default='buffers/', help='Path of the buffer')
p.add_argument('--classifier_dir', type=str, default='weights/', help='Path of the weights')
p.add_argument('--model_dir', type=str, default=None, help='Path of the pre-trained model')
p.add_argument('--algo', type=str, default='Ess-InfoGAIL', help='Which algorithm to use')
p.add_argument('--cuda', action='store_true', help='Whether to use GPU')
p.add_argument('--use_obs_norm', action='store_true', help='Whether to normalize the observations')
p.add_argument('--auto_lr', type=bool, default=True, help='Whether to use automatic learning rates')
p.add_argument('--rollout_length', type=int, default=5000)
p.add_argument('--batch_size', type=int, default=1000)
p.add_argument('--num_steps', type=int, default=2000000)
p.add_argument('--eval_interval', type=int, default=200000)
p.add_argument('--obs_horizon', type=int, default=8)
p.add_argument('--obs_his_steps', type=int, default=1)
p.add_argument('--im_ratio', type=int, default=20)
p.add_argument('--surrogate_loss_coef', type=float, default=4.0)
p.add_argument('--disc_grad_penalty', type=float, default=0.1)
p.add_argument('--value_loss_coef', type=float, default=5.0)
p.add_argument('--disc_coef', type=float, default=20)
p.add_argument('--us_coef', type=float, default=1.0)
p.add_argument('--ss_coef', type=float, default=4.0)
p.add_argument('--info_max_coef1', type=float, default=3.0)
p.add_argument('--info_max_coef2', type=float, default=0.05)
p.add_argument('--info_max_coef3', type=float, default=0.5)
p.add_argument('--begin_weight', type=int, default=20)
p.add_argument('--reward_i_coef', type=float, default=1.0)
p.add_argument('--reward_us_coef', type=float, default=0.1)
p.add_argument('--reward_ss_coef', type=float, default=0.1)
p.add_argument('--reward_t_coef', type=float, default=0.005)
p.add_argument('--epoch_ppo', type=int, default=20)
p.add_argument('--epoch_disc', type=int, default=50)
p.add_argument('--epoch_prior', type=int, default=1)
p.add_argument('--lr_actor', type=float, default=3e-3)
p.add_argument('--lr_critic', type=float, default=3e-3)
p.add_argument('--lr_prior', type=float, default=3e-3)
p.add_argument('--lr_disc', type=float, default=5e-3)
p.add_argument('--lr_q', type=float, default=1e-2)
args = p.parse_args()
run(args)