forked from rll-research/BPref
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_PrefPPO.py
194 lines (175 loc) · 8.78 KB
/
train_PrefPPO.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
import gym
import argparse
import yaml
import os
from collections import OrderedDict
from stable_baselines3 import PPO_REWARD
from stable_baselines3.ppo import MlpPolicy
from stable_baselines3.common.env_util import make_vec_dmcontrol_env, make_vec_metaworld_env
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
from stable_baselines3.common.vec_env import VecNormalize
from reward_model import RewardModel
def linear_schedule(initial_value: Union[float, str]) -> Callable[[float], float]:
"""
Linear learning rate schedule.
:param initial_value: (float or str)
:return: (function)
"""
if isinstance(initial_value, str):
initial_value = float(initial_value)
def func(progress_remaining: float) -> float:
"""
Progress will decrease from 1 (beginning) to 0
:param progress_remaining: (float)
:return: (float)
"""
return progress_remaining * initial_value
return func
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="walker_walk", help="environment ID")
parser.add_argument("-tb", "--tensorboard-log", help="Tensorboard log dir", default="logs/PrefPPO/", type=str)
parser.add_argument("--seed", help="Random generator seed", type=int, default=123)
parser.add_argument("--n-envs", help="# of parallel environments", type=int, default=16)
parser.add_argument("--n-steps", help="# of steps to run for each environment per update", type=int, default=500)
parser.add_argument("--lr", help="learning rate", type=float, default=3e-4)
parser.add_argument("--total-timesteps", help="total timesteps", type=int, default=2000000)
parser.add_argument("-b", "--batch-size", help="batch size", type=int, default=64)
parser.add_argument("--ent-coef", help="coeff for entropy", type=float, default=0.0)
parser.add_argument("--hidden-dim", help="dim of hidden features", type=int, default=1024)
parser.add_argument("--num-layer", help="# of layers", type=int, default=2)
parser.add_argument("--use-sde", help="Whether to use generalized State Dependent Exploration", type=int, default=1)
parser.add_argument("--sde-freq", help="Sample a new noise matrix every n steps", type=int, default=4)
parser.add_argument("--gae-lambda", help="Factor for trade-off of bias vs variance", type=float, default=0.92)
parser.add_argument("--clip-init", help="Initial value of clipping", type=float, default=0.4)
parser.add_argument("--n-epochs", help="Number of epoch when optimizing the surrogate loss", type=int, default=20)
parser.add_argument("--normalize", help="Normalization", type=int, default=1)
parser.add_argument("--unsuper-step", help="# of steps for unsupervised learning", type=int, default=32000)
parser.add_argument("--unsuper-n-epochs", help="# of steps for unsupervised learning", type=int, default=50)
# reward learning
parser.add_argument("--re-lr", help="Learning rate of reward fn", type=float, default=0.0003)
parser.add_argument("--re-segment", help="Size of segment", type=int, default=50)
parser.add_argument("--re-act", help="Last activation for reward fn", type=str, default='tanh')
parser.add_argument("--re-num-interaction", help="# of interactions", type=int, default=16000)
parser.add_argument("--re-num-feed", help="# of feedbacks", type=int, default=1)
parser.add_argument("--re-batch", help="Batch size", type=int, default=128)
parser.add_argument("--re-update", help="Gradient update of reward fn", type=int, default=100)
parser.add_argument("--re-feed-type", help="type of feedback", type=int, default=0)
parser.add_argument("--re-large-batch", help="size of buffer for ensemble uncertainty", type=int, default=10)
parser.add_argument("--re-max-feed", help="# of total feedback", type=int, default=1400)
parser.add_argument("--teacher-beta", type=float, default=-1)
parser.add_argument("--teacher-gamma", type=float, default=1.0)
parser.add_argument("--teacher-eps-mistake", type=float, default=0.0)
parser.add_argument("--teacher-eps-skip", type=float, default=0.0)
parser.add_argument("--teacher-eps-equal", type=float, default=0.0)
args = parser.parse_args()
metaworld_flag = False
max_ep_len = 1000
if 'metaworld' in args.env:
metaworld_flag = True
max_ep_len = 500
env_name = args.env
if args.normalize == 1:
args.tensorboard_log += 'normalized_' + env_name
else:
args.tensorboard_log += env_name
args.tensorboard_log += '/teacher_' + str(args.teacher_beta)
args.tensorboard_log += '_' + str(args.teacher_gamma)
args.tensorboard_log += '_' + str(args.teacher_eps_mistake)
args.tensorboard_log += '_' + str(args.teacher_eps_skip)
args.tensorboard_log += '_' + str(args.teacher_eps_equal)
args.tensorboard_log += '/lr_'+str(args.lr)
args.tensorboard_log += '_reward_lr' + str(args.re_lr)
args.tensorboard_log += '_seg' + str(args.re_segment)
args.tensorboard_log += '_act' + str(args.re_act)
args.tensorboard_log += '_inter' + str(args.re_num_interaction)
args.tensorboard_log += '_type' + str(args.re_feed_type)
args.tensorboard_log += '_large' + str(args.re_large_batch)
args.tensorboard_log += '_rebatch' + str(args.re_batch)
args.tensorboard_log += '_reupdate' + str(args.re_update)
args.tensorboard_log += '_batch_' + str(args.batch_size)
args.tensorboard_log += '_nenvs_' + str(args.n_envs)
args.tensorboard_log += '_nsteps_' + str(args.n_steps)
args.tensorboard_log += '_ent_' + str(args.ent_coef)
args.tensorboard_log += '_hidden_' + str(args.hidden_dim)
args.tensorboard_log += '_sde_' + str(args.use_sde)
args.tensorboard_log += '_sdefreq_' + str(args.sde_freq)
args.tensorboard_log += '_gae_' + str(args.gae_lambda)
args.tensorboard_log += '_clip_' + str(args.clip_init)
args.tensorboard_log += '_nepochs_' + str(args.n_epochs)
args.tensorboard_log += '_maxfeed_' + str(args.re_max_feed)
args.tensorboard_log += '_unsuper_' + str(args.unsuper_step)
args.tensorboard_log += '_update_' + str(args.unsuper_n_epochs)
args.tensorboard_log += '_seed_' + str(args.seed)
# extra params
if args.use_sde == 0:
use_sde = False
else:
use_sde = True
clip_range = linear_schedule(args.clip_init)
# Parallel environments
if metaworld_flag:
env = make_vec_metaworld_env(
args.env,
n_envs=args.n_envs,
monitor_dir=args.tensorboard_log,
seed=args.seed)
else:
env = make_vec_dmcontrol_env(
args.env,
n_envs=args.n_envs,
monitor_dir=args.tensorboard_log,
seed=args.seed)
# instantiating the reward model
reward_model = RewardModel(
env.envs[0].observation_space.shape[0],
env.envs[0].action_space.shape[0],
size_segment=args.re_segment,
activation=args.re_act,
lr=args.re_lr,
mb_size=args.re_batch,
teacher_beta=args.teacher_beta,
teacher_gamma=args.teacher_gamma,
teacher_eps_mistake=args.teacher_eps_mistake,
teacher_eps_skip=args.teacher_eps_skip,
teacher_eps_equal=args.teacher_eps_equal,
large_batch=args.re_large_batch)
if args.normalize == 1:
env = VecNormalize(env, norm_reward=False)
# network arch
net_arch = [dict(pi=[args.hidden_dim]*args.num_layer,
vf=[args.hidden_dim]*args.num_layer)]
policy_kwargs = dict(net_arch=net_arch)
# train model
model = PPO_REWARD(
reward_model,
MlpPolicy, env,
tensorboard_log=args.tensorboard_log,
seed=args.seed,
learning_rate=args.lr,
batch_size=args.batch_size,
n_steps=args.n_steps,
ent_coef=args.ent_coef,
policy_kwargs=policy_kwargs,
use_sde=use_sde,
sde_sample_freq=args.sde_freq,
gae_lambda=args.gae_lambda,
clip_range=clip_range,
n_epochs=args.n_epochs,
num_interaction=args.re_num_interaction,
num_feed=args.re_num_feed,
feed_type=args.re_feed_type,
re_update=args.re_update,
metaworld_flag=metaworld_flag,
max_feed=args.re_max_feed,
unsuper_step=args.unsuper_step,
unsuper_n_epochs=args.unsuper_n_epochs,
size_segment=args.re_segment,
max_ep_len=max_ep_len,
verbose=1)
# save args
with open(os.path.join(args.tensorboard_log, "args.yml"), "w") as f:
ordered_args = OrderedDict([(key, vars(args)[key]) for key in sorted(vars(args).keys())])
yaml.dump(ordered_args, f)
model.learn(total_timesteps=args.total_timesteps, unsuper_flag=1)
model.reward_model.save(args.tensorboard_log, args.total_timesteps)