forked from rll-research/BPref
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_SAC.py
executable file
·188 lines (156 loc) · 6.63 KB
/
train_SAC.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
#!/usr/bin/env python3
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import copy
import math
import os
import sys
import time
import pickle as pkl
import utils
import hydra
from logger import Logger
from replay_buffer import ReplayBuffer
class Workspace(object):
def __init__(self, cfg):
self.work_dir = os.getcwd()
print(f'workspace: {self.work_dir}')
self.cfg = cfg
self.logger = Logger(self.work_dir,
save_tb=cfg.log_save_tb,
log_frequency=cfg.log_frequency,
agent=cfg.agent.name)
utils.set_seed_everywhere(cfg.seed)
self.device = torch.device(cfg.device)
self.log_success = False
self.step = 0
# make env
if 'metaworld' in cfg.env:
self.env = utils.make_metaworld_env(cfg)
self.log_success = True
else:
self.env = utils.make_env(cfg)
cfg.agent.params.obs_dim = self.env.observation_space.shape[0]
cfg.agent.params.action_dim = self.env.action_space.shape[0]
cfg.agent.params.action_range = [
float(self.env.action_space.low.min()),
float(self.env.action_space.high.max())
]
self.agent = hydra.utils.instantiate(cfg.agent)
# no relabel
self.replay_buffer = ReplayBuffer(
self.env.observation_space.shape,
self.env.action_space.shape,
int(cfg.replay_buffer_capacity),
self.device)
meta_file = os.path.join(self.work_dir, 'metadata.pkl')
pkl.dump({'cfg': self.cfg}, open(meta_file, "wb"))
def evaluate(self):
average_episode_reward = 0
if self.log_success:
success_rate = 0
for episode in range(self.cfg.num_eval_episodes):
obs = self.env.reset()
self.agent.reset()
done = False
episode_reward = 0
if self.log_success:
episode_success = 0
while not done:
with utils.eval_mode(self.agent):
action = self.agent.act(obs, sample=False)
obs, reward, done, extra = self.env.step(action)
episode_reward += reward
if self.log_success:
episode_success = max(episode_success, extra['success'])
average_episode_reward += episode_reward
if self.log_success:
success_rate += episode_success
average_episode_reward /= self.cfg.num_eval_episodes
if self.log_success:
success_rate /= self.cfg.num_eval_episodes
success_rate *= 100.0
self.logger.log('eval/episode_reward', average_episode_reward,
self.step)
if self.log_success:
self.logger.log('eval/success_rate', success_rate,
self.step)
self.logger.dump(self.step)
def run(self):
episode, episode_reward, done = 0, 0, True
if self.log_success:
episode_success = 0
start_time = time.time()
fixed_start_time = time.time()
while self.step < self.cfg.num_train_steps:
if done:
if self.step > 0:
self.logger.log('train/duration',
time.time() - start_time, self.step)
self.logger.log('train/total_duration',
time.time() - fixed_start_time, self.step)
start_time = time.time()
self.logger.dump(
self.step, save=(self.step > self.cfg.num_seed_steps))
# evaluate agent periodically
if self.step > 0 and self.step % self.cfg.eval_frequency == 0:
self.logger.log('eval/episode', episode, self.step)
self.evaluate()
self.logger.log('train/episode_reward', episode_reward,
self.step)
if self.log_success:
self.logger.log('train/episode_success', episode_success,
self.step)
obs = self.env.reset()
self.agent.reset()
done = False
episode_reward = 0
if self.log_success:
episode_success = 0
episode_step = 0
episode += 1
self.logger.log('train/episode', episode, self.step)
# sample action for data collection
if self.step < self.cfg.num_seed_steps:
action = self.env.action_space.sample()
else:
with utils.eval_mode(self.agent):
action = self.agent.act(obs, sample=True)
# run training update
if self.step == (self.cfg.num_seed_steps + self.cfg.num_unsup_steps) and self.cfg.num_unsup_steps > 0:
# reset Q due to unsuperivsed exploration
self.agent.reset_critic()
# update agent
self.agent.update_after_reset(
self.replay_buffer, self.logger, self.step,
gradient_update=self.cfg.reset_update,
policy_update=True)
elif self.step > self.cfg.num_seed_steps + self.cfg.num_unsup_steps:
self.agent.update(self.replay_buffer, self.logger, self.step)
# unsupervised exploration
elif self.step > self.cfg.num_seed_steps:
self.agent.update_state_ent(self.replay_buffer, self.logger, self.step,
gradient_update=1, K=self.cfg.topK)
next_obs, reward, done, extra = self.env.step(action)
# allow infinite bootstrap
done = float(done)
done_no_max = 0 if episode_step + 1 == self.env._max_episode_steps else done
episode_reward += reward
if self.log_success:
episode_success = max(episode_success, extra['success'])
self.replay_buffer.add(
obs, action,
reward, next_obs, done,
done_no_max)
obs = next_obs
episode_step += 1
self.step += 1
self.agent.save(self.work_dir, self.step)
@hydra.main(config_path='config/train.yaml', strict=True)
def main(cfg):
workspace = Workspace(cfg)
workspace.run()
if __name__ == '__main__':
main()