-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
175 lines (150 loc) · 6.95 KB
/
train.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
import time
import numpy as np
from collections import deque
import common.utils as utils
def evaluate(env_name, test_env, agent, logger, video, num_eval_episodes, start_time, action_repeat, num_episode, step):
def eval(env_name, num, env, agent, logger, video, num_episodes, step, start_time, action_repeat, num_episode):
all_ep_rewards = []
all_ep_length = []
do_carla_metrics = False
if env_name.startswith('carla'):
# carla metrics:
reason_each_episode_ended = []
distance_driven_each_episode = []
crash_intensity = 0.
steer = 0.
brake = 0.
count = 0
do_carla_metrics = True
# loop num_episodes
for episode in range(num_episodes):
obs = env.reset()
dist_driven_this_episode = 0.
video.init(enabled=(episode == 0))
done, episode_reward, episode_length = False, 0, 0
# evaluate once
while not done:
with utils.eval_mode(agent):
action = agent.select_action(obs, deterministic=True)
obs, reward, done, info = env.step(action)
episode_reward += reward
episode_length += 1
# metrics:
if do_carla_metrics:
dist_driven_this_episode += info['distance']
crash_intensity += info['crash_intensity']
steer += abs(info['steer'])
brake += info['brake']
count += 1
video.record(env)
video.save('%s-Env%d.mp4' % (step * action_repeat, num))
# record the score
all_ep_rewards.append(episode_reward)
all_ep_length.append(episode_length)
# record log
# metrics:
if do_carla_metrics:
reason_each_episode_ended.append(info['reason_episode_ended'])
distance_driven_each_episode.append(dist_driven_this_episode)
mean, std, best = np.mean(all_ep_rewards), np.std(all_ep_rewards), np.max(all_ep_rewards)
if do_carla_metrics:
print('METRICS--------------------------')
print("reason_each_episode_ended: {}".format(reason_each_episode_ended))
print("distance: {}".format(distance_driven_each_episode))
print('crash_intensity: {}'.format(crash_intensity / count))
print('steer: {}'.format(steer / count))
print('brake: {}'.format(brake / count))
print('---------------------------------')
test_info = {
("TestEpRet%d" % num): mean,
("TestStd%d" % num): std,
("TestBestEpLen%d" % num): best,
"DistanceEp": np.mean(distance_driven_each_episode),
"Crash_intensity" : crash_intensity / count,
"Steer" : steer / count,
"Brake" : brake / count,
'Episode': num_episode,
'Time': (time.time() - start_time) / 3600,
}
else:
test_info = {
("TestEpRet%d" % num): mean,
("TestStd%d" % num): std,
("TestBestEpLen%d" % num): best,
'Episode': num_episode,
'Time': (time.time() - start_time) / 3600,
}
utils.log(logger, 'eval_agent', test_info, step)
#video.save('%d.mp4' % step)
if isinstance(test_env, list):
# test_env is a list
for num, t_env in enumerate(test_env):
eval(env_name, num, t_env, agent, logger, video, num_eval_episodes,
step, start_time, action_repeat, num_episode)
else:
# test_env is an environment
eval(env_name, 0, test_env, agent, logger, video, num_eval_episodes,
step, start_time, action_repeat, num_episode)
def train_agent(env_name, train_envs, test_env, agent, replay_buffer, logger, video, model_dir,
total_steps, init_steps, eval_freq, action_repeat, num_updates,
num_eval_episodes, test, save_model, save_model_freq, device, **kwargs):
epoch_start_times = start_time = time.time()
env_id, num_sources, best_return = 0, len(train_envs), 0.0
episode, episode_reward, episode_step, done = 0, 0, 0, False
env = train_envs[env_id]
o = env.reset() #(9,,)
replay_buffer.add_obs(o)
# import pdb
# pdb.set_trace()
for step in range(1, total_steps + 1):
# sample action for data collection
if step < init_steps:
a = env.action_space.sample()
else:
with utils.eval_mode(agent):
a = agent.select_action(o)
# run training update
if step >= init_steps and step % num_updates == 0:
for _ in range(num_updates):
agent.update(replay_buffer, logger, step, step % 500 == 0)
o2, r, done, infos = env.step(a)
agent.total_time_steps += 1
episode_reward += r
episode_step += 1
# allow infinit bootstrap
d_bool = 0 if episode_step == env._max_episode_steps else float(done)
replay_buffer.add(o, a, r, o2, done, d_bool, episode_step, env_id)
o = o2
if done:
train_info = dict(EpRet=episode_reward, EpLen=episode_step, EpNum=episode)
utils.log(logger, 'train_agent', train_info, step)
# logger['tb'].dump(step)
print("Total T: {} Reward: {:.3f} Episode Num: {} Episode T: {} Time: {}".format(
agent.total_time_steps, episode_reward, episode, episode_step, utils.calc_time(start_time)))
if best_return < episode_reward:
best_return = episode_reward
agent.save(model_dir, 'best_in_env_%d' % env_id)
episode += 1
env_id = episode % num_sources
env = train_envs[env_id]
o, done, episode_reward, episode_step = env.reset(), False, 0, 0
replay_buffer.add_obs(o)
# evaluate agent periodically
if step % eval_freq == 0 and step > init_steps and test:
evaluate(env_name, test_env, agent, logger, video, num_eval_episodes,
start_time, action_repeat, episode, step)
print('Update Extr Times: %s Update Critic Times: %d Update Actor Times: %d' % (
agent.update_extr_total_steps, agent.update_critic_steps, agent.update_actor_steps
))
agent.print_log(logger['sp'],
test_env,
step // eval_freq,
step,
action_repeat,
test,
start_time,
eval_freq * action_repeat / (time.time() - epoch_start_times))
if save_model and step == total_steps:
agent.save(model_dir, step)
# if save_model and step // eval_freq == save_model_freq:
# agent.save(model_dir, step)