-
Notifications
You must be signed in to change notification settings - Fork 173
/
main.py
145 lines (134 loc) · 6.62 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
import argparse
import torch
import os
import numpy as np
from gym.spaces import Box, Discrete
from pathlib import Path
from torch.autograd import Variable
from tensorboardX import SummaryWriter
from utils.make_env import make_env
from utils.buffer import ReplayBuffer
from utils.env_wrappers import SubprocVecEnv, DummyVecEnv
from algorithms.attention_sac import AttentionSAC
def make_parallel_env(env_id, n_rollout_threads, seed):
def get_env_fn(rank):
def init_env():
env = make_env(env_id, discrete_action=True)
env.seed(seed + rank * 1000)
np.random.seed(seed + rank * 1000)
return env
return init_env
if n_rollout_threads == 1:
return DummyVecEnv([get_env_fn(0)])
else:
return SubprocVecEnv([get_env_fn(i) for i in range(n_rollout_threads)])
def run(config):
model_dir = Path('./models') / config.env_id / config.model_name
if not model_dir.exists():
run_num = 1
else:
exst_run_nums = [int(str(folder.name).split('run')[1]) for folder in
model_dir.iterdir() if
str(folder.name).startswith('run')]
if len(exst_run_nums) == 0:
run_num = 1
else:
run_num = max(exst_run_nums) + 1
curr_run = 'run%i' % run_num
run_dir = model_dir / curr_run
log_dir = run_dir / 'logs'
os.makedirs(log_dir)
logger = SummaryWriter(str(log_dir))
torch.manual_seed(run_num)
np.random.seed(run_num)
env = make_parallel_env(config.env_id, config.n_rollout_threads, run_num)
model = AttentionSAC.init_from_env(env,
tau=config.tau,
pi_lr=config.pi_lr,
q_lr=config.q_lr,
gamma=config.gamma,
pol_hidden_dim=config.pol_hidden_dim,
critic_hidden_dim=config.critic_hidden_dim,
attend_heads=config.attend_heads,
reward_scale=config.reward_scale)
replay_buffer = ReplayBuffer(config.buffer_length, model.nagents,
[obsp.shape[0] for obsp in env.observation_space],
[acsp.shape[0] if isinstance(acsp, Box) else acsp.n
for acsp in env.action_space])
t = 0
for ep_i in range(0, config.n_episodes, config.n_rollout_threads):
print("Episodes %i-%i of %i" % (ep_i + 1,
ep_i + 1 + config.n_rollout_threads,
config.n_episodes))
obs = env.reset()
model.prep_rollouts(device='cpu')
for et_i in range(config.episode_length):
# rearrange observations to be per agent, and convert to torch Variable
torch_obs = [Variable(torch.Tensor(np.vstack(obs[:, i])),
requires_grad=False)
for i in range(model.nagents)]
# get actions as torch Variables
torch_agent_actions = model.step(torch_obs, explore=True)
# convert actions to numpy arrays
agent_actions = [ac.data.numpy() for ac in torch_agent_actions]
# rearrange actions to be per environment
actions = [[ac[i] for ac in agent_actions] for i in range(config.n_rollout_threads)]
next_obs, rewards, dones, infos = env.step(actions)
replay_buffer.push(obs, agent_actions, rewards, next_obs, dones)
obs = next_obs
t += config.n_rollout_threads
if (len(replay_buffer) >= config.batch_size and
(t % config.steps_per_update) < config.n_rollout_threads):
if config.use_gpu:
model.prep_training(device='gpu')
else:
model.prep_training(device='cpu')
for u_i in range(config.num_updates):
sample = replay_buffer.sample(config.batch_size,
to_gpu=config.use_gpu)
model.update_critic(sample, logger=logger)
model.update_policies(sample, logger=logger)
model.update_all_targets()
model.prep_rollouts(device='cpu')
ep_rews = replay_buffer.get_average_rewards(
config.episode_length * config.n_rollout_threads)
for a_i, a_ep_rew in enumerate(ep_rews):
logger.add_scalar('agent%i/mean_episode_rewards' % a_i,
a_ep_rew * config.episode_length, ep_i)
if ep_i % config.save_interval < config.n_rollout_threads:
model.prep_rollouts(device='cpu')
os.makedirs(run_dir / 'incremental', exist_ok=True)
model.save(run_dir / 'incremental' / ('model_ep%i.pt' % (ep_i + 1)))
model.save(run_dir / 'model.pt')
model.save(run_dir / 'model.pt')
env.close()
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("env_id", help="Name of environment")
parser.add_argument("model_name",
help="Name of directory to store " +
"model/training contents")
parser.add_argument("--n_rollout_threads", default=12, type=int)
parser.add_argument("--buffer_length", default=int(1e6), type=int)
parser.add_argument("--n_episodes", default=50000, type=int)
parser.add_argument("--episode_length", default=25, type=int)
parser.add_argument("--steps_per_update", default=100, type=int)
parser.add_argument("--num_updates", default=4, type=int,
help="Number of updates per update cycle")
parser.add_argument("--batch_size",
default=1024, type=int,
help="Batch size for training")
parser.add_argument("--save_interval", default=1000, type=int)
parser.add_argument("--pol_hidden_dim", default=128, type=int)
parser.add_argument("--critic_hidden_dim", default=128, type=int)
parser.add_argument("--attend_heads", default=4, type=int)
parser.add_argument("--pi_lr", default=0.001, type=float)
parser.add_argument("--q_lr", default=0.001, type=float)
parser.add_argument("--tau", default=0.001, type=float)
parser.add_argument("--gamma", default=0.99, type=float)
parser.add_argument("--reward_scale", default=100., type=float)
parser.add_argument("--use_gpu", action='store_true')
config = parser.parse_args()
run(config)