-
Notifications
You must be signed in to change notification settings - Fork 0
/
callee_SAC_norm.py
226 lines (195 loc) · 8.6 KB
/
callee_SAC_norm.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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import os
import torch
import argparse
import numpy as np
from copy import deepcopy
from typing import Optional, Tuple
from torch.utils.tensorboard import SummaryWriter
from tianshou.utils import TensorboardLogger, WandbLogger
from tianshou.env import SubprocVectorEnv
from tianshou.utils.net.common import Net
from tianshou.trainer import offpolicy_trainer, OffpolicyTrainer
from tianshou.data import ReplayBuffer
from tianshou.policy import BasePolicy, DQNPolicy
from tianshou.policy import DiscreteSACPolicy
from enviroment.briscola_gym.briscola_features import BriscolaEnv
from tianshou.env import PettingZooEnv
import shortuuid
from agents.pg import PGPolicy
from agents.random import RandomPolicy
from agents.ac import Actor, Critic
from multi_agents_rl.buffer import MultiAgentVectorReplayBuffer
from multi_agents_rl.collector import MultiAgentCollector
from multi_agents_rl.mapolicy import MultiAgentPolicyManager
import gymnasium
from multi_agents_rl.observation_wrapper import VectorEnvNormObs
def env_func():
return PettingZooEnv(BriscolaEnv(use_role_ids=True, normalize_reward=False))
def get_agent(args, is_fixed=False):
net = Net(args.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
actor = Actor(net, args.action_shape, softmax_output=False,
device=args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(args.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
critic1 = Critic(net_c1, last_size=args.action_shape,
device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(args.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
critic2 = Critic(net_c2, last_size=args.action_shape,
device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
if args.auto_alpha:
target_entropy = 0.98 * np.log(np.prod(args.action_shape))
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = DiscreteSACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
estimation_step=args.n_step,
reward_normalization=args.rew_norm
).to(args.device)
return policy
def get_random_agent(args):
agent = RandomPolicy(
observation_space=args.state_shape,
action_space=args.action_shape,
device=args.device
)
return agent
def selfplay(args): # always train first agent, start from random policy
train_envs = SubprocVectorEnv(
[env_func for _ in range(args.num_parallel_env)])
test_envs = SubprocVectorEnv(
[env_func for _ in range(args.num_parallel_env)])
train_envs = VectorEnvNormObs(train_envs)
test_envs = VectorEnvNormObs(test_envs, update_obs_rms=False)
test_envs.set_obs_rms(train_envs.get_obs_rms())
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# Env
briscola = BriscolaEnv(use_role_ids=True, normalize_reward=False)
env = PettingZooEnv(briscola)
args.state_shape = briscola.observation_space_shape
args.action_shape = briscola.action_space_shape
# initialize agents and ma-policy
id_agent_learning = 0
callee = get_agent(args)
agents = [get_random_agent(args), callee,
get_random_agent(args), get_random_agent(args), get_random_agent(args)]
# {'caller': 0, 'callee': 1, 'good_1': 2, 'good_2': 3, 'good_3': 4}
LERNING_AGENTS_ID = ['callee']
policy = MultiAgentPolicyManager(agents, env, LERNING_AGENTS_ID)
# collector
train_collector = MultiAgentCollector(
policy, train_envs, LERNING_AGENTS_ID,
MultiAgentVectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=False)
test_collector = MultiAgentCollector(
policy, test_envs, LERNING_AGENTS_ID, exploration_noise=False)
# Log
args.algo_name = "SAC_callee_vs_random NORM"
log_name = os.path.join(
args.algo_name, f'{str(args.seed)}_{shortuuid.uuid()[:8]}')
log_path = os.path.join(args.logdir, log_name)
os.makedirs(log_path, exist_ok=True)
print(f'Saving results in path: {log_path}')
print('-----------------------------------')
if args.logger == "wandb":
logger = WandbLogger(
train_interval=1,
update_interval=1,
save_interval=1,
name=log_name.replace(os.path.sep, "__"),
run_id=args.resume_id,
config=args,
project=args.wandb_project,
)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
if args.logger == "tensorboard":
logger = TensorboardLogger(writer)
else: # wandb
logger.load(writer)
def save_best_fn(policy):
model_save_path = os.path.join(
log_path, f'callee_policy_best_epoch.pth')
torch.save(policy.policies['callee'].state_dict(), model_save_path)
def reward_metric(rews):
return rews[:, id_agent_learning]
trainer = OffpolicyTrainer(policy,
train_collector,
test_collector,
max_epoch=args.epoch,
step_per_epoch=args.step_per_epoch,
step_per_collect=0, # NOTE this is keep but ignore by our collector
episode_per_collect=args.episode_per_collect,
episode_per_test=args.test_num,
batch_size=args.batch_size,
save_best_fn=save_best_fn,
reward_metric=reward_metric,
logger=logger,
update_per_step=args.update_per_step,
test_in_train=False)
trainer.run()
model_save_path = os.path.join(log_path, 'callee_policy_last_epoch.pth')
torch.save(policy.policies['callee'].state_dict(), model_save_path)
def get_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=42)
NUM_GAMES = 15000
parser.add_argument('--buffer-size', type=int, default=8*NUM_GAMES)
parser.add_argument('--actor-lr', type=float, default=1e-5) # 1e-4
parser.add_argument('--critic-lr', type=float, default=1e-4) # 1e-3
parser.add_argument('--alpha-lr', type=float, default=3e-4)
parser.add_argument('--gamma', type=float, default=1.0)
parser.add_argument('--tau', type=float, default=0.005)
parser.add_argument('--alpha', type=float, default=0.05)
parser.add_argument('--auto-alpha', action="store_true", default=False)
parser.add_argument('--epoch', type=int, default=5)
parser.add_argument('--step-per-epoch', type=int, default=5*8*NUM_GAMES)
parser.add_argument('--episode-per-collect', type=int, default=NUM_GAMES)
parser.add_argument('--update-per-step', type=float,
default=0.01) # NOTE 8*1500 gradient steps
parser.add_argument('--batch-size', type=int, default=512)
parser.add_argument('--hidden-sizes', type=int,
nargs='*', default=[256, 256])
parser.add_argument('--num-parallel-env', type=int, default=20)
parser.add_argument('--rew-norm', action="store_true", default=False)
parser.add_argument('--n-step', type=int, default=1)
parser.add_argument('--num-generation', type=int, default=200)
parser.add_argument('--test-num', type=int, default=30000)
parser.add_argument('--logdir', type=str,
default='log/')
parser.add_argument('--render', type=float, default=0.1)
parser.add_argument(
'--device', type=str, default='cuda' if torch.cuda.is_available() else 'cpu'
)
parser.add_argument(
"--logger",
type=str,
default="tensorboard",
choices=["tensorboard", "wandb"],
)
parser.add_argument("--wandb-project", type=str, default="FORL_briscola")
parser.add_argument("--resume-id", type=str, default=None)
return parser
def get_args() -> argparse.Namespace:
parser = get_parser()
return parser.parse_known_args()[0]
# train the agent and watch its performance in a match!
args = get_args()
selfplay(args)