-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
193 lines (144 loc) · 6.25 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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
"""
File này dùng để training model
"""
from .agent.base_agent import Agent, RandomAgent, PretrainedAgent
from utils.memory import StateMemory, ReplayBuffer
from agent.DQL_agent import DQNAgent
from agent.Qmix_agent import QMIXAgent
from magent2.environments import battle_v4
import os
import torch
from torch.utils.data import DataLoader
from time import time
import argparse
class Trainer :
"""
Sử dụng blue để huấn luyện
"""
def __init__(self, env, red_agent: Agent, blue_agent:Agent, buffer, batch_size = 64, is_self_play = False):
self.red_agent = red_agent
self.blue_agent = blue_agent
self.buffer = buffer
self.batch_size = batch_size
self.env = env
self.is_self_play = is_self_play
def agent_give_action(self, name: str, observation):
if self.is_self_play :
return self.blue_agent.get_action(observation)
if name == "blue":
return self.blue_agent.get_action(observation)
return self.red_agent.get_action(observation)
def update_memory(self, is_longterm: bool = False):
"""
Tạo ra một vòng lặp lưu trữ và cập nhật dữ liệu cho từng agent
"""
self.env.reset()
prev_obs = {}
prev_actions = {}
red_reward = 0
blue_reward = 0
prev_team = "red"
n_kills = {"red": 0, "blue": 0}
# vong lap 1
for idx, agent in enumerate(self.env.agent_iter()):
prev_ob, reward, termination, truncation, _ = self.env.last()
team = agent.split("_")[0]
n_kills[team] += (reward > 4.5)
if truncation or termination:
prev_action = None
else:
if agent.split("_")[0] == "red":
prev_action = self.agent_give_action("red", prev_ob)
red_reward += reward
else:
prev_action = self.agent_give_action("blue", prev_ob)
blue_reward += reward
prev_obs[agent] = prev_ob
prev_actions[agent] = prev_action
self.env.step(prev_action)
if (idx + 1) % self.env.num_agents == 0: break
# vong lap 2
for agent in self.env.agent_iter():
obs, reward, termination, truncation, _ = self.env.last()
team = agent.split("_")[0]
n_kills[team] += (reward > 4.5)
if truncation or termination:
action = None
else:
if agent.split("_")[0] == "red" :
action = self.agent_give_action("red", obs)
red_reward += reward
else:
action = self.agent_give_action("blue", obs)
blue_reward += reward
self.env.step(action)
if isinstance(self.buffer, StateMemory):
if team != prev_team :
self.buffer.ensemble()
prev_team = team
idx = int(agent.split("_")[1]) % self.buffer.grouped_agents
self.buffer.push(
idx,
prev_obs[agent],
prev_actions[agent],
reward,
obs,
termination
)
else:
self.buffer.push(
prev_obs[agent],
prev_actions[agent],
reward,
obs,
termination
)
prev_obs[agent] = obs
prev_actions[agent] = action
return blue_reward - red_reward, n_kills, blue_reward # red thắng
def save_model (self, file_path):
torch.save(self.blue_agent.q_net.state_dict(), file_path)
print(f"Model saved to {file_path}")
def train(self, episodes=500, target_update_freq=2, is_type = "dqn"):
gap_rewards = []
for eps in range(episodes):
start = time()
gap_reward, n_kills, blue_reward = self.update_memory()
if is_type == "qmix":
self.buffer.ensemble()
dataloader = DataLoader(self.buffer, batch_size = self.batch_size, shuffle = True)
# print(f"Out of dataloader {len(self.buffer)}")
self.blue_agent.train(dataloader)
self.blue_agent.decay_epsilon()
if eps % target_update_freq == 0:
self.blue_agent.update_target_network()
end = time() - start
# wandb.log({
# "episode": eps,
# "gap_rewards": gap_reward,
# "epsilon": self.blue_agent.epsilon,
# "time": end,
# "red_kill": n_kills["red"],
# "blue_kill": n_kills["blue"]
# })
gap_rewards.append(gap_reward)
print(f"Episode {eps}, Gap Reward: {gap_reward}, Total Reward: {blue_reward}, Epsilon: {self.blue_agent.epsilon:.2f}, Time: {end}, Kill: {n_kills}")
self.env.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Train a Double Deep Q for MAgent")
parser.add_argument("-mode", type=str, required=True, help="self-play if you want to train with self-play, otherwise random")
parser.add_argument("-save_dir", type=str, required=True, help="Path to save model")
args = parser.parse_args()
is_self_play = args.is_self_play == "self-play"
save_dir = args.save_dir
env = battle_v4.env(map_size=45, render_mode="rgb_array", attack_opponent_reward=0.5)
device = "cuda" if torch.cuda.is_available() else "cpu"
observation_shape = env.observation_space("red_0").shape
action_shape = env.action_space("red_0").n
num_agents = 27
blue_agent = DQNAgent(observation_shape,action_shape, device=device)
red_agent = RandomAgent(action_shape)
buffer = ReplayBuffer(capacity=10000)
trainer = Trainer(env, red_agent, blue_agent, buffer, batch_size = 64, is_self_play=is_self_play)
trainer.train(episodes = 70)
trainer.save_model(save_dir)