-
Notifications
You must be signed in to change notification settings - Fork 3
/
run_new_dqn_phase.py
159 lines (128 loc) · 6.08 KB
/
run_new_dqn_phase.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
import gym
from environment import TSCEnv
from world import World
from generator import PhaseVehicleGenerator, LaneVehicleGenerator,PressureRewardGenerator
from agent.new_dqn_agent import DQNAgent
from metric import TravelTimeMetric, ThroughputMetric, SpeedScoreMetric,MaxWaitingTimeMetric
import argparse
import os
import numpy as np
import logging
from datetime import datetime
import utils as u
import ntpath
# parse args
parser = argparse.ArgumentParser(description='Run Example')
parser.add_argument('config_file', type=str, help='path of config file')
parser.add_argument('--thread', type=int, default=1, help='number of threads')
parser.add_argument('--steps', type=int, default=3600, help='number of steps')
#parser.add_argument('--action_interval', type=int, default=20, help='how often agent make decisions')
#parser.add_argument('--episodes', type=int, default=200, help='training episodes')
parser.add_argument('--save_model', action="store_true", default=False)
parser.add_argument('--load_model', action="store_true", default=False)
parser.add_argument("--save_rate", type=int, default=20, help="save model once every time this many episodes are completed")
parser.add_argument('--save_dir', type=str, default="model/dqn", help='directory in which model should be saved')
parser.add_argument('--log_dir', type=str, default="log/dqn", help='directory in which logs should be saved')
parser.add_argument('--parameters', type=str, default="agent/configs_new_dqn/default.json", help='path to the file with informations about the model')
args = parser.parse_args()
if not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
logger = logging.getLogger('main')
logger.setLevel(logging.DEBUG)
file_name_time = f"{datetime.now().strftime('%Y%m%d-%H%M%S')}"
file_name = f"{args.log_dir}/{file_name_time}"
fh = logging.FileHandler(file_name+'.log')
fh.setLevel(logging.DEBUG)
sh = logging.StreamHandler()
sh.setLevel(logging.INFO)
logger.addHandler(fh)
logger.addHandler(sh)
#Config File
parameters = u.get_info_file(args.parameters)
episodes = parameters['episodes']
parameters['log_path'] = args.log_dir
action_interval = parameters['action_interval']
#start wandb
u.wand_init("TLC - Results C2",f"new_dqn ts: {ntpath.basename(args.parameters)[:-5]}", "new_dqn ts")
# create world
world = World(args.config_file, thread_num=args.thread)
# create agents
agents = []
for i in world.intersections:
action_space = gym.spaces.Discrete(len(i.phases))
agents.append(DQNAgent(
action_space,
PhaseVehicleGenerator(world, i, ["phase_vehicles"], in_only=True, average=None, scale=0.0125),
PressureRewardGenerator(world, i, scale=.005, negative=True),
i.id,
parameters,
world
))
if args.load_model:
agents[-1].load_model(args.save_dir)
main_agent = agents[0]
# Create metric
metric = [TravelTimeMetric(world), ThroughputMetric(world), SpeedScoreMetric(world), MaxWaitingTimeMetric(world)]
# create env
env = TSCEnv(world, agents, metric)
# train dqn_agent
def train(args, env):
total_decision_num = 0
for e in range(episodes):
last_obs = env.reset()
if e % args.save_rate == args.save_rate - 1:
env.eng.set_save_replay(True)
env.eng.set_replay_file("replay_%s.txt" % e)
else:
env.eng.set_save_replay(False)
episodes_rewards = [0 for i in agents]
episodes_decision_num = 0
i = 0
while i < args.steps:
if i % action_interval == 0:
actions = []
for agent_id, agent in enumerate(agents):
if total_decision_num > agent.learning_start:
actions.append(main_agent.get_action(last_obs[agent_id]))
else:
actions.append(main_agent.sample())
rewards_list = []
for _ in range(action_interval):
obs, rewards, dones, _ = env.step(actions)
i += 1
rewards_list.append(rewards)
rewards = np.mean(rewards_list, axis=0)
for agent_id, agent in enumerate(agents):
u.append_new_line(file_name+f"_{agent_id}",[[last_obs[agent_id],-1], actions[agent_id], rewards[agent_id], [obs[agent_id],-1],e,i])
main_agent.remember(last_obs[agent_id], actions[agent_id], rewards[agent_id], obs[agent_id])
episodes_rewards[agent_id] += rewards[agent_id]
episodes_decision_num += 1
total_decision_num += 1
last_obs = obs
#for agent_id, agent in enumerate(agents):
if total_decision_num > main_agent.learning_start and total_decision_num % main_agent.update_model_freq == main_agent.update_model_freq - 1:
main_agent.replay()
if total_decision_num > main_agent.learning_start and total_decision_num % main_agent.update_target_model_freq == main_agent.update_target_model_freq - 1:
main_agent.update_target_network()
if all(dones):
break
if e % args.save_rate == args.save_rate - 1:
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
main_agent.save_model(args.save_dir)
eval_dict = {}
logger.info(f"episode:{e}/{episodes-1}, steps:{i}")
eval_dict["episode"]=e
eval_dict["steps"]=i
for agent_id, agent in enumerate(agents):
logger.info("\tagent:{}, mean_episode_reward:{}".format(agent_id, episodes_rewards[agent_id] / episodes_decision_num))
for metric in env.metric:
logger.info(f"\t{metric.name}: {metric.eval()}")
eval_dict[metric.name]=metric.eval()
eval_dict["epsilon"]=main_agent.epsilon
eval_dict["mean_episode_reward"]=episodes_rewards[0] / episodes_decision_num
u.wand_log(eval_dict)
#for agent in agents:
main_agent.save_model(args.save_dir)
if __name__ == '__main__':
train(args, env)