-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathddpg.py
172 lines (140 loc) · 6.87 KB
/
ddpg.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
import argparse
import time
import torch
import torch.nn.functional as F
import numpy as np
import tqdm
import gym
import utils
import run
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def exploration_noise(action, random_process, eps):
return action + eps*random_process.sample().astype(np.float32)
def soft_update(target, source, tau):
for target_param, param in zip(target.parameters(), source.parameters()):
target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
def hard_update(target, source):
for target_param, param in zip(target.parameters(), source.parameters()):
target_param.data.copy_(param.data)
def ddpg(agent, env, args):
agent.to(device)
# initialize target networks
target_agent = type(agent)()
target_agent.to(device)
hard_update(target_agent.actor, agent.actor)
hard_update(target_agent.critic, agent.critic)
random_process = utils.OrnsteinUhlenbeckProcess(size=env.action_space.shape, sigma=args.sigma, theta=args.theta)
eps = args.eps_start
buffer = utils.ReplayBuffer(args.buffer_size)
critic_optimizer = torch.optim.Adam(agent.critic.parameters(), lr=args.critic_lr)
actor_optimizer = torch.optim.Adam(agent.actor.parameters(), lr=args.actor_lr)
save_dir = utils.make_process_dirs('ddpg_run')
# use warmp up steps to add random transitions to the buffer
state = env.reset()
done = False
for _ in range(args.warmup_steps):
if done: state = env.reset(); done = False
rand_action = env.action_space.sample()
next_state, reward, done, info = env.step(rand_action)
buffer.push(state, rand_action, reward, next_state, done)
state = next_state
for episode in range(args.num_episodes):
state = env.reset()
random_process.reset_states()
done = False
for step in range(args.max_episode_steps):
if done: break
# collect new experience
action = agent(state)
noisy_action = exploration_noise(action, random_process, eps)
next_state, reward, done, info = env.step(noisy_action)
if args.render: env.render()
buffer.push(state, noisy_action, reward, next_state, done)
state = next_state
batch = buffer.sample(args.batch_size)
# batch will be None if not enough experience has been collected yet
if not batch:
continue
# prepare transitions for models
state_batch, action_batch, reward_batch, next_state_batch, done_batch = zip(*batch)
cat_tuple = lambda t : torch.cat(t).to(device)
list_to_tensor = lambda t : torch.tensor(t).unsqueeze(0).to(device)
state_batch = cat_tuple(state_batch)
next_state_batch = cat_tuple(next_state_batch)
action_batch = cat_tuple(action_batch)
reward_batch = list_to_tensor(reward_batch).T
done_batch = list_to_tensor(done_batch).T
# critic update
target_action_s2 = target_agent.actor(next_state_batch)
target_action_value_s2 = target_agent.critic(next_state_batch, target_action_s2)
td_target = reward_batch + args.gamma*(1.-done_batch)*target_action_value_s2
agent_critic_pred = agent.critic(state_batch, action_batch)
critic_loss = F.mse_loss(td_target, agent_critic_pred)
critic_optimizer.zero_grad()
critic_loss.backward()
critic_optimizer.step()
# actor update
agent_actions = agent.actor(state_batch)
actor_loss = -agent.critic(state_batch, agent_actions).mean()
actor_optimizer.zero_grad()
actor_loss.backward()
actor_optimizer.step()
# move target model towards training model
soft_update(target_agent.actor, agent.actor, args.tau)
soft_update(target_agent.critic, agent.critic, args.tau)
eps = max(args.eps_final, eps - (args.eps_start - args.eps_final)/args.eps_anneal)
if episode % args.eval_interval == 0:
agent.eval()
returns = run.run(agent, env, args.eval_episodes, args.max_episode_steps, verbosity=0)
mean_return = returns.mean()
print(f"Episodes of training: {episode+1}, mean reward in test mode: {mean_return}")
agent.train()
agent.save(save_dir)
return agent
def parse_args():
parser = argparse.ArgumentParser(description='Train agent with DDPG')
parser.add_argument('--env', type=str, default='Pendulum-v0', help='training environment')
parser.add_argument('--num_episodes', type=int, default=500,
help='number of episodes for training')
parser.add_argument('--max_episode_steps', type=int, default=250,
help='maximum steps per episode')
parser.add_argument('--batch_size', type=int, default=128,
help='training batch size')
parser.add_argument('--tau', type=float, default=.001,
help='for model parameter % update')
parser.add_argument('--actor_lr', type=float, default=1e-4,
help='actor learning rate')
parser.add_argument('--critic_lr', type=float, default=1e-3,
help='critic learning rate')
parser.add_argument('--gamma', type=float, default=.99,
help='gamma, the discount factor')
parser.add_argument('--eps_start', type=float, default=1.)
parser.add_argument('--eps_final', type=float, default=1e-3)
parser.add_argument('--eps_anneal', type=float, default=1e6)
parser.add_argument('--theta', type=float, default=.15,
help='theta for Ornstein Uhlenbeck process computation')
parser.add_argument('--sigma', type=float, default=.2,
help='sigma for Ornstein Uhlenbeck process computation')
parser.add_argument('--buffer_size', type=int, default=100000,
help='replay buffer size')
parser.add_argument('--eval_interval', type=int, default=15,
help='how often to test the agent without exploration (in episodes)')
parser.add_argument('--eval_episodes', type=int, default=10,
help='how many episodes to run for when testing')
parser.add_argument('--warmup_steps', type=int, default=1000,
help='warmup length, in steps')
parser.add_argument('--render', type=int, default=0)
return parser.parse_args()
if __name__ == "__main__":
from agent import PendulumAgent, MountaincarAgent
args = parse_args()
if args.env == 'Pendulum-v0':
agent = PendulumAgent()
env = gym.make(args.env)
elif args.env == 'MountainCarContinuous-v0':
agent = MountaincarAgent()
env = gym.make(args.env)
else:
print(f"CL Arg --env {args.env} not recognized")
exit(1)
agent = ddpg(agent, env, args)