-
Notifications
You must be signed in to change notification settings - Fork 0
/
TD3.py
241 lines (194 loc) · 10.9 KB
/
TD3.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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
"""
The Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm.
* Only continuous action spaces are supported.
references:
- cleanrl: https://docs.cleanrl.dev/rl-algorithms/td3/
- cleanrl codes: https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/td3_continuous_action.py
- original papers:
* https://arxiv.org/abs/1802.09477
! Note: the code is completed with the help of copilot.
"""
import numpy as np
import torch
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from stable_baselines3.common.buffers import ReplayBuffer
import os
import random
import datetime
import time
class TD3:
"""
The Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm.
"""
def __init__(self, env, actor_class, critic_class, exp_name="td3", seed=1, cuda=0, gamma=0.99, buffer_size=1000000,
rb_optimize_memory=False, exploration_noise=0.1, policy_regular_noise=0.2, noise_clip=0.5,
actor_lr=3e-4, critic_lr=3e-4, batch_size=256, policy_frequency=2, tau=0.005, write_frequency=100,
save_folder="./td3/"):
"""
Initialize the TD3 algorithm.
:param env: the gymnasium-based environment
:param actor_class: the actor class
:param critic_class: the critic class
:param exp_name: the name of the experiment
:param seed: the random seed
:param cuda: the cuda device
:param gamma: the discount factor
:param buffer_size: the size of the replay buffer
:param rb_optimize_memory: whether to optimize the memory usage of the replay buffer
:param exploration_noise: the exploration noise
:param policy_regular_noise: the policy smoothing regularization noise scale.
:param noise_clip: the noise clip
:param actor_lr: the learning rate of the actor
:param critic_lr: the learning rate of the critic
:param batch_size: the batch size
:param policy_frequency: the policy update frequency
:param tau: the soft update coefficient
:param write_frequency: the write frequency
:param save_folder: the folder to save the model
"""
self.exp_name = exp_name
self.seed = seed
# set the random seeds
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
self.device = torch.device("cuda:{}".format(cuda) if torch.cuda.is_available() else "cpu")
self.env = env
# initialize the actor and critic networks
self.actor = actor_class(self.env).to(self.device)
self.actor_target = actor_class(self.env).to(self.device)
# twin critic
self.qf_1 = critic_class(self.env).to(self.device)
self.qf_2 = critic_class(self.env).to(self.device)
self.qf_1_target = critic_class(self.env).to(self.device)
self.qf_2_target = critic_class(self.env).to(self.device)
# copy the parameters of the policy networks to the target networks
self.actor_target.load_state_dict(self.actor.state_dict())
self.qf_1_target.load_state_dict(self.qf_1.state_dict())
self.qf_2_target.load_state_dict(self.qf_2.state_dict())
# initialize the optimizers
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=actor_lr)
self.q_optimizer = optim.Adam(list(self.qf_1.parameters()) + list(self.qf_2.parameters()), lr=critic_lr)
# + modify the observation space to be float32
self.env.observation_space.dtype = np.float32
# initialize the replay buffer
self.replay_buffer = ReplayBuffer(
buffer_size,
self.env.observation_space,
self.env.action_space,
self.device,
optimize_memory_usage=rb_optimize_memory,
handle_timeout_termination=False
)
self.gamma = gamma
self.batch_size = batch_size
self.exploration_noise = exploration_noise
self.policy_regular_noise = policy_regular_noise
self.noise_clip = noise_clip
self.policy_frequency = policy_frequency
self.tau = tau
# * for the tensorboard writer
run_name = "{}-{}-{}-{}".format(exp_name, env.unwrapped.spec.id, seed,
datetime.datetime.fromtimestamp(time.time()).strftime('%Y-%m-%d-%H-%M-%S'))
os.makedirs("./runs/", exist_ok=True)
self.writer = SummaryWriter(os.path.join("./runs/", run_name))
self.write_frequency = write_frequency
self.save_folder = save_folder
os.makedirs(self.save_folder, exist_ok=True)
def learn(self, total_timesteps=1000000, learning_starts=25000):
obs, _ = self.env.reset()
for global_step in range(total_timesteps):
if global_step < learning_starts:
action = self.env.action_space.sample()
else:
with torch.no_grad():
action = self.actor(torch.Tensor(obs).to(self.device))
action += torch.normal(0, self.actor.action_scale * self.exploration_noise)
action = action.cpu().numpy().clip(self.env.action_space.low, self.env.action_space.high)
next_obs, reward, terminated, truncated, info = self.env.step(action)
if "episode" in info:
print(f"global_step={global_step}, episodic_return={info['episode']['r']}")
self.writer.add_scalar("charts/episodic_return", info["episode"]["r"], global_step)
self.writer.add_scalar("charts/episodic_length", info["episode"]["l"], global_step)
self.replay_buffer.add(obs, next_obs, action, reward, terminated, info)
if not terminated:
obs = next_obs
else:
obs, _ = self.env.reset()
if global_step > learning_starts:
self.optimize(global_step)
self.env.close()
self.writer.close()
def optimize(self, global_step):
data = self.replay_buffer.sample(self.batch_size)
with torch.no_grad():
clipped_noise = (torch.rand_like(data.actions, device=self.device) * self.policy_regular_noise).clamp(
-self.noise_clip, self.noise_clip) * self.actor.action_scale
next_state_actions = (self.actor_target(data.next_observations) + clipped_noise).clamp(
self.env.action_space.low[0].item(), self.env.action_space.high[0].item())
qf1_next_target = self.qf_1_target(data.next_observations, next_state_actions)
qf2_next_target = self.qf_2_target(data.next_observations, next_state_actions)
min_q_next_target = torch.min(qf1_next_target, qf2_next_target)
next_q_value = data.rewards.flatten() + (1 - data.dones.flatten()) * self.gamma * min_q_next_target.view(-1)
qf1_a_values = self.qf_1(data.observations, data.actions).view(-1)
qf2_a_values = self.qf_2(data.observations, data.actions).view(-1)
qf1_loss = F.mse_loss(qf1_a_values, next_q_value)
qf2_loss = F.mse_loss(qf2_a_values, next_q_value)
qf_loss = qf1_loss + qf2_loss
# optimize the model
self.q_optimizer.zero_grad()
qf1_loss.backward()
self.q_optimizer.step()
if global_step % self.policy_frequency == 0:
actor_loss = -self.qf_1(data.observations, self.actor(data.observations)).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# update the target network
for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
for param, target_param in zip(self.qf_1.parameters(), self.qf_1_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
for param, target_param in zip(self.qf_2.parameters(), self.qf_2_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
if global_step % self.write_frequency == 0:
self.writer.add_scalar("losses/qf_1_values", qf1_a_values.mean().item(), global_step)
self.writer.add_scalar("losses/qf_2_values", qf2_a_values.mean().item(), global_step)
self.writer.add_scalar("losses/qf_1_loss", qf1_loss.item(), global_step)
self.writer.add_scalar("losses/qf_2_loss", qf2_loss.item(), global_step)
self.writer.add_scalar("losses/qf_loss", qf_loss.item() / 2.0, global_step)
self.writer.add_scalar("losses/actor_loss", actor_loss.item(), global_step)
def save(self, indicator="best"):
if indicator.startswith("best") or indicator.startswith("final"):
torch.save(self.actor.state_dict(),
os.path.join(self.save_folder,
"actor-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed)))
torch.save(self.qf_1.state_dict(),
os.path.join(self.save_folder,
"qf_1-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed)))
torch.save(self.qf_2.state_dict(),
os.path.join(self.save_folder,
"qf_2-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed)))
else:
# for normally saved models.
torch.save(self.actor.state_dict(),
os.path.join(self.save_folder,
"actor-{}-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed,
datetime.datetime.fromtimestamp(
time.time()).strftime(
'%Y-%m-%d-%H-%M-%S'))))
torch.save(self.qf_1.state_dict(),
os.path.join(self.save_folder,
"qf_1-{}-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed,
datetime.datetime.fromtimestamp(
time.time()).strftime(
'%Y-%m-%d-%H-%M-%S'))))
torch.save(self.qf_2.state_dict(),
os.path.join(self.save_folder,
"qf_2-{}-{}-{}-{}.pth".format(self.exp_name, indicator, self.seed,
datetime.datetime.fromtimestamp(
time.time()).strftime(
'%Y-%m-%d-%H-%M-%S'))))