-
Notifications
You must be signed in to change notification settings - Fork 1
/
TD3.py
217 lines (169 loc) · 6.8 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
import copy
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import wandb
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
activation_dict = {
"relu": torch.nn.ReLU(),
"tanh": torch.nn.Tanh(),
}
class Actor(nn.Module):
def __init__(self, state_dim, action_dim, max_action):
super(Actor, self).__init__()
self.l1 = nn.Linear(state_dim, 256)
self.l2 = nn.Linear(256, 256)
self.l3 = nn.Linear(256, action_dim)
self.max_action = max_action
def forward(self, state):
a = F.relu(self.l1(state))
a = F.relu(self.l2(a))
return self.max_action * torch.tanh(self.l3(a))
class Critic(nn.Module):
def __init__(self, state_dim, action_dim, activation='relu', layer_norm=False):
super(Critic, self).__init__()
self.activation = activation_dict[activation]
self.output_activation = torch.nn.Identity()
self.layer_norm = nn.LayerNorm(256, 256)
# Q1 architecture
self.l1 = nn.Linear(state_dim + action_dim, 256)
self.l2 = nn.Linear(256, 256)
self.l3 = nn.Linear(256, 1)
# Q2 architecture
self.l4 = nn.Linear(state_dim + action_dim, 256)
self.l5 = nn.Linear(256, 256)
self.l6 = nn.Linear(256, 1)
def forward(self, state, action):
sa = torch.cat([state, action], 1)
q1 = self.activation(self.l1(sa))
q1 = self.activation(self.l2(q1))
q1 = self.output_activation(self.l3(q1))
q2 = self.activation(self.l4(sa))
q2 = self.activation(self.l5(q2))
q2 = self.output_activation(self.l6(q2))
return q1, q2
def Q1(self, state, action):
sa = torch.cat([state, action], 1)
q1 = self.activation(self.l1(sa))
q1 = self.activation(self.l2(q1))
q1 = self.output_activation(self.l3(q1))
return q1
class EBM(nn.Module):
def __init__(self, state_dim, action_dim, neg_samples=256):
super().__init__()
self.l1 = nn.Linear(state_dim + action_dim, 256)
self.l2 = nn.Linear(256, 256)
self.l3 = nn.Linear(256, 1)
self.neg_samples = neg_samples
def forward(self, state, action):
sa = torch.cat([state, action], -1)
e1 = F.tanh(self.l1(sa))
e1 = F.tanh(self.l2(e1))
e1 = self.l3(e1)
return e1
class TD3_BC(object):
def __init__(
self,
state_dim,
action_dim,
max_action,
discount=0.99,
tau=0.005,
policy_noise=0.2,
noise_clip=0.5,
policy_freq=2,
alpha=1.0,
critic_activation='relu',
layer_norm=False,
neg_samples=200,
):
self.actor = Actor(state_dim, action_dim, max_action).to(device)
self.actor_target = copy.deepcopy(self.actor)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-5)
self.critic = Critic(state_dim, action_dim, critic_activation, layer_norm).to(device)
self.critic_target = copy.deepcopy(self.critic)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4)
self.ebm = EBM(state_dim, action_dim, neg_samples).to(device)
self.ebm_optimizer = torch.optim.Adam(self.ebm.parameters(), lr=3e-4)
self.max_action = max_action
self.discount = discount
self.tau = tau
self.policy_noise = policy_noise
self.noise_clip = noise_clip
self.policy_freq = policy_freq
self.alpha = alpha
self.total_it = 0
def select_action(self, state):
state = torch.FloatTensor(state.reshape(1, -1)).to(device)
return self.actor(state).cpu().data.numpy().flatten()
def learn_ebm(self, replay_buffer, batch_size=256):
state, action, next_state, reward, not_done = replay_buffer.sample(batch_size)
negative_actions = torch.rand(size=(state.shape[0], self.ebm.neg_samples, action.shape[-1])) * 2 - 1
target_actions = torch.cat([action.unsqueeze(dim=1), negative_actions.to(device)], dim=1)
energy = self.ebm(state.unsqueeze(1).repeat(1, target_actions.shape[1], 1), target_actions).squeeze()
logits = -energy
loss = (-logits[:, 0] + torch.logsumexp(logits, dim=-1)).mean()
self.ebm_optimizer.zero_grad()
loss.backward()
self.ebm_optimizer.step()
return loss.item()
def eval_ebm(self, replay_buffer, batch_size=256):
state, action, next_state, reward, not_done = replay_buffer.sample(batch_size)
negative_actions = torch.rand(size=(state.shape[0], self.ebm.neg_samples, action.shape[-1])) * 2 - 1
target_actions = torch.cat([action.unsqueeze(dim=1), negative_actions.to(device)], dim=1)
energy = self.ebm(state.unsqueeze(1).repeat(1, target_actions.shape[1], 1), target_actions).squeeze()
energy_positive = energy[:, 0].mean()
energy_negtive = energy[:, 1:].mean()
return energy_positive, energy_negtive
def train(self, replay_buffer, batch_size=256):
self.total_it += 1
# Sample replay buffer
state, action, next_state, reward, not_done = replay_buffer.sample(batch_size)
with torch.no_grad():
# Select action according to policy and add clipped noise
noise = (
torch.randn_like(action) * self.policy_noise
).clamp(-self.noise_clip, self.noise_clip)
next_action = (
self.actor_target(next_state) + noise
).clamp(-self.max_action, self.max_action)
# Compute the target Q value
target_Q1, target_Q2 = self.critic_target(next_state, next_action)
target_Q = torch.min(target_Q1, target_Q2)
target_Q = reward + not_done * self.discount * target_Q
current_Q1, current_Q2 = self.critic(state, action)
critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
wandb.log({'Q': current_Q1.mean().item()})
# Delayed policy updates
if self.total_it % self.policy_freq == 0:
# Compute actor loss
pi = self.actor(state)
Q = self.critic.Q1(state, pi)
E = self.ebm(state, pi)
actor_loss = -self.alpha * Q.mean() + E.mean()
# Optimize the actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# Update the frozen target models
for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
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)
return {'Q': current_Q1.mean().item()}
def save(self, filename):
torch.save(self.critic.state_dict(), filename + "_critic")
torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer")
torch.save(self.actor.state_dict(), filename + "_actor")
torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer")
def load(self, filename):
self.critic.load_state_dict(torch.load(filename + "_critic"))
self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer"))
self.critic_target = copy.deepcopy(self.critic)
self.actor.load_state_dict(torch.load(filename + "_actor"))
self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
self.actor_target = copy.deepcopy(self.actor)