forked from yufeiwang63/RLlab
-
Notifications
You must be signed in to change notification settings - Fork 0
/
DDPG_torch.py
138 lines (108 loc) · 6.87 KB
/
DDPG_torch.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
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import numpy as np
from helper_functions import SlidingMemory, PERMemory
from torch_networks import DDPG_actor_network, DDPG_critic_network, NAF_network
class DDPG():
def __init__(self, state_dim, action_dim, mem_size, train_batch_size, gamma, actor_lr, critic_lr,
action_low, action_high, tau, noise, if_PER = True):
self.mem_size, self.train_batch_size = mem_size, train_batch_size
self.gamma, self.actor_lr, self.critic_lr = gamma, actor_lr, critic_lr
self.global_step = 0
self.tau, self.explore = tau, noise
self.state_dim, self.action_dim = state_dim, action_dim
self.action_high, self.action_low = action_high, action_low
self.replay_mem = PERMemory(mem_size) if if_PER else SlidingMemory(mem_size)
#self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.device = 'cpu'
self.if_PER = if_PER
self.actor_policy_net = DDPG_actor_network(state_dim, action_dim, action_low, action_high).to(self.device)
self.actor_target_net = DDPG_actor_network(state_dim, action_dim,action_low, action_high).to(self.device)
#self.critic_policy_net = DDPG_critic_network(state_dim, action_dim).to(self.device)
#self.critic_target_net = DDPG_critic_network(state_dim, action_dim).to(self.device)
self.critic_policy_net = NAF_network(state_dim, action_dim, action_low, action_high).to(self.device)
self.critic_target_net = NAF_network(state_dim, action_dim, action_low, action_high).to(self.device)
self.actor_optimizer = optim.Adam(self.actor_policy_net.parameters(), self.actor_lr)
self.critic_optimizer = optim.Adam(self.critic_policy_net.parameters(), self.critic_lr)
self.hard_update(self.actor_target_net, self.actor_policy_net)
self.hard_update(self.critic_target_net, self.critic_policy_net)
def soft_update(self, 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(self, target, source):
for target_param, param in zip(target.parameters(), source.parameters()):
target_param.data.copy_(param.data)
# training process
def train(self, pre_state, action, reward, next_state, if_end):
self.replay_mem.add(pre_state, action, reward, next_state, if_end)
if self.replay_mem.num() < self.mem_size:
return
self.explore.decaynoise()
# sample $self.train_batch_size$ samples from the replay memory, and use them to train
if not self.if_PER:
train_batch = self.replay_mem.sample(self.train_batch_size)
else:
train_batch, idx_batch, weight_batch = self.replay_mem.sample(self.train_batch_size)
weight_batch = torch.tensor(weight_batch, dtype = torch.float).unsqueeze(1)
# adjust dtype to suit the gym default dtype
pre_state_batch = torch.tensor([x[0] for x in train_batch], dtype=torch.float, device = self.device)
action_batch = torch.tensor([x[1] for x in train_batch], dtype = torch.float, device = self.device)
# view to make later computation happy
reward_batch = torch.tensor([x[2] for x in train_batch], dtype=torch.float, device = self.device).view(self.train_batch_size,1)
next_state_batch = torch.tensor([x[3] for x in train_batch], dtype=torch.float, device = self.device)
if_end = [x[4] for x in train_batch]
if_end = torch.tensor(np.array(if_end).astype(float),device = self.device, dtype=torch.float).view(self.train_batch_size,1)
# use the target_Q_network to get the target_Q_value
with torch.no_grad():
next_action_batch = self.actor_target_net(next_state_batch)
#print(next_action_batch)
q_target_ = self.critic_target_net(next_state_batch, next_action_batch)
q_target = self.gamma * q_target_ * (1 - if_end) + reward_batch
q_pred = self.critic_policy_net(pre_state_batch, action_batch)
if self.if_PER:
TD_error_batch = np.abs(q_target.numpy() - q_pred.detach().numpy())
self.replay_mem.update(idx_batch, TD_error_batch)
self.critic_optimizer.zero_grad()
closs = (q_pred - q_target) ** 2
if self.if_PER:
closs *= weight_batch
closs = torch.mean(closs)
closs.backward()
torch.nn.utils.clip_grad_norm(self.critic_policy_net.parameters(), 1)
self.critic_optimizer.step()
self.actor_optimizer.zero_grad()
aloss = - self.critic_policy_net(pre_state_batch, self.actor_policy_net(pre_state_batch))
aloss = aloss.mean()
aloss.backward()
torch.nn.utils.clip_grad_norm(self.actor_policy_net.parameters(), 1)
self.actor_optimizer.step()
# update target network
self.soft_update(self.actor_target_net, self.actor_policy_net, self.tau)
self.soft_update(self.critic_target_net, self.critic_policy_net, self.tau)
self.global_step += 1
if self.global_step >0 and self.global_step % 10000 == 0:
torch.save(self.actor_policy_net.state_dict(), './record/ddpg_actor_param.txt')
torch.save(self.critic_policy_net.state_dict(), './record/ddpg_critic_param.txt')
# store the (pre_s, action, reward, next_state, if_end) tuples in the replay memory
def perceive(self, pre_s, action, reward, next_state, if_end):
self.replay_mem.append([pre_s, action, reward, next_state, if_end])
if len(self.replay_mem) > self.mem_size:
self.replay_mem.popleft()
# use the policy net to choose the action with the highest Q value
def action(self, s, add_noise = True):
s = torch.tensor(s, dtype=torch.float, device = self.device).unsqueeze(0)
with torch.no_grad():
action = self.actor_policy_net(s)
noise = self.explore.noise() if add_noise else 0.0
# use item() to get the vanilla number instead of a tensor
#return [np.clip(np.random.normal(action.item(), self.explore_rate), self.action_low, self.action_high)]
return np.clip(action.numpy()[0] + noise, self.action_low, self.action_high)
# choose an action according to the epsilon-greedy method
#def e_action(self, s):
# p = random.random()
# if p <= 1.0 / np.log(self.global_step + 3):
# return random.randint(0, self.action_dim - 1)
# else:
# return self.action(s)