-
Notifications
You must be signed in to change notification settings - Fork 2
/
ddpg.py
151 lines (131 loc) · 5.31 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
# -*- coding: utf-8 -*-
import gym
import roboschool
import numpy as np
from itertools import count
from OpenGL import GLU
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import visdom
from libs import replay_memory, utils
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
vis = visdom.Visdom()
def soft_update(dst, src, tau=0.001):
for dp, sp in zip(dst.parameters(), src.parameters()):
dp.data.copy_(dp.data * (1.0 - tau) + sp.data * tau)
class OUNoise:
def __init__(self, n_action, mu=0, theta=0.15, sigma=0.2):
self.n_action = n_action
self.mu = mu
self.theta = theta
self.sigma = sigma
self.reset()
def reset(self):
self.state = np.ones(self.n_action, dtype=np.float32) * self.mu
def __call__(self):
self.state += self.theta * (self.mu - self.state) + self.sigma * np.random.randn(len(self.state)).astype(np.float32)
return self.state
class Actor(nn.Module):
def __init__(self, n_action, n_state, init_w=3e-3):
super(Actor, self).__init__()
self.fc1 = nn.Linear(n_state, 128)
self.fc2 = nn.Linear(128, 128)
self.fc3 = nn.Linear(128, n_action)
self.fc3.weight.data.uniform_(-init_w, init_w)
self.fc3.bias.data.uniform_(-init_w, init_w)
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = F.tanh(self.fc3(x))
return x
class Critic(nn.Module):
def __init__(self, n_action, n_state, init_w=3e-3):
super(Critic, self).__init__()
self.fc1 = nn.Linear(n_state + n_action, 128)
self.fc2 = nn.Linear(128, 128)
self.fc3 = nn.Linear(128, 1)
self.fc3.weight.data.uniform_(-init_w, init_w)
self.fc3.bias.data.uniform_(-init_w, init_w)
def forward(self, x):
s, a = x
x = F.relu(self.fc1(torch.cat([s, a], 1)))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
BATCH_SIZE = 64
EPS_START = 1.0
EPS_END = 0.1
EPS_DECAY = 100000
env = gym.make("RoboschoolInvertedDoublePendulum-v1")
noise = OUNoise(env.action_space.shape[0])
actor = Actor(env.action_space.shape[0], env.observation_space.shape[0]).to(device)
target_actor = Actor(env.action_space.shape[0], env.observation_space.shape[0]).to(device)
target_actor.load_state_dict(actor.state_dict())
target_actor.eval()
critic = Critic(env.action_space.shape[0], env.observation_space.shape[0]).to(device)
target_critic = Critic(env.action_space.shape[0], env.observation_space.shape[0]).to(device)
target_critic.load_state_dict(critic.state_dict())
target_critic.eval()
actor_optimizer = optim.Adam(actor.parameters(), lr=0.0001)
critic_optimizer = optim.Adam(critic.parameters(), lr=0.001)
def optimize_model(memory, batch_size, criterion=nn.MSELoss(), gamma=0.999):
if len(memory) < batch_size:
return
transitions = memory.sample(batch_size)
batch = utils.Transition(*zip(*transitions))
next_state_batch = torch.stack(batch.next_state).to(device)
state_batch = torch.stack(batch.state).to(device)
action_batch = torch.stack(batch.action).to(device)
reward_batch = torch.stack(batch.reward).to(device)
done_batch = torch.stack(batch.done).to(device)
state_action_values = critic([state_batch, action_batch])
next_state_action_values = target_critic([next_state_batch, target_actor(next_state_batch)]).detach()
expected_state_action_values = (next_state_action_values * gamma * (1.0 - done_batch)) + reward_batch
critic_loss = criterion(state_action_values, expected_state_action_values)
critic_optimizer.zero_grad()
critic_loss.backward()
critic_optimizer.step()
actor_loss = -critic([state_batch, actor(state_batch)]).mean()
actor_optimizer.zero_grad()
actor_loss.backward()
actor_optimizer.step()
soft_update(target_actor, actor)
soft_update(target_critic, critic)
steps_done = 0
n_episodes = 20000
warmup = 1000
memory = replay_memory.ReplayMemory(50000)
win = vis.line(X=np.array([0]), Y=np.array([0.0]),
opts=dict(title='Score'))
for n in range(n_episodes):
# Initialize the environment and state
state = env.reset().astype(np.float32)
noise.reset()
sum_rwd = 0
for t in count():
# Perform an action
if steps_done < warmup:
action = np.random.uniform(-1.0, 1.0, env.action_space.shape[0]).astype(np.float32)
else:
eps = EPS_END + (EPS_START - EPS_END) * np.exp(-1. * steps_done / EPS_DECAY)
action = actor(torch.tensor(state).to(device)).detach().cpu().numpy() + eps * noise()
next_state, reward, done, _ = env.step(action)
next_state = next_state.astype(np.float32)
env.render()
reward = torch.tensor([reward])
done = torch.tensor([float(done)])
memory.push(torch.from_numpy(state), torch.from_numpy(action), reward,
torch.from_numpy(next_state), done)
state = next_state.copy()
# Perform one step of the optimization (on the target network)
optimize_model(memory, BATCH_SIZE)
sum_rwd += reward.numpy()
steps_done += 1
if done:
break
print("Episode: %d, Total Reward: %f" % (n, sum_rwd))
vis.line(X=np.array([n]), Y=np.array([sum_rwd]), win=win, update='append')
print('Complete')
env.close()