-
Notifications
You must be signed in to change notification settings - Fork 49
/
main.py
194 lines (149 loc) · 5.52 KB
/
main.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
import copy
from collections import namedtuple
from itertools import count
import math
import random
import numpy as np
import time
import gym
from wrappers import *
from memory import ReplayMemory
from models import *
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchvision.transforms as T
Transition = namedtuple('Transion',
('state', 'action', 'next_state', 'reward'))
def select_action(state):
global steps_done
sample = random.random()
eps_threshold = EPS_END + (EPS_START - EPS_END)* \
math.exp(-1. * steps_done / EPS_DECAY)
steps_done += 1
if sample > eps_threshold:
with torch.no_grad():
return policy_net(state.to('cuda')).max(1)[1].view(1,1)
else:
return torch.tensor([[random.randrange(4)]], device=device, dtype=torch.long)
def optimize_model():
if len(memory) < BATCH_SIZE:
return
transitions = memory.sample(BATCH_SIZE)
"""
zip(*transitions) unzips the transitions into
Transition(*) creates new named tuple
batch.state - tuple of all the states (each state is a tensor)
batch.next_state - tuple of all the next states (each state is a tensor)
batch.reward - tuple of all the rewards (each reward is a float)
batch.action - tuple of all the actions (each action is an int)
"""
batch = Transition(*zip(*transitions))
actions = tuple((map(lambda a: torch.tensor([[a]], device='cuda'), batch.action)))
rewards = tuple((map(lambda r: torch.tensor([r], device='cuda'), batch.reward)))
non_final_mask = torch.tensor(
tuple(map(lambda s: s is not None, batch.next_state)),
device=device, dtype=torch.uint8)
non_final_next_states = torch.cat([s for s in batch.next_state
if s is not None]).to('cuda')
state_batch = torch.cat(batch.state).to('cuda')
action_batch = torch.cat(actions)
reward_batch = torch.cat(rewards)
state_action_values = policy_net(state_batch).gather(1, action_batch)
next_state_values = torch.zeros(BATCH_SIZE, device=device)
next_state_values[non_final_mask] = target_net(non_final_next_states).max(1)[0].detach()
expected_state_action_values = (next_state_values * GAMMA) + reward_batch
loss = F.smooth_l1_loss(state_action_values, expected_state_action_values.unsqueeze(1))
optimizer.zero_grad()
loss.backward()
for param in policy_net.parameters():
param.grad.data.clamp_(-1, 1)
optimizer.step()
def get_state(obs):
state = np.array(obs)
state = state.transpose((2, 0, 1))
state = torch.from_numpy(state)
return state.unsqueeze(0)
def train(env, n_episodes, render=False):
for episode in range(n_episodes):
obs = env.reset()
state = get_state(obs)
total_reward = 0.0
for t in count():
action = select_action(state)
if render:
env.render()
obs, reward, done, info = env.step(action)
total_reward += reward
if not done:
next_state = get_state(obs)
else:
next_state = None
reward = torch.tensor([reward], device=device)
memory.push(state, action.to('cpu'), next_state, reward.to('cpu'))
state = next_state
if steps_done > INITIAL_MEMORY:
optimize_model()
if steps_done % TARGET_UPDATE == 0:
target_net.load_state_dict(policy_net.state_dict())
if done:
break
if episode % 20 == 0:
print('Total steps: {} \t Episode: {}/{} \t Total reward: {}'.format(steps_done, episode, t, total_reward))
env.close()
return
def test(env, n_episodes, policy, render=True):
env = gym.wrappers.Monitor(env, './videos/' + 'dqn_pong_video')
for episode in range(n_episodes):
obs = env.reset()
state = get_state(obs)
total_reward = 0.0
for t in count():
action = policy(state.to('cuda')).max(1)[1].view(1,1)
if render:
env.render()
time.sleep(0.02)
obs, reward, done, info = env.step(action)
total_reward += reward
if not done:
next_state = get_state(obs)
else:
next_state = None
state = next_state
if done:
print("Finished Episode {} with reward {}".format(episode, total_reward))
break
env.close()
return
if __name__ == '__main__':
# set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# hyperparameters
BATCH_SIZE = 32
GAMMA = 0.99
EPS_START = 1
EPS_END = 0.02
EPS_DECAY = 1000000
TARGET_UPDATE = 1000
RENDER = False
lr = 1e-4
INITIAL_MEMORY = 10000
MEMORY_SIZE = 10 * INITIAL_MEMORY
# create networks
policy_net = DQN(n_actions=4).to(device)
target_net = DQN(n_actions=4).to(device)
target_net.load_state_dict(policy_net.state_dict())
# setup optimizer
optimizer = optim.Adam(policy_net.parameters(), lr=lr)
steps_done = 0
# create environment
env = gym.make("PongNoFrameskip-v4")
env = make_env(env)
# initialize replay memory
memory = ReplayMemory(MEMORY_SIZE)
# train model
train(env, 400)
torch.save(policy_net, "dqn_pong_model")
policy_net = torch.load("dqn_pong_model")
test(env, 1, policy_net, render=False)