-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_dqn.py
185 lines (160 loc) · 6.02 KB
/
run_dqn.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
from algos.dqn import DQN
from algos.buffer import ReplayBuffer
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
from collections import deque
import datetime
import copy
import retro
import os
import matplotlib.pyplot as plt
import time
Tensor = torch.cuda.DoubleTensor
torch.set_default_tensor_type(Tensor)
device = torch.device("cuda:0")
# simulation setup
config = {
'env_name': 'Castlevania-aria-of-sorrow-2ndboss_2',
'image_h': 160,
'image_w': 240,
'size_action': 8, # 7 basic button + 1 combined button (up + B)
'kernel_size': 4,
'stride': 4,
'frame_skip': 2,
'double_q': True,
'lr': 0.0001,
'copy_steps': 1000,
'discount': 0.99,
'eps_max': 1.,
'eps_min': 0.02,
'exploration_steps': 10000,
'batch_size': 32,
'replay_buffer_size': 50000,
'steps_before_train': 1000,
'seed': 1,
'max_episode': 10000,
'load_from_previous': False,
'save_current': False,
}
# prepare environment
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
retro.data.Integrations.add_custom_path(os.path.join(SCRIPT_DIR, "envs"))
print(config['env_name'] in retro.data.list_games(inttype=retro.data.Integrations.ALL))
env = retro.make(config['env_name'], inttype=retro.data.Integrations.ALL)
action_map = [0, 4, 5, 6, 7, 8, 10, [4, 0]]
# action dims:
# 0 - B (attack)
# 1 - unknown
# 2 - map
# 3 - unknown
# 4 - up
# 5 - down
# 6 - left
# 7 - right
# 8 - A (jump)
# 9 - unknown
# 10 - upper left button
# 11 - unknown
# running simulation
dqn = DQN(config)
if config['load_from_previous'] is True:
Q = torch.load('./model/Q.pth.tar')
dqn.Q.load_state_dict(Q['state_dict'])
Q_tar = torch.load('./model/Q_tar.pth.tar')
dqn.Q_tar.load_state_dict(Q_tar['state_dict'])
buffer = ReplayBuffer(config)
train_writer = SummaryWriter(log_dir='tensorboard/dqn_{env:}_{date:%Y-%m-%d_%H:%M:%S}'.format(
env=config['env_name'],
date=datetime.datetime.now()))
frame_skip = config['frame_skip']
obs = env.reset()
obs_tensor = dqn.phi(obs)
obs_queue = deque([obs_tensor] * frame_skip, maxlen=frame_skip)
next_obs_queue = deque([obs_tensor] * frame_skip, maxlen=frame_skip)
steps = 0
steps_before_train = config['steps_before_train']
for i_episode in range(config['max_episode']):
obs = env.reset()
done = False
t = 0
ret = 0.
while done is False:
obs_tensor_skip = torch.cat(list(obs_queue)).to('cuda:0')[None, :].double() / 255.0
action = dqn.act_probabilistic(obs_tensor_skip)
action_onehot = np.zeros(12, dtype='int8')
action_onehot[action_map[action]] = 1
reward_skip = 0.
done_skip = False
env.render()
for ii in range(frame_skip):
obs_queue.append(dqn.phi(obs))
next_obs, reward, done, info = env.step(action_onehot)
if done is True:
obs_queue = deque([dqn.phi(obs)] * frame_skip, maxlen=frame_skip)
next_obs_queue = deque([dqn.phi(next_obs)] * frame_skip, maxlen=frame_skip)
break
next_obs_queue.append(dqn.phi(next_obs))
reward_skip += reward
done_skip = done_skip or done
obs = copy.deepcopy(next_obs)
if t > 2:
buffer.append_memory(obs=torch.cat(list(obs_queue)),
action=torch.from_numpy(np.array([action])).to(device),
reward=torch.from_numpy(np.array([reward_skip])).to(device),
next_obs=torch.cat(list(next_obs_queue)),
done=done_skip)
if steps > steps_before_train:
dqn.update(buffer)
t += 1
steps += 1
ret += reward_skip
if done:
print("Episode {} return {} (total steps: {})".format(i_episode, ret, steps))
train_writer.add_scalar('Performance/episodic_return', ret, i_episode)
if config['save_current'] is True:
torch.save({'state_dict': dqn.Q.state_dict()}, './model/Q_{}.pth.tar'.format(config['env_name']))
torch.save({'state_dict': dqn.Q_tar.state_dict()}, './model/Q_tar_{}.pth.tar'.format(config['env_name']))
env.close()
train_writer.close()
def test_model(episodes):
dqn = DQN(config)
Q = torch.load('./model/Q_{}.pth.tar'.format(config['env_name']))
dqn.Q.load_state_dict(Q['state_dict'])
Q_tar = torch.load('./model/Q_tar_{}.pth.tar'.format(config['env_name']))
dqn.Q_tar.load_state_dict(Q_tar['state_dict'])
steps = 0
for i_episode in range(episodes):
obs = env.reset()
obs_tensor = dqn.phi(obs)
obs_queue = deque([obs_tensor] * frame_skip, maxlen=frame_skip)
next_obs_queue = deque([obs_tensor] * frame_skip, maxlen=frame_skip)
done = False
t = 0
ret = 0.
while done is False:
obs_tensor_skip = torch.cat(list(obs_queue)).to('cuda:0')[None, :].double() / 255.0
action = dqn.act_deterministic(obs_tensor_skip)
action_onehot = np.zeros(12, dtype='int8')
action_onehot[action_map[action]] = 1
reward_skip = 0.
done_skip = False
env.render()
time.sleep(0.01)
for ii in range(frame_skip):
obs_queue.append(dqn.phi(obs))
next_obs, reward, done, info = env.step(action_onehot)
if done is True:
obs_queue = deque([dqn.phi(obs)] * frame_skip, maxlen=frame_skip)
next_obs_queue = deque([dqn.phi(next_obs)] * frame_skip, maxlen=frame_skip)
break
next_obs_queue.append(dqn.phi(next_obs))
reward_skip += reward
done_skip = done_skip or done
obs = copy.deepcopy(next_obs)
t += 1
steps += 1
ret += reward_skip
if done:
print("Testing episode {} return {} (total steps: {})".format(i_episode, ret, steps))
train_writer.add_scalar('TestPerformance/episodic_return', ret, i_episode)