-
Notifications
You must be signed in to change notification settings - Fork 0
/
actor_critic.py
251 lines (185 loc) · 7.46 KB
/
actor_critic.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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
import sys
import argparse
import pickle
from collections import deque
import gym
import torch
import torch.optim as optim
from torch.autograd import Variable
import numpy as np
from models import MODELS
from utils import preprocess_state
def parse_arguments():
parser = argparse.ArgumentParser()
parser.add_argument('-g', '--game', action='store', dest='game', default='Breakout-v0')
parser.add_argument('-m', '--model', action='store', dest='model', default='a2c-lstm')
parser.add_argument('-w', '--warmstart', action='store_true', dest='warm_start', default=False)
parser.add_argument('-r', '--render', action='store_true', dest='render', default=False)
args = parser.parse_args()
game = args.game
model_name = args.model
warm_start = args.warm_start
render = args.render
return game, model_name, warm_start, render
def initialize(game, model_name, warm_start):
# Initialize environment
env = gym.make(game)
num_actions = env.action_space.n
# Initialize constants
num_frames = 4
# Cold start
if not warm_start:
# Initialize model
model = MODELS[model_name](in_channels=num_frames, num_actions=num_actions)
optimizer = optim.RMSprop(model.parameters(), lr=1.0e-4, weight_decay=0.01)
# Initialize statistics
running_reward = None
running_rewards = []
# Warm start
if warm_start:
data_file = 'results/{}_{}.p'.format(game, model_name)
try:
with open(data_file, 'rb') as f:
running_rewards = pickle.load(f)
running_reward = running_rewards[-1]
prior_eps = len(running_rewards)
model_file = 'saved_models/{}_{}_ep_{}.p'.format(game, model_name, prior_eps)
with open(model_file, 'rb') as f:
saved_model = pickle.load(f)
model, optimizer = saved_model
except OSError:
print('Saved file not found. Creating new cold start model.')
model = MODELS[model_name](in_channels=num_frames, num_actions=num_actions)
optimizer = optim.RMSprop(model.parameters(), lr=1.0e-4, weight_decay=0.01)
running_reward = None
running_rewards = []
cuda = torch.cuda.is_available()
if cuda:
model = model.cuda()
return env, model, optimizer, cuda, running_reward, running_rewards
def select_action(model, state, cuda):
num_frames, height, width = state.shape
state = torch.FloatTensor(state.reshape(-1, num_frames, height, width))
if cuda:
state = state.cuda()
probs, state_value = model(Variable(state))
m = torch.distributions.Categorical(probs)
action = m.sample()
log_prob = m.log_prob(action)
return action.data[0], log_prob, state_value
def select_action_lstm(model, state, hc, cuda):
hx, cx = hc
num_frames, height, width = state.shape
state = torch.FloatTensor(state.reshape(-1, num_frames, height, width))
if cuda:
state = state.cuda()
probs, state_value, (hx, cx) = model((Variable(state), (hx, cx)))
m = torch.distributions.Categorical(probs)
action = m.sample()
log_prob = m.log_prob(action)
return action.data[0], log_prob, state_value, (hx, cx)
def backpropagate(model, optimizer, gamma, cuda):
current_reward = 0
saved_actions = model.saved_actions
policy_losses = []
value_losses = []
rewards = deque()
for r in model.rewards[::-1]:
current_reward = r + gamma * current_reward
rewards.appendleft(current_reward)
rewards = list(rewards)
rewards = torch.Tensor(rewards)
if cuda:
rewards = rewards.cuda()
# z-score rewards
rewards = (rewards - rewards.mean()) / (rewards.std() + np.finfo(float).eps)
for (log_prob, state_value), r in zip(saved_actions, rewards):
reward = r - state_value.data[0]
policy_losses.append(-log_prob * Variable(reward))
r = torch.Tensor([r])
if cuda:
r = r.cuda()
value_losses.append(torch.nn.functional.smooth_l1_loss(state_value.view(1), Variable(r)))
optimizer.zero_grad()
loss = torch.stack(policy_losses).sum() + torch.stack(value_losses).sum()
loss.backward()
# Clip gradient at 20,000
# torch.nn.utils.clip_grad_norm(model.parameters(), 20000)
optimizer.step()
del model.rewards[:]
del model.saved_actions[:]
def main():
# Parse arguments
game, model_name, warm_start, render = parse_arguments()
# initialize enviroment/model
data = initialize(game, model_name, warm_start)
env, model, optimizer, cuda, running_reward, running_rewards = data
# Initialize constants
max_episodes = 500000
max_frames = 10000
gamma = 0.95
num_frames = 4
for ep in range(len(running_rewards), max_episodes):
# Anneal temperature from 1.8 down to 0.8 over 20,000 episodes
model.temperature = max(0.8, 1.8 - 1.0 * ((ep) / 2.0e4))
# Reset LSTM hidden units when episode begins
if model_name == 'a2c-lstm':
cx = Variable(torch.zeros(1, 100))
hx = Variable(torch.zeros(1, 100))
if cuda:
cx = cx.cuda()
hx = hx.cuda()
state = env.reset()
state = preprocess_state(state)
state = np.stack([state]*num_frames)
reward_sum = 0.0
for frame in range(max_frames):
# render frame if render argument was passed
if render:
env.render()
# Select action
if model_name == 'a2c-lstm':
result = select_action_lstm(model, state, (hx, cx), cuda)
action, log_prob, state_value, (hx, cx) = result
else:
result = select_action(model, state, cuda)
action, log_prob, state_value = result
model.saved_actions.append((log_prob, state_value))
# Perform step
next_state, reward, done, info = env.step(action)
# Add reward to reward buffer
model.rewards.append(reward)
reward_sum += reward
# Compute latest state
next_state = preprocess_state(next_state)
# Evict oldest frame add new frame to state
next_state = np.stack([next_state]*num_frames)
next_state[1:, :, :] = state[:-1, :, :]
state = next_state
if done:
break
# Compute/display episode statistics
if running_reward is None:
running_reward = reward_sum
else:
running_reward = running_reward * 0.99 + reward_sum * 0.01
running_rewards.append(running_reward)
verbose_str = 'Episode {} complete'.format(ep+1)
verbose_str += '\tReward total:{}'.format(reward_sum)
verbose_str += '\tRunning mean: {:.4}'.format(running_reward)
sys.stdout.write('\r' + verbose_str)
sys.stdout.flush()
# Update model
backpropagate(model, optimizer, gamma, cuda)
# Save model every 1000 episodes
if (ep+1) % 1000 == 0:
model_file = 'saved_models/{}_{}_ep_{}.p'.format(game, model_name, ep+1)
with open(model_file, 'wb') as f:
pickle.dump((model.cpu(), optimizer), f)
if cuda:
model = model.cuda()
data_file = 'results/{}_{}.p'.format(game, model_name)
with open(data_file, 'wb') as f:
pickle.dump(running_rewards, f)
if __name__ == '__main__':
main()