-
Notifications
You must be signed in to change notification settings - Fork 157
/
train_drqn_ale.py
305 lines (282 loc) · 8.99 KB
/
train_drqn_ale.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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
"""An example of training a Deep Recurrent Q-Network (DRQN).
DRQN is a DQN with a recurrent Q-network, described in
https://arxiv.org/abs/1507.06527.
To train DRQN for 50M timesteps on Breakout, run:
python train_drqn_ale.py --recurrent
To train DQRN using a recurrent model on flickering 1-frame Breakout, run:
python train_drqn_ale.py --recurrent --flicker --no-frame-stack
"""
import argparse
import gym
import gym.wrappers
import numpy as np
import torch
from torch import nn
import pfrl
from pfrl import experiments, explorers, replay_buffers, utils
from pfrl.q_functions import DiscreteActionValueHead
from pfrl.wrappers import atari_wrappers
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--env",
type=str,
default="BreakoutNoFrameskip-v4",
help="OpenAI Atari domain to perform algorithm on.",
)
parser.add_argument(
"--outdir",
type=str,
default="results",
help=(
"Directory path to save output files."
" If it does not exist, it will be created."
),
)
parser.add_argument("--seed", type=int, default=0, help="Random seed [0, 2 ** 31)")
parser.add_argument(
"--gpu", type=int, default=0, help="GPU to use, set to -1 if no GPU."
)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load", type=str, default=None)
parser.add_argument(
"--final-exploration-frames",
type=int,
default=10**6,
help="Timesteps after which we stop " + "annealing exploration rate",
)
parser.add_argument(
"--final-epsilon",
type=float,
default=0.01,
help="Final value of epsilon during training.",
)
parser.add_argument(
"--eval-epsilon",
type=float,
default=0.001,
help="Exploration epsilon used during eval episodes.",
)
parser.add_argument(
"--steps",
type=int,
default=5 * 10**7,
help="Total number of timesteps to train the agent.",
)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument(
"--replay-start-size",
type=int,
default=5 * 10**4,
help="Minimum replay buffer size before " + "performing gradient updates.",
)
parser.add_argument(
"--target-update-interval",
type=int,
default=3 * 10**4,
help="Frequency (in timesteps) at which " + "the target network is updated.",
)
parser.add_argument("--demo-n-episodes", type=int, default=30)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument(
"--eval-interval",
type=int,
default=250000,
help="Frequency (in timesteps) of evaluation phase.",
)
parser.add_argument(
"--update-interval",
type=int,
default=4,
help="Frequency (in timesteps) of network updates.",
)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=(
"Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--lr", type=float, default=2.5e-4, help="Learning rate.")
parser.add_argument(
"--recurrent",
action="store_true",
default=False,
help="Use a recurrent model. See the code for the model definition.",
)
parser.add_argument(
"--flicker",
action="store_true",
default=False,
help=(
"Use so-called flickering Atari, where each"
" screen is blacked out with probability 0.5."
),
)
parser.add_argument(
"--no-frame-stack",
action="store_true",
default=False,
help=(
"Disable frame stacking so that the agent can only see the current screen."
),
)
parser.add_argument(
"--episodic-update-len",
type=int,
default=10,
help="Maximum length of sequences for updating recurrent models",
)
parser.add_argument(
"--batch-size",
type=int,
default=32,
help=(
"Number of transitions (in a non-recurrent case)"
" or sequences (in a recurrent case) used for an"
" update."
),
)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
flicker=args.flicker,
frame_stack=not args.no_frame_stack,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training"
)
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
print("Observation space", env.observation_space)
print("Action space", env.action_space)
n_frames = env.observation_space.shape[0]
n_actions = env.action_space.n
if args.recurrent:
# Q-network with LSTM
q_func = pfrl.nn.RecurrentSequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.ReLU(),
nn.LSTM(input_size=3136, hidden_size=512),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores whole episodes
rbuf = replay_buffers.EpisodicReplayBuffer(10**6)
else:
# Q-network without LSTM
q_func = nn.Sequential(
nn.Conv2d(n_frames, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.Flatten(),
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
DiscreteActionValueHead(),
)
# Replay buffer that stores transitions separately
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
opt = torch.optim.Adam(q_func.parameters(), lr=1e-4, eps=1e-4)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.DoubleDQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator="mean",
phi=phi,
minibatch_size=args.batch_size,
episodic_update_len=args.episodic_update_len,
recurrent=args.recurrent,
)
if args.load:
agent.load(args.load)
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.demo_n_episodes,
)
print(
"n_runs: {} mean: {} median: {} stdev {}".format(
args.demo_n_episodes,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
)
)
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
eval_env=eval_env,
)
if __name__ == "__main__":
main()