/
atari_1step_qlearning.py
465 lines (389 loc) · 15.5 KB
/
atari_1step_qlearning.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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
# -*- coding: utf-8 -*-
"""
Teaching a machine to play an Atari game (Pacman by default) by implementing
a 1-step Q-learning with TFLearn, TensorFlow and OpenAI gym environment. The
algorithm is described in "Asynchronous Methods for Deep Reinforcement Learning"
paper. OpenAI's gym environment is used here for providing the Atari game
environment for handling games logic and states. This example is originally
adapted from Corey Lynch's repo (url below).
Requirements:
- gym environment (pip install gym)
- gym Atari environment (pip install gym[atari])
References:
- Asynchronous Methods for Deep Reinforcement Learning. Mnih et al, 2015.
Links:
- Paper: http://arxiv.org/pdf/1602.01783v1.pdf
- OpenAI's gym: https://gym.openai.com/
- Original Repo: https://github.com/coreylynch/async-rl
"""
from __future__ import division, print_function, absolute_import
import threading
import random
import numpy as np
import time
from skimage.transform import resize
from skimage.color import rgb2gray
from collections import deque
import gym
import tensorflow as tf
import tflearn
# Fix for TF 0.12
try:
writer_summary = tf.summary.FileWriter
merge_all_summaries = tf.summary.merge_all
histogram_summary = tf.summary.histogram
scalar_summary = tf.summary.scalar
except Exception:
writer_summary = tf.train.SummaryWriter
merge_all_summaries = tf.merge_all_summaries
histogram_summary = tf.histogram_summary
scalar_summary = tf.scalar_summary
# Change that value to test instead of train
testing = False
# Model path (to load when testing)
test_model_path = '/path/to/your/qlearning.tflearn.ckpt'
# Atari game to learn
# You can also try: 'Breakout-v0', 'Pong-v0', 'SpaceInvaders-v0', ...
game = 'MsPacman-v0'
# Learning threads
n_threads = 8
# =============================
# Training Parameters
# =============================
# Max training steps
TMAX = 80000000
# Current training step
T = 0
# Consecutive screen frames when performing training
action_repeat = 4
# Async gradient update frequency of each learning thread
I_AsyncUpdate = 5
# Timestep to reset the target network
I_target = 40000
# Learning rate
learning_rate = 0.001
# Reward discount rate
gamma = 0.99
# Number of timesteps to anneal epsilon
anneal_epsilon_timesteps = 400000
# =============================
# Utils Parameters
# =============================
# Display or not gym evironment screens
show_training = True
# Directory for storing tensorboard summaries
summary_dir = '/tmp/tflearn_logs/'
summary_interval = 100
checkpoint_path = 'qlearning.tflearn.ckpt'
checkpoint_interval = 2000
# Number of episodes to run gym evaluation
num_eval_episodes = 100
# =============================
# TFLearn Deep Q Network
# =============================
def build_dqn(num_actions, action_repeat):
"""
Building a DQN.
"""
inputs = tf.placeholder(tf.float32, [None, action_repeat, 84, 84])
# Inputs shape: [batch, channel, height, width] need to be changed into
# shape [batch, height, width, channel]
net = tf.transpose(inputs, [0, 2, 3, 1])
net = tflearn.conv_2d(net, 32, 8, strides=4, activation='relu')
net = tflearn.conv_2d(net, 64, 4, strides=2, activation='relu')
net = tflearn.fully_connected(net, 256, activation='relu')
q_values = tflearn.fully_connected(net, num_actions)
return inputs, q_values
# =============================
# ATARI Environment Wrapper
# =============================
class AtariEnvironment(object):
"""
Small wrapper for gym atari environments.
Responsible for preprocessing screens and holding on to a screen buffer
of size action_repeat from which environment state is constructed.
"""
def __init__(self, gym_env, action_repeat):
self.env = gym_env
self.action_repeat = action_repeat
# Agent available actions, such as LEFT, RIGHT, NOOP, etc...
self.gym_actions = range(gym_env.action_space.n)
# Screen buffer of size action_repeat to be able to build
# state arrays of size [1, action_repeat, 84, 84]
self.state_buffer = deque()
def get_initial_state(self):
"""
Resets the atari game, clears the state buffer.
"""
# Clear the state buffer
self.state_buffer = deque()
x_t = self.env.reset()
x_t = self.get_preprocessed_frame(x_t)
s_t = np.stack([x_t for i in range(self.action_repeat)], axis=0)
for i in range(self.action_repeat-1):
self.state_buffer.append(x_t)
return s_t
def get_preprocessed_frame(self, observation):
"""
0) Atari frames: 210 x 160
1) Get image grayscale
2) Rescale image 110 x 84
3) Crop center 84 x 84 (you can crop top/bottom according to the game)
"""
return resize(rgb2gray(observation), (110, 84))[13:110 - 13, :]
def step(self, action_index):
"""
Excecutes an action in the gym environment.
Builds current state (concatenation of action_repeat-1 previous
frames and current one). Pops oldest frame, adds current frame to
the state buffer. Returns current state.
"""
x_t1, r_t, terminal, info = self.env.step(self.gym_actions[action_index])
x_t1 = self.get_preprocessed_frame(x_t1)
previous_frames = np.array(self.state_buffer)
s_t1 = np.empty((self.action_repeat, 84, 84))
s_t1[:self.action_repeat-1, :] = previous_frames
s_t1[self.action_repeat-1] = x_t1
# Pop the oldest frame, add the current frame to the queue
self.state_buffer.popleft()
self.state_buffer.append(x_t1)
return s_t1, r_t, terminal, info
# =============================
# 1-step Q-Learning
# =============================
def sample_final_epsilon():
"""
Sample a final epsilon value to anneal towards from a distribution.
These values are specified in section 5.1 of http://arxiv.org/pdf/1602.01783v1.pdf
"""
final_epsilons = np.array([.1, .01, .5])
probabilities = np.array([0.4, 0.3, 0.3])
return np.random.choice(final_epsilons, 1, p=list(probabilities))[0]
def actor_learner_thread(thread_id, env, session, graph_ops, num_actions,
summary_ops, saver):
"""
Actor-learner thread implementing asynchronous one-step Q-learning, as specified
in algorithm 1 here: http://arxiv.org/pdf/1602.01783v1.pdf.
"""
global TMAX, T
# Unpack graph ops
s = graph_ops["s"]
q_values = graph_ops["q_values"]
st = graph_ops["st"]
target_q_values = graph_ops["target_q_values"]
reset_target_network_params = graph_ops["reset_target_network_params"]
a = graph_ops["a"]
y = graph_ops["y"]
grad_update = graph_ops["grad_update"]
summary_placeholders, assign_ops, summary_op = summary_ops
# Wrap env with AtariEnvironment helper class
env = AtariEnvironment(gym_env=env,
action_repeat=action_repeat)
# Initialize network gradients
s_batch = []
a_batch = []
y_batch = []
final_epsilon = sample_final_epsilon()
initial_epsilon = 1.0
epsilon = 1.0
print("Thread " + str(thread_id) + " - Final epsilon: " + str(final_epsilon))
time.sleep(3*thread_id)
t = 0
while T < TMAX:
# Get initial game observation
s_t = env.get_initial_state()
terminal = False
# Set up per-episode counters
ep_reward = 0
episode_ave_max_q = 0
ep_t = 0
while True:
# Forward the deep q network, get Q(s,a) values
readout_t = q_values.eval(session=session, feed_dict={s: [s_t]})
# Choose next action based on e-greedy policy
a_t = np.zeros([num_actions])
if random.random() <= epsilon:
action_index = random.randrange(num_actions)
else:
action_index = np.argmax(readout_t)
a_t[action_index] = 1
# Scale down epsilon
if epsilon > final_epsilon:
epsilon -= (initial_epsilon - final_epsilon) / anneal_epsilon_timesteps
# Gym excecutes action in game environment on behalf of actor-learner
s_t1, r_t, terminal, info = env.step(action_index)
# Accumulate gradients
readout_j1 = target_q_values.eval(session = session,
feed_dict = {st : [s_t1]})
clipped_r_t = np.clip(r_t, -1, 1)
if terminal:
y_batch.append(clipped_r_t)
else:
y_batch.append(clipped_r_t + gamma * np.max(readout_j1))
a_batch.append(a_t)
s_batch.append(s_t)
# Update the state and counters
s_t = s_t1
T += 1
t += 1
ep_t += 1
ep_reward += r_t
episode_ave_max_q += np.max(readout_t)
# Optionally update target network
if T % I_target == 0:
session.run(reset_target_network_params)
# Optionally update online network
if t % I_AsyncUpdate == 0 or terminal:
if s_batch:
session.run(grad_update, feed_dict={y: y_batch,
a: a_batch,
s: s_batch})
# Clear gradients
s_batch = []
a_batch = []
y_batch = []
# Save model progress
if t % checkpoint_interval == 0:
saver.save(session, "qlearning.ckpt", global_step=t)
# Print end of episode stats
if terminal:
stats = [ep_reward, episode_ave_max_q/float(ep_t), epsilon]
for i in range(len(stats)):
session.run(assign_ops[i],
{summary_placeholders[i]: float(stats[i])})
print("| Thread %.2i" % int(thread_id), "| Step", t,
"| Reward: %.2i" % int(ep_reward), " Qmax: %.4f" %
(episode_ave_max_q/float(ep_t)),
" Epsilon: %.5f" % epsilon, " Epsilon progress: %.6f" %
(t/float(anneal_epsilon_timesteps)))
break
def build_graph(num_actions):
# Create shared deep q network
s, q_network = build_dqn(num_actions=num_actions,
action_repeat=action_repeat)
network_params = tf.trainable_variables()
q_values = q_network
# Create shared target network
st, target_q_network = build_dqn(num_actions=num_actions,
action_repeat=action_repeat)
target_network_params = tf.trainable_variables()[len(network_params):]
target_q_values = target_q_network
# Op for periodically updating target network with online network weights
reset_target_network_params = \
[target_network_params[i].assign(network_params[i])
for i in range(len(target_network_params))]
# Define cost and gradient update op
a = tf.placeholder("float", [None, num_actions])
y = tf.placeholder("float", [None])
action_q_values = tf.reduce_sum(tf.multiply(q_values, a), reduction_indices=1)
cost = tflearn.mean_square(action_q_values, y)
optimizer = tf.train.RMSPropOptimizer(learning_rate)
grad_update = optimizer.minimize(cost, var_list=network_params)
graph_ops = {"s": s,
"q_values": q_values,
"st": st,
"target_q_values": target_q_values,
"reset_target_network_params": reset_target_network_params,
"a": a,
"y": y,
"grad_update": grad_update}
return graph_ops
# Set up some episode summary ops to visualize on tensorboard.
def build_summaries():
episode_reward = tf.Variable(0.)
scalar_summary("Reward", episode_reward)
episode_ave_max_q = tf.Variable(0.)
scalar_summary("Qmax Value", episode_ave_max_q)
logged_epsilon = tf.Variable(0.)
scalar_summary("Epsilon", logged_epsilon)
# Threads shouldn't modify the main graph, so we use placeholders
# to assign the value of every summary (instead of using assign method
# in every thread, that would keep creating new ops in the graph)
summary_vars = [episode_reward, episode_ave_max_q, logged_epsilon]
summary_placeholders = [tf.placeholder("float")
for i in range(len(summary_vars))]
assign_ops = [summary_vars[i].assign(summary_placeholders[i])
for i in range(len(summary_vars))]
summary_op = merge_all_summaries()
return summary_placeholders, assign_ops, summary_op
def get_num_actions():
"""
Returns the number of possible actions for the given atari game
"""
# Figure out number of actions from gym env
env = gym.make(game)
num_actions = env.action_space.n
return num_actions
def train(session, graph_ops, num_actions, saver):
"""
Train a model.
"""
# Set up game environments (one per thread)
envs = [gym.make(game) for i in range(n_threads)]
summary_ops = build_summaries()
summary_op = summary_ops[-1]
# Initialize variables
session.run(tf.initialize_all_variables())
writer = writer_summary(summary_dir + "/qlearning", session.graph)
# Initialize target network weights
session.run(graph_ops["reset_target_network_params"])
# Start n_threads actor-learner training threads
actor_learner_threads = \
[threading.Thread(target=actor_learner_thread,
args=(thread_id, envs[thread_id], session,
graph_ops, num_actions, summary_ops, saver))
for thread_id in range(n_threads)]
for t in actor_learner_threads:
t.start()
time.sleep(0.01)
# Show the agents training and write summary statistics
last_summary_time = 0
while True:
if show_training:
for env in envs:
env.render()
now = time.time()
if now - last_summary_time > summary_interval:
summary_str = session.run(summary_op)
writer.add_summary(summary_str, float(T))
last_summary_time = now
for t in actor_learner_threads:
t.join()
def evaluation(session, graph_ops, saver):
"""
Evaluate a model.
"""
saver.restore(session, test_model_path)
print("Restored model weights from ", test_model_path)
monitor_env = gym.make(game)
monitor_env.monitor.start("qlearning/eval")
# Unpack graph ops
s = graph_ops["s"]
q_values = graph_ops["q_values"]
# Wrap env with AtariEnvironment helper class
env = AtariEnvironment(gym_env=monitor_env,
action_repeat=action_repeat)
for i_episode in xrange(num_eval_episodes):
s_t = env.get_initial_state()
ep_reward = 0
terminal = False
while not terminal:
monitor_env.render()
readout_t = q_values.eval(session=session, feed_dict={s : [s_t]})
action_index = np.argmax(readout_t)
s_t1, r_t, terminal, info = env.step(action_index)
s_t = s_t1
ep_reward += r_t
print(ep_reward)
monitor_env.monitor.close()
def main(_):
with tf.Session() as session:
num_actions = get_num_actions()
graph_ops = build_graph(num_actions)
saver = tf.train.Saver(max_to_keep=5)
if testing:
evaluation(session, graph_ops, saver)
else:
train(session, graph_ops, num_actions, saver)
if __name__ == "__main__":
tf.app.run()