-
Notifications
You must be signed in to change notification settings - Fork 24
/
rl_utils.py
238 lines (184 loc) · 7.06 KB
/
rl_utils.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import threading
import numpy as np
import scipy.signal
from collections import namedtuple
import utils as ut
logger = ut.logging.get_logger()
Batch = namedtuple("Batch", ["si", "a", "adv", "r", "features", "c", "z"])
def discount(x, gamma):
return scipy.signal.lfilter(
[1], [1, -gamma], x[:,::-1], axis=1)[:,::-1]
def flatten_first_two(x):
return np.reshape(x, [-1] + list(x.shape)[2:])
def multiple_process_rollout(rollout, gamma, lambda_=1.0):
"""
given a rollout, compute its returns and the advantage
"""
batch_si = np.asarray(rollout['states'])
batch_a = np.asarray(rollout['actions'])
rewards = np.asarray(rollout['rewards'])
vpred_t = np.hstack(
[rollout['values'][:,:,0], np.expand_dims(rollout['r'], -1)])
rewards_plus_v = np.hstack(
[rollout['rewards'], np.expand_dims(rollout['r'], -1)])
batch_r = discount(rewards_plus_v, gamma)[:,:-1]
delta_t = rewards + gamma * vpred_t[:,1:] - vpred_t[:,:-1]
batch_adv = discount(delta_t, gamma * lambda_)
features = rollout['features'][:,0]
if 'conditions' in rollout:
batch_c = np.asarray(rollout['conditions'])
batch_z = None
else:
batch_c = None
batch_z = np.asarray(rollout['z'])
#batch_a = flatten_first_two(batch_a)
#batch_r = flatten_first_two(batch_r)
#batch_si = flatten_first_two(batch_si)
#batch_adv = flatten_first_two(batch_adv)
#features = features[:,:,0,:]
return Batch(batch_si, batch_a, batch_adv, batch_r, features, batch_c, batch_z)
class PartialRollout(object):
"""
a piece of a complete rollout. We run our agent, and process its experience
once it has processed enough steps.
"""
def __init__(self):
self.states = []
self.actions = []
self.rewards = []
self.values = []
self.r = 0.0
self.features = []
self.conditions = None
self.z = None
def add(self, state, action, reward, value, features, conditions=None, z=None):
self.states += [state]
self.actions += [action]
self.rewards += [reward]
self.values += [value]
self.features += [features]
if conditions is not None:
if self.conditions is None:
self.conditions = []
self.conditions += [conditions]
if z is not None:
if self.z is None:
self.z = []
self.z += [z]
class WorkerThread(threading.Thread):
def __init__(self, env, policy,
traj_enqueues, traj_placeholders, traj_size,
replay_enqueue, replay_placeholder, replay_size):
threading.Thread.__init__(self)
self.env = env
self.sess = None
self.daemon = True
self.policy = policy
self.last_features = None
self.summary_writer = None
self.num_local_steps = env.episode_length
self.traj_enqueues = traj_enqueues
self.traj_placeholders = traj_placeholders
self.traj_size = traj_size
self.replay_enqueue = replay_enqueue
self.replay_placeholder = replay_placeholder
self.replay_size = replay_size
def start_thread(self, sess, summary_writer):
self.sess = sess
self.summary_writer = summary_writer
self.start()
def run(self):
with self.sess.as_default():
self._run()
def _run(self):
rollout_provider = env_runner(
self.env, self.policy,
self.num_local_steps, self.summary_writer)
while True:
out = next(rollout_provider)
feed_dict = {
self.traj_placeholders['actions']: out.actions,
self.traj_placeholders['states']: out.states,
self.traj_placeholders['rewards']: out.rewards,
self.traj_placeholders['values']: out.values,
self.traj_placeholders['features']: out.features,
self.traj_placeholders['r']: out.r,
}
if self.env.conditional:
feed_dict.update({
self.traj_placeholders['conditions']: out.conditions,
})
else:
feed_dict.update({
self.traj_placeholders['z']: out.z,
})
for k, v in feed_dict.items():
if isinstance(v, list):
feed_dict[k] = np.array(v)
fetches = [
self.traj_enqueues,
]
if self.replay_enqueue is not None:
fetches.append(self.replay_enqueue)
feed_dict.update({
self.replay_placeholder: out.states[-1],
})
out = self.sess.run(fetches, feed_dict)
class ReplayThread(threading.Thread):
def __init__(self, replay, replay_dequeue):
threading.Thread.__init__(self)
self.replay = replay
self.replay_dequeue = replay_dequeue
def start_thread(self, sess):
self.sess = sess
self.start()
def run(self):
with self.sess.as_default():
self._run()
def _run(self):
while True:
generated = self.sess.run(self.replay_dequeue)
self.replay.push(generated)
def env_runner(env, policy, num_local_steps, summary_writer):
last_state, condition, z = env.reset()
last_features = policy.get_initial_features(1, flat=True)
length = 0
rewards = 0
while True:
rollout = PartialRollout()
last_action = env.initial_action
for _ in range(num_local_steps):
c, h = last_features
fetched = policy.act(
last_state, last_action, c, h, condition, z)
action, value_, features = fetched[0], fetched[1], fetched[2:4]
action = [np.argmax(action[name]) for name in env.acs]
state, reward, terminal, info = env.step(action)
# collect the experience
rollout.add(last_state, action, reward,
value_, last_features, condition, z)
length += 1
# TODO: discriminator communication to get reward
rewards += reward
last_state = state
last_action = action
last_features = features
if info:
summary = tf.Summary()
for k, v in info.items():
summary.value.add(tag=k, simple_value=float(v))
summary_writer.add_summary(summary, policy.global_step.eval())
summary_writer.flush()
last_state, condition, z = env.reset()
logger.debug(
"Episode finished. Sum of rewards: {:.5f}." \
"Length: {}.".format(rewards, length))
length = 0
rewards = 0
rollout.states += [state]
# once we have enough experience, yield it,
# and have the ThreadRunner place it on a queue
yield rollout