-
Notifications
You must be signed in to change notification settings - Fork 0
/
bcq.py
244 lines (192 loc) · 8.88 KB
/
bcq.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
import logging
import pathlib
import numpy as np
import chainer
import chainer.functions as F
from chainer import optimizers
from chainer.dataset import concat_examples
class BCQ(object):
def __init__(self, critic_builder, perturbator_builder, vae_builder, state_dim, action_dim, *,
gamma=0.99, tau=0.5 * 1e-3, lmb=0.75, num_action_samples=10, num_q_ensembles=2, batch_size=100, device=-1):
self._logger = logging.getLogger(self.__class__.__name__)
self._q_ensembles = []
self._target_q_ensembles = []
self._q_optimizers = []
for _ in range(num_q_ensembles):
q_function = critic_builder(state_dim, action_dim)
target_q_function = critic_builder(state_dim, action_dim)
q_optimizer = optimizers.Adam()
q_optimizer.setup(q_function)
self._q_ensembles.append(q_function)
self._target_q_ensembles.append(target_q_function)
self._q_optimizers.append(q_optimizer)
self._perturbator = perturbator_builder(state_dim, action_dim)
self._target_perturbator = perturbator_builder(state_dim, action_dim)
self._perturbator_optimizer = optimizers.Adam()
self._perturbator_optimizer.setup(self._perturbator)
self._vae = vae_builder(state_dim, action_dim)
self._vae_optimizer = optimizers.Adam()
self._vae_optimizer.setup(self._vae)
if not device < 0:
for q_function in self._q_ensembles:
q_function.to_device(device=device)
for target_q_function in self._target_q_ensembles:
target_q_function.to_device(device=device)
self._perturbator.to_device(device=device)
self._target_perturbator.to_device(device=device)
self._vae.to_device(device=device)
self._gamma = 0.99
self._tau = tau
self._lambda = lmb
self._num_q_ensembles = num_q_ensembles
self._num_action_samples = num_action_samples
self._batch_size = batch_size
self._device = device
self._initialized = False
def train(self, iterator, **kwargs):
if not self._initialized:
self._initialize_all_networks()
self._initialized = True
batch = concat_examples(iterator.next(), device=self._device)
vae_update_status = self._train_vae(batch)
q_update_status = self._q_update(batch)
perturbator_update_status = self._perturbator_update(batch)
self._update_all_target_networks(tau=self._tau)
status = {}
status.update(vae_update_status)
status.update(q_update_status)
status.update(perturbator_update_status)
return status
def compute_action(self, s):
with chainer.using_config('enable_backprop', False), chainer.using_config('train', False):
s = np.float32(s)
if s.ndim == 1:
s = np.reshape(s, newshape=(1, ) + s.shape)
state = chainer.Variable(s)
if not self._device < 0:
state.to_gpu()
s_rep = F.repeat(x=state, repeats=100, axis=0)
a_rep = self._perturbator(s_rep, self._vae._decode(s_rep))
max_index = F.argmax(self._q_ensembles[0](s_rep, a_rep), axis=0)
a = a_rep[max_index]
if not self._device < 0:
a.to_cpu()
if a.shape[0] == 1:
return np.squeeze(a.data, axis=0)
else:
return a.data
def save_models(self, outdir, prefix):
for index, q_func in enumerate(self._q_ensembles):
q_filepath = pathlib.Path(
outdir, 'q{}_iter-{}'.format(index, prefix))
q_func.to_cpu()
q_func.save(q_filepath)
if not self._device < 0:
q_func.to_device(device=self._device)
perturbator_filepath = pathlib.Path(
outdir, 'perturbator_iter-{}'.format(prefix))
vae_filepath = pathlib.Path(outdir, 'vae_iter-{}'.format(prefix))
self._perturbator.to_cpu()
self._vae.to_cpu()
self._perturbator.save(perturbator_filepath)
self._vae.save(vae_filepath)
if not self._device < 0:
self._perturbator.to_device(device=self._device)
self._vae.to_device(device=self._device)
def load_models(self, q_param_filepaths, perturbator_filepath, vae_filepath):
for index, q_func in enumerate(self._q_ensembles):
q_func.to_cpu()
if q_param_filepaths:
q_func.load(q_param_filepaths[index])
if not self._device < 0:
q_func.to_device(device=self._device)
self._perturbator.to_cpu()
self._vae.to_cpu()
if perturbator_filepath:
self._perturbator.load(perturbator_filepath)
if vae_filepath:
self._vae.load(vae_filepath)
if not self._device < 0:
self._perturbator.to_device(device=self._device)
self._vae.to_device(device=self._device)
def _train_vae(self, batch):
status = {}
(s, a, _, _, _) = batch
reconstructed_action, mean, ln_var = self._vae((s, a))
reconstruction_loss = F.mean_squared_error(reconstructed_action, a)
latent_loss = 0.5 * \
F.gaussian_kl_divergence(mean, ln_var, reduce='mean')
vae_loss = reconstruction_loss + latent_loss
self._vae_optimizer.target.cleargrads()
vae_loss.backward()
vae_loss.unchain_backward()
self._vae_optimizer.update()
xp = chainer.backend.get_array_module(vae_loss)
status['vae_loss'] = xp.array(vae_loss.array)
return status
def _q_update(self, batch):
(s, a, _, _, _) = batch
target_q_value = self._compute_target_q_value(batch)
for optimizer in self._q_optimizers:
optimizer.target.cleargrads()
loss = 0.0
for q in self._q_ensembles:
loss += F.mean_squared_error(target_q_value, q(s, a))
loss.backward()
loss.unchain_backward()
for optimizer in self._q_optimizers:
optimizer.update()
xp = chainer.backend.get_array_module(loss)
status = {}
status['q_loss'] = xp.array(loss.array)
return status
def _perturbator_update(self, batch):
(s, _, _, _, _) = batch
sampled_actions = self._vae._decode(s)
perturbed_actions = self._perturbator(s, sampled_actions)
loss = -F.mean(self._q_ensembles[0](s, perturbed_actions))
self._perturbator_optimizer.target.cleargrads()
loss.backward()
loss.unchain_backward()
self._perturbator_optimizer.update()
status = {}
xp = chainer.backend.get_array_module(loss)
status['perturbator loss'] = xp.array(loss.array)
return status
def _compute_target_q_value(self, batch):
with chainer.using_config('train', False), \
chainer.using_config('enable_backprop', False):
(_, _, r, s_next, non_terminal) = batch
r = F.reshape(r, shape=(*r.shape, 1))
non_terminal = F.reshape(
non_terminal, shape=(*non_terminal.shape, 1))
s_next_rep = F.repeat(
x=s_next, repeats=self._num_action_samples, axis=0)
a_next_rep = self._vae._decode(s_next_rep)
perturbed_action = self._target_perturbator(s_next_rep, a_next_rep)
q_values = F.stack([q_target(s_next_rep, perturbed_action)
for q_target in self._target_q_ensembles])
assert q_values.shape == (
self._num_q_ensembles, self._batch_size * self._num_action_samples, 1)
weighted_q_minmax = self._lambda * F.min(q_values, axis=0) \
+ (1 - self._lambda) * F.max(q_values, axis=0)
assert weighted_q_minmax.shape == (
self._batch_size * self._num_action_samples, 1)
next_q_value = F.max(
F.reshape(weighted_q_minmax, shape=(self._batch_size, -1)), axis=1, keepdims=True)
assert next_q_value.shape == (self._batch_size, 1)
target_q_value = r + self._gamma * next_q_value * non_terminal
target_q_value.unchain()
assert target_q_value.shape == (self._batch_size, 1)
return target_q_value
def _initialize_all_networks(self):
self._update_all_target_networks(tau=1.0)
def _update_all_target_networks(self, tau):
for target_q, q in zip(self._target_q_ensembles, self._q_ensembles):
self._update_target_network(target_q, q, tau)
self._update_target_network(
self._target_perturbator, self._perturbator, tau)
def _update_target_network(self, target, origin, tau):
for target_param, origin_param in zip(target.params(), origin.params()):
target_param.data = tau * origin_param.data + \
(1.0 - tau) * target_param.data