-
Notifications
You must be signed in to change notification settings - Fork 0
/
policy_gradient.py
291 lines (236 loc) · 10.2 KB
/
policy_gradient.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
import random
import collections
import numpy as np
from array import array
import time
from timings import Timings
import tensorflow as tf
from tf_helpers import *
class QModelBase(object):
def __init__(self):
self.state_size = None
self.n_actions = None
self.output_size = None
self.label = "none"
self.session: tf.Session = None
self.keep_prob = None
self.x = None
self.action = None
self.mask = None
self.grad_update = None
self.tensorboard_dir = 'tf-logs'
def add_summaries(self):
self.score = tf.placeholder(tf.float32, shape=[])
self.score_summary = tf.summary.scalar("score-10k", self.score)
self.merged_summaries = tf.summary.merge_all()
self.writer = tf.summary.FileWriter(self.tensorboard_dir + "/" + self.label, self.session.graph)
def add_score(self, score: float, step: int):
ss = self.session.run([self.merged_summaries], feed_dict = {self.score: score})
self.writer.add_summary(ss[0], step)
def close_session(self):
self.session.close()
def get_action(self, state: np.ndarray) -> (int, np.ndarray):
feed_dict = {self.x: state}
if self.keep_prob is not None:
feed_dict[self.keep_prob] = 1
action = self.session.run(fetches = [self.action], feed_dict = feed_dict)
return (action[0], [])
def train_one_batch(self, state: np.ndarray, mask: np.ndarray, rewards: np.ndarray):
feed_dict = { self.x : state, self.mask: mask, self.y : rewards}
if self.keep_prob is not None:
feed_dict[self.keep_prob] = .5
self.session.run(self.grad_update, feed_dict = feed_dict)
class QModel(QModelBase):
def __init__(self,
session: tf.Session,
label: str,
n_actions: int,
state_size: int,
layer_size: int,
dropout: bool,
res_blocks: int,
res_layers: int,
layers: int,
layer_norm: bool,
learn_rate: float,
tensorboard_dir):
super().__init__()
state_frames = 1
self.label = label
self.session = session
self.tensorboard_dir = tensorboard_dir
input_size = state_size * state_frames
print("input_size %d" % input_size)
x = tf.placeholder(tf.float32, [None, input_size], name = "X")
tip = x
if (dropout):
keep_prob = tf.placeholder(tf.float32, name = "keep_prob")
else:
keep_prob = None
for i in range(res_blocks):
tip = add_residual_block(tip, layer_size, res_layers, layer_norm = layer_norm, keep_prob = keep_prob, name = ("res-block-%d" % i))
if res_blocks == 0:
for i in range(0, layers):
#tip = add_dense(tip, layer_size, name = ("layer-%d" % i))
tip = add_dense_norm(tip, layer_size, layer_norm = layer_norm, keep_prob = keep_prob, name = ("layer-%d" % i))
output_size = n_actions
action_scores = tf.identity(tf.layers.dense(tip, output_size), name = "action_scores")
action = tf.argmax(action_scores, axis=1, name = "action")
print("output_layer.size: ", action_scores.shape)
y = tf.placeholder(tf.float32, [None, output_size], name = "Y")
error_mask = tf.placeholder(tf.float32, [None, output_size], name = "Y_mask")
error = tf.multiply(tf.square(action_scores - y), error_mask)
network_params = tf.trainable_variables()
optimizer = tf.train.AdamOptimizer(learning_rate = learn_rate, beta1 = 0.9, beta2 = 0.999, epsilon = 1e-8, use_locking = False)
grad_update = optimizer.minimize(error, var_list = network_params, name = "grad_update")
self.x = x
self.output_layer = action_scores
self.output_size = output_size
self.action = action
self.n_actions = n_actions
self.input_size = input_size
self.y = y
self.mask = error_mask
self.error = error
self.network_params = network_params
self.optimizer = optimizer
self.grad_update = grad_update
self.state_frames = state_frames
self.state_size = state_size
self.keep_prob = keep_prob
self.add_summaries()
self.session.run(tf.global_variables_initializer())
class QModelFromSavedFile(QModelBase):
def __init__(self, session, path):
super().__init__()
saver = tf.train.import_meta_graph(path + ".meta")
saver.restore(session, path)
g = tf.get_default_graph()
self.x = g.get_tensor_by_name("X:0")
self.action_scores = g.get_tensor_by_name("action_scores:0")
self.action = g.get_tensor_by_name("action:0")
self.mask = g.get_tensor_by_name("Y_mask:0")
self.keep_prob = g.get_tensor_by_name("keep_prob:0")
self.state_size = int(self.x.shape[1])
self.n_actions = self.action_scores.shape[1]
self.output_size = self.action_scores.shape[1]
print("state_size %d n_actions %d output_size %d" % (self.state_size, self.n_actions, self.output_size))
self.label = "none"
self.session: tf.Session = session
self.grad_update = g.get_operation_by_name("grad_update")
self.tensorboard_dir = 'tf-logs'
class StateActionReward(object) :
def __init__(self, state: np.ndarray, action: int, reward: float):
self.state = state
self.action = action
self.reward = reward
self.next_state = None
def add_reward(self, reward) :
self.reward += reward
def set_next_state(self, next_state) :
self.next_state = next_state
def training_sample(self, model: QModel) -> (np.ndarray, np.ndarray) :
state_size = model.state_size
n_actions = model.n_actions
if (model.output_size == model.n_actions):
mask = np.zeros(shape = [1, n_actions ])
output = np.zeros(shape = [1, n_actions])
mask[0, self.action] = 1
output[0, self.action] = self.reward
return self.state, mask, output
else:
mask = np.zeros(shape = [1, n_actions + state_size * n_actions])
output = np.zeros(shape = [1, n_actions + state_size * n_actions])
mask[0, self.action] = 1
output[0, self.action] = self.reward
state_start = self.action * state_size + n_actions
state_end = state_start + state_size
output[0, state_start : state_end] = self.next_state
mask[0, state_start : state_end] = 1
return self.state, mask, output
class PolicyGradientOptimizer(object) :
def __init__(self,
model: QModel,
input_size,
n_actions,
max_memory_size: int,
action_buffer_size: int,
gamma: float,
exploration: float,
timings: Timings) :
self.model = model
self.state_size = model.state_size
self.input_size = input_size
self.n_actions = n_actions
self.max_memory_size = max_memory_size
self.action_buffer_size = action_buffer_size
self.gamma = gamma
self.action_buffer = collections.deque()
self.memory = []
self.exploration = exploration
self.timings = timings
def memory_size(self) -> int :
return len(self.memory)
def ready(self) -> bool :
return (self.memory_size() > 10240)
def add_action(self, state: np.ndarray, action: int, reward: float) :
self.action_buffer.appendleft(StateActionReward(state, action, reward))
if (reward != 0.0) :
for sar in self.action_buffer :
reward *= self.gamma
sar.add_reward(reward)
# add buffer events
while len(self.action_buffer) > self.action_buffer_size :
sar = self.action_buffer.pop()
sar.set_next_state(self.action_buffer[-1].state)
self.add_to_memory(sar)
# trim memory
while len(self.memory) > self.max_memory_size :
self.memory.pop()
# if the memory is not yet full append to the existing memory. If it is full
# then replace a randomly chosen sample.
def add_to_memory(self, sar):
if len(self.memory) < self.max_memory_size:
self.memory.append(sar)
else:
idx = np.random.randint(0, len(self.memory))
self.memory[idx] = sar
# this is for when we have ended an episode and the new set of actions and rewards are independent
# of the the previous actions. For instance, a new ball in pong.
def flush_action_buffer(self):
n = len(self.action_buffer)
while len(self.action_buffer) > 0:
sar = self.action_buffer.pop()
if (len(self.action_buffer) > 0):
sar.set_next_state(self.action_buffer[-1].state)
self.add_to_memory(sar)
self.action_buffer = collections.deque()
return n
def train(self, n_epochs: int, batch_size: int) :
for i in range(n_epochs) :
self.train_one_batch(batch_size)
def get_action(self, input: np.ndarray) -> (int, np.ndarray) :
a, state = self.model.get_action(input)
return a, state
def train_one_batch(self, batch_size: int) :
if not self.ready() :
return
t = time.time()
input_rows = []
output_rows = []
mask_rows = []
for i in range(batch_size) :
idx = np.random.randint(0, len(self.memory))
sar = self.memory[idx]
irow, mask, orow = sar.training_sample(self.model)
input_rows.append(irow)
mask_rows.append(mask)
output_rows.append(orow)
self.timings.add('assemble sar', time.time() - t)
t = time.time()
input_block = np.vstack(input_rows)
output_block = np.vstack(output_rows)
mask_block = np.vstack(mask_rows)
self.timings.add('vstack', time.time() - t)
self.model.train_one_batch(input_block, mask_block, output_block)
self.timings.add('train_one_batch', time.time() - t)