-
Notifications
You must be signed in to change notification settings - Fork 5
/
PolicyNetwork.py
344 lines (276 loc) · 12.1 KB
/
PolicyNetwork.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
import silence_tensorflow.auto
import numpy as np
import tensorflow as tf
from Environment import d, State, Environment
from components import BiGRU, GRU, Perceptron, SLP, Embedder, Attention
from util import train_test_split, save_checkpoint, write_model_name
from tensorflow import keras
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras import utils as np_utils
from tqdm import tqdm
class PolicyNetwork(tf.keras.Model):
def __init__(self, T, saved_model_name: str = '', env: Environment = None):
super(PolicyNetwork, self).__init__()
self.T = T
self.env = env
self.beam_size = 1
self.lr = 1e-3
self.ita_discount = 0.9
self.opt = tf.keras.optimizers.Adam(learning_rate=self.lr)
self.save_model_dir = './saved_models/'
if saved_model_name:
self.load_saved_model(saved_model_name)
def call(self, x):
q_vector, H_t = x
return self.sub_forward(q_vector, H_t)
def load_saved_model(self, saved_model_name):
try:
self.model = keras.models.load_model(
self.save_model_dir + saved_model_name)
except:
print('Load failed. Initialise new network.')
def initialise_models(self):
self.GRU = GRU()
self.Perceptron = Perceptron()
self.Attention = Attention()
self.BiGRU = BiGRU()
self.SLP = SLP(self.T)
self.Embedder = Embedder()
def initialise(self):
if not self.env:
self.env = Environment(self.KG)
if not hasattr(self, 'model'):
self.model = PolicyNetwork(self.T, env=self.env)
self.model.initialise_models()
def train(self, inputs, epochs=10, attention=True, perceptron=True):
KG, dataset, T = inputs
train_set, test_set = train_test_split(dataset)
# Hyperparameters configuration
self.T = T
self.KG = KG
self.initialise()
self.model.use_attention = attention
self.model.use_perceptron = perceptron
train_acc = []
train_losses = []
val_acc = []
val_losses = []
for i in range(epochs):
print("\n\n>>>>>>>>>>>> EPOCH: ", i + 1, " / ", epochs)
train_accuracy, train_loss = self.run_train_op(train_set)
val_accuracy, val_loss = self.run_val_op(test_set)
train_acc.append(train_accuracy)
train_losses.append(train_loss)
val_acc.append(val_accuracy)
val_losses.append(val_loss)
# Save Model
model_name = 'model'
if attention and perceptron:
model_name += '_combined'
model_type = 'combined'
elif attention and not perceptron:
model_name += '_att'
model_type = 'attention'
elif perceptron and not attention:
model_name += '_per'
model_type = 'perceptron'
model_name += str(i + 1)
write_model_name(model_name, model_type)
# Save Results
results_file = self.save_model_dir + \
"{}_results.csv".format(model_type)
with open(results_file, "a+") as f:
f.write("epoch {}, {}, {}, {}, {}\n".format(
i, train_acc, train_losses, val_acc, val_losses))
return (train_acc, train_losses), (val_acc, val_losses)
def predict(self, inputs, attention=True, perceptron=True):
KG, dataset, T = inputs
# Hyperparameters configuration
self.T = T
self.KG = KG
self.initialise()
self.model.use_attention = attention
self.model.use_perceptron = perceptron
val_acc, predictions = self.run_val_op(dataset, predictions=True)
return val_acc, predictions
def run_train_op(self, train_set, predictions=False):
# Hyperparameters configuration
self.model.beam_size = 1
y_hat = []
losses = []
print('============ TRAINING ============')
for inputs in tqdm(train_set):
with tf.GradientTape(persistent=True) as tape:
prediction, outputs = self.forward(inputs)
y_hat.append(prediction)
if all(x is None for x in outputs):
continue
loss = -outputs[-1]
gradients = tape.gradient(loss, self.model.trainable_variables)
self.opt.apply_gradients(
zip(gradients, self.model.trainable_variables))
losses.append(loss)
acc = np.mean([y_hat[i] == train_set[i][-1]
for i in range(len(y_hat))])
loss = np.mean(losses)
return acc, loss
def run_val_op(self, val_set, predictions=False):
# Hyperparameters configuration
self.model.beam_size = 32
y_hat = []
losses = []
print('\n============ VALIDATING ============')
for inputs in tqdm(val_set):
prediction, outputs = self.forward(inputs)
y_hat.append(prediction)
if all(x is None for x in outputs):
continue
loss = -outputs[-1]
losses.append(loss)
acc = np.mean([y_hat[i] == val_set[i][-1] for i in range(len(y_hat))])
loss = np.mean(losses)
return acc, loss
def forward(self, inputs):
q, e_s, ans = inputs
T = self.T
temp_q = np.empty((0, 50)).astype(np.float32)
for w in q:
embeded_word = self.model.Embedder.embed_word(w)
if embeded_word is not None and embeded_word.all():
temp_q = np.append(
temp_q, embeded_word.reshape((1, 50)), axis=0)
q = temp_q
q = tf.convert_to_tensor(
value=q, dtype=tf.float32) # Embedding Module
q = tf.reshape(q, [1, *q.shape])
r_0 = np.zeros(d).astype(np.float32)
q_vector = self.model.bigru(q) # BiGRU Module
self.model.env.start_new_query(State(q, e_s, e_s, set()), ans)
prediction, actions_onehot, action_probs, discount_r = self.model(
[q_vector, self.model.gru(r_0)])
outputs = [actions_onehot, action_probs, discount_r]
return prediction, outputs
def sub_forward(self, q_vector, H_t_t):
# OUTPUTS
rewards = []
action_probs = []
actions_onehot = []
# Trajectories
S_t = {} # T x States; state
q_t = {} # T x d x n; question
H_t = {} # T x d; encoded history
r_t = {} # T x d; relation
a_t = {} # T x d x 2(relation, next_node)
q_t_star = {} # T x d; attention weighted question
H_t[0] = np.zeros(d).astype(np.float32)
r_t[0] = np.zeros(d).astype(np.float32)
H_t[1] = H_t_t
S_t[1] = self.env.current_state
for t in range(1, self.T+1):
q_t[t] = self.slp(q_vector, t) # Single-Layer Perceptron Module
possible_actions = self.env.get_possible_actions()
# Reached terminal node
if not possible_actions:
break
action_space = self.beam_search(possible_actions)
temp_action_space = action_space.copy()
semantic_scores = []
for action in action_space:
# Attention Layer: Generate Similarity Scores between q and r and current point of attention
r_star = self.Embedder.embed_relation(action[0])
if r_star is not None and r_star.all():
r_star = tf.Variable(r_star)
if self.use_attention:
q_t_star[t] = self.attention(r_star, q_t[t])
else:
q_t_star[t] = tf.reduce_sum(q_t[t], 0)
# Perceptron Module: Generate Semantic Score for action given q
if self.use_perceptron:
score = self.perceptron(r_star, H_t[t], q_t_star[t])
else:
r_star = tf.nn.l2_normalize(r_star, 0)
temp_q_t_star = tf.nn.l2_normalize(q_t_star[t], 0)
score = tf.reduce_sum(
tf.math.multiply(r_star, temp_q_t_star))
semantic_scores.append(score)
else:
temp_action_space.remove(action)
continue
if not semantic_scores:
break
# Softmax Module: Leading to selection of action according to policy
action_distribution = tf.nn.softmax(semantic_scores)
index, action = self.sample_action(
temp_action_space, action_distribution)
a_t[t] = action
r_t[t] = self.Embedder.embed_relation(action[0])
H_t[t+1] = self.gru(r_t[t])
# Take action, advance state, and get reward
# q_t & H_t passed in order to generate the new State object within Environment
new_state, new_reward = self.env.transit(action, t, q_t, H_t)
S_t[t+1] = new_state
# Record action, state and reward
rewards.append(new_reward)
action_probs.append(action_distribution)
actions_onehot.append(np_utils.to_categorical(
index, num_classes=len(temp_action_space)))
prediction = S_t[len(S_t)].e_t
if not rewards:
return [prediction, None, None, None]
discount_r = self.discount_rewards(rewards)
output = []
action_probs = pad_sequences(
action_probs, padding='post', dtype='float32')
actions_onehot = pad_sequences(
actions_onehot, padding='post', dtype='float32')
return [prediction, actions_onehot, action_probs, discount_r]
def discount_rewards(self, rewards, normalize=False):
discounted_r = tf.Variable(0, dtype=tf.float32)
for t in reversed(range(0, len(rewards))):
discounted_r = self.ita_discount * discounted_r + rewards[t]
return discounted_r
def REINFORCE_loss_function(self, outputs):
actions_onehot, action_probs, rewards = outputs
action_prob = tf.reduce_sum(action_probs * actions_onehot, axis=1)
# Log likelihood of probabilities
log_action_prob = tf.math.log(tf.cast(action_prob, dtype=tf.float32))
loss = - log_action_prob * rewards
return tf.reduce_mean(loss)
# TRAINABLE
def bigru(self, q):
# Returns: q_vector
return self.BiGRU(q)
# TRAINABLE
def slp(self, q_vector, t):
# Returns: q_t = Tanh(Wt * q_vector + b_t)
return self.SLP(q_vector, t)
# TRAINABLE
def gru(self, r_t):
# Returns: H_t_plus_1 = GRU(H_t, r_t)
return self.GRU(r_t)
# TRAINABLE
def attention(self, r_star, q_t):
# Returns: q_t_star[t]
return self.Attention(r_star, q_t)
# TRAINABLE
def perceptron(self, r_star, H_t, q_t_star):
# Returns: S(a_t, q) = r_star * W_L2 * ReLU(W_L1 * [H_t; q_t_star])
return self.Perceptron(r_star, H_t, q_t_star)
def beam_search(self, possible_actions, beam_size=None):
if not beam_size:
beam_size = self.beam_size
actions_scores = []
for action in possible_actions:
expected_reward = self.env.get_action_reward(action)
actions_scores.append((action, expected_reward))
sorted_actions = sorted(actions_scores, key=lambda x: x[1])[:beam_size]
beamed_actions = [action_score[0] for action_score in sorted_actions]
return beamed_actions
def sample_action(self, actions, probabilities):
# Convert probabilities to log_probabilities and reshape it to [1, action_space]
rescaled_probas = tf.expand_dims(tf.math.log(
probabilities), 0) # shape [1, action_space]
# Draw one example from the distribution (we could draw more)
index = tf.compat.v1.multinomial(rescaled_probas, num_samples=1)
index = tf.squeeze(index, [0]).numpy()[0]
return index, actions[index]