forked from pararthshah/qa-memnn
/
memnn_theano_v3.py
484 lines (387 loc) · 18.1 KB
/
memnn_theano_v3.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
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
import numpy as np
import theano
import theano.tensor as T
import sys, random, pprint
from theano_util import *
from keras.activations import tanh, hard_sigmoid
from keras.initializations import glorot_uniform, orthogonal
from keras.utils.theano_utils import shared_zeros, alloc_zeros_matrix
def inspect_inputs(i, node, fn):
print i, node, "inputs:", [input[0] for input in fn.inputs],
def inspect_outputs(i, node, fn):
print i, node, "outputs:", [output[0] for output in fn.outputs]
class MemNN:
def __init__(self, n_words=1000, n_embedding=100, lr=0.01, margin=0.1, momentum=0.9, word_to_id=None):
self.n_embedding = n_embedding
self.n_lstm_embed = n_embedding
self.word_embed = n_embedding
self.lr = lr
self.momentum = momentum
self.margin = margin
self.n_words = n_words
self.n_D = 3 * self.n_words + 3
self.word_to_id = word_to_id
self.id_to_word = dict((v, k) for k, v in word_to_id.iteritems())
# Question
x = T.vector('x')
phi_x = T.vector('phi_x')
# True statements
phi_f1_1 = T.vector('phi_f1_1')
phi_f2_1 = T.vector('phi_f2_1')
# False statements
phi_f1_2 = T.vector('phi_f1_2')
phi_f2_2 = T.vector('phi_f2_2')
# Supporting memories
m0 = T.vector('m0')
m1 = T.vector('m1')
phi_m0 = T.vector('phi_m0')
phi_m1 = T.vector('phi_m1')
# True word
r = T.vector('r')
# Word sequence
words = T.ivector('words')
# Scoring function
self.U_O = init_shared_normal(n_embedding, self.n_D, 0.01)
# Word embedding
self.L = glorot_uniform((self.n_words, self.word_embed))
self.Lprime = glorot_uniform((self.n_words, self.n_lstm_embed))
# LSTM
self.W_i = glorot_uniform((self.word_embed, self.n_lstm_embed))
self.U_i = orthogonal((self.n_lstm_embed, self.n_lstm_embed))
self.b_i = shared_zeros((self.n_lstm_embed))
self.W_f = glorot_uniform((self.word_embed, self.n_lstm_embed))
self.U_f = orthogonal((self.n_lstm_embed, self.n_lstm_embed))
self.b_f = shared_zeros((self.n_lstm_embed))
self.W_c = glorot_uniform((self.word_embed, self.n_lstm_embed))
self.U_c = orthogonal((self.n_lstm_embed, self.n_lstm_embed))
self.b_c = shared_zeros((self.n_lstm_embed))
self.W_o = glorot_uniform((self.word_embed, self.n_lstm_embed))
self.U_o = orthogonal((self.n_lstm_embed, self.n_lstm_embed))
self.b_o = shared_zeros((self.n_lstm_embed))
mem_cost = self.calc_cost(phi_x, phi_f1_1, phi_f1_2, phi_f2_1, phi_f2_2, phi_m0)
lstm_output = self.lstm_cost(words)
self.predict_function_r = theano.function(inputs = [words], outputs = lstm_output, allow_input_downcast=True)
lstm_cost = -T.sum(T.mul(r, T.log(lstm_output)))
cost = mem_cost + lstm_cost
params = [
self.U_O,
self.W_i, self.U_i, self.b_i,
self.W_f, self.U_f, self.b_f,
self.W_c, self.U_c, self.b_c,
self.W_o, self.U_o, self.b_o,
self.L, self.Lprime
]
grads = T.grad(cost, params)
# Parameter updates
updates = self.get_updates(params, grads, method='adagrad')
l_rate = T.scalar('l_rate')
# Theano functions
self.train_function = theano.function(
inputs = [phi_x, phi_f1_1, phi_f1_2, phi_f2_1, phi_f2_2,
phi_m0, r, words,
theano.Param(l_rate, default=self.lr)],
outputs = cost,
updates = updates,
on_unused_input='warn',
allow_input_downcast=True,
)
#mode='FAST_COMPILE')
#mode='DebugMode')
#mode=theano.compile.MonitorMode(pre_func=inspect_inputs,post_func=inspect_outputs))
# Candidate statement for prediction
phi_f = T.vector('phi_f')
score_o = self.calc_score_o(phi_x, phi_f)
self.predict_function_o = theano.function(inputs = [phi_x, phi_f], outputs = score_o)
def get_updates(self, params, grads, method=None, **kwargs):
self.rho = 0.95
self.epsilon = 1e-6
accumulators = [shared_zeros(p.get_value().shape) for p in params]
updates=[]
if method == 'adadelta':
print "Using ADADELTA"
delta_accumulators = [shared_zeros(p.get_value().shape) for p in params]
for p, g, a, d_a in zip(params, grads, accumulators, delta_accumulators):
new_a = self.rho * a + (1 - self.rho) * g ** 2 # update accumulator
updates.append((a, new_a))
# use the new accumulator and the *old* delta_accumulator
update = g * T.sqrt(d_a + self.epsilon) / T.sqrt(new_a + self.epsilon)
new_p = p - self.lr * update
updates.append((p, new_p)) # apply constraints
# update delta_accumulator
new_d_a = self.rho * d_a + (1 - self.rho) * update ** 2
updates.append((d_a, new_d_a))
elif method == 'adam':
# unimplemented
print "Using ADAM"
elif method == 'adagrad':
print "Using ADAGRAD"
for p, g, a in zip(params, grads, accumulators):
new_a = a + g ** 2 # update accumulator
updates.append((a, new_a))
new_p = p - self.lr * g / T.sqrt(new_a + self.epsilon)
updates.append((p, new_p)) # apply constraints
else: # Default
print "Using MOMENTUM"
l_rate = kwargs['l_rate']
for param, gparam in zip(params, gradient):
param_update = theano.shared(param.get_value()*0., broadcastable=param.broadcastable)
updates.append((param, param - param_update * l_rate))
updates.append((param_update, self.momentum*param_update + (1. - self.momentum)*gparam))
return updates
def _step(self,
xi_t, xf_t, xc_t, xo_t,
h_tm1, c_tm1,
u_i, u_f, u_o, u_c):
i_t = hard_sigmoid(xi_t + T.dot(h_tm1, u_i))
f_t = hard_sigmoid(xf_t + T.dot(h_tm1, u_f))
c_t = f_t * c_tm1 + i_t * tanh(xc_t + T.dot(h_tm1, u_c))
o_t = hard_sigmoid(xo_t + T.dot(h_tm1, u_o))
h_t = o_t * tanh(c_t)
return h_t, c_t
# words: word index in n_words
def lstm_cost(self, words):
x = self.L[words]
# Each element of x is (word_embed,) shape
xi = T.dot(x, self.W_i) + self.b_i
xf = T.dot(x, self.W_f) + self.b_f
xc = T.dot(x, self.W_c) + self.b_c
xo = T.dot(x, self.W_o) + self.b_o
[outputs, memories], updates = theano.scan(
self._step,
sequences=[xi, xf, xc, xo],
outputs_info=[
alloc_zeros_matrix(self.n_lstm_embed),
alloc_zeros_matrix(self.n_lstm_embed),
],
non_sequences=[
self.U_i, self.U_f, self.U_o, self.U_c,
],
truncate_gradient=-1
)
r = T.dot(self.Lprime, outputs[-1])
return T.nnet.softmax(r)
def calc_score_o(self, phi_x, phi_y_yp_t):
return T.dot(self.U_O.dot(phi_x), self.U_O.dot(phi_y_yp_t))
# phi_f1_1 = phi_f1 - phi_f1bar + phi_t1_1
# phi_f1_2 = phi_f1bar - phi_f1 + phi_t1_2
def calc_cost(self, phi_x, phi_f1_1, phi_f1_2, phi_f2_1, phi_f2_2, phi_m0):
score1_1 = self.calc_score_o(phi_x, phi_f1_1)
score1_2 = self.calc_score_o(phi_x, phi_f1_2)
score2_1 = self.calc_score_o(phi_x + phi_m0, phi_f2_1)
score2_2 = self.calc_score_o(phi_x + phi_m0, phi_f2_2)
s_o_cost = (
T.maximum(0, self.margin - score1_1) + T.maximum(0, self.margin + score1_2) +
T.maximum(0, self.margin - score2_1) + T.maximum(0, self.margin + score2_2)
)
return s_o_cost
def construct_phi(self, phi_type, bow=None, word_id=None, ids=None):
# type 0: question (phi_x)
# type 1: supporting memory (phi_m*)
# type 2: candidate memory (phi_y)
# type 3: word vector
# type 4: write-time features
# type 5: 0s
assert(phi_type >= 0 and phi_type < 6)
phi = np.zeros((3*self.n_words + 3,))
if phi_type < 3:
assert(bow is not None)
phi[phi_type*self.n_words:(phi_type+1)*self.n_words] = bow
elif phi_type == 3:
assert(word_id != None and word_id < self.n_words)
phi[2*self.n_words + word_id] = 1
elif phi_type == 5:
pass
else:
assert(ids != None and len(ids) == 3)
if ids[0] > ids[1]: phi[3*self.n_words] = 1
if ids[0] > ids[2]: phi[3*self.n_words+1] = 1
if ids[1] > ids[2]: phi[3*self.n_words+2] = 1
return phi
def make_one_hot(self, index):
v = np.zeros((self.n_words))
v[index] = 1.0
return v
# returns (phi_y - phi_yp + phi_t)
def construct_wt_phi(self, index_x, index_y, index_yp, y, yp):
phi_y = self.construct_phi(2, bow=y)
phi_yp = self.construct_phi(2, bow=yp)
phi_t = self.construct_phi(4, ids=[index_x, index_y, index_yp])
return phi_y - phi_yp + phi_t
def neg_sample(self, c, num):
assert(c < num)
assert(num > 1)
f = random.randint(0, num-2)
if f == c:
f = num-1
return f
def find_m0(self, index_x, phi_x, statements, ignore=None):
max_score = float("-inf")
index_m0 = 0
m0 = statements[0]
for i in xrange(1,len(statements)):
if ignore and i == ignore:
continue
s = statements[i]
phi_s = self.construct_wt_phi(index_x, i, index_m0, s, m0)
if self.predict_function_o(phi_x, phi_s) >= 0:
index_m0 = i
m0 = s
return index_m0, m0
def train(self, dataset_seq, dataset_bow, questions, n_epochs=100, lr_schedule=None):
l_rate = self.lr
for epoch in xrange(n_epochs):
costs = []
if lr_schedule != None and epoch in lr_schedule:
l_rate = lr_schedule[epoch]
random.shuffle(questions)
for i, question in enumerate(questions):
article_no = question[0]
article = dataset_bow[article_no]
line_no = question[1]
question_phi = question[2]
correct_stmts = question[4].split(' ')
correct_stmt1 = int(correct_stmts[0])
is_single_statement = len(correct_stmts) == 1
correct_stmt2 = None
if not is_single_statement:
correct_stmt2 = int(correct_stmts[1])
question_seq = question[-1]
if line_no <= 1:
continue
# The question
phi_x = self.construct_phi(0, bow=question_phi)
# Find m0
index_m0, m0 = self.find_m0(line_no, phi_x, article[:line_no])
phi_m0 = self.construct_phi(1, bow=m0)
# Find m1
index_m1, m1 = self.find_m0(index_m0, phi_x + phi_m0, article[:line_no], ignore=index_m0)
phi_m1 = self.construct_phi(1, bow=m1)
# False statement 1
false_stmt1 = index_m0
if false_stmt1 == correct_stmt1:
false_stmt1 = self.neg_sample(correct_stmt1, line_no)
phi_f1_1 = self.construct_wt_phi(line_no, correct_stmt1, false_stmt1, article[correct_stmt1], article[false_stmt1])
phi_f1_2 = self.construct_wt_phi(line_no, false_stmt1, correct_stmt1, article[false_stmt1], article[correct_stmt1])
# False statement 2
phi_f2_1 = None
phi_f2_2 = None
if not is_single_statement:
false_stmt2 = index_m1
if false_stmt2 == correct_stmt2:
false_stmt2 = self.neg_sample(correct_stmt2, line_no)
phi_f2_1 = self.construct_wt_phi(line_no, correct_stmt2, false_stmt2, article[correct_stmt2], article[false_stmt2])
phi_f2_2 = self.construct_wt_phi(line_no, false_stmt2, correct_stmt2, article[false_stmt2], article[correct_stmt2])
else:
phi_f2_1 = self.construct_phi(5)
phi_f2_2 = self.construct_phi(5)
# Correct word
correct_word = question[3]
r = self.make_one_hot(correct_word)
words = np.asarray(dataset_seq[article_no][index_m0] + dataset_seq[article_no][index_m1] + question_seq)
cost = self.train_function(phi_x, phi_f1_1, phi_f1_2, phi_f2_1, phi_f2_2,
phi_m0, r, words)
#print "%d: %f" % (i, cost)
costs.append(cost)
print "Epoch %d: %f" % (epoch, np.mean(costs))
def find_word(self, words):
probs = self.predict_function_r(words)
return np.argmax(probs)
def predict(self, dataset_seq, dataset_bow, questions):
correct_answers = 0
wrong_answers = 0
fake_correct_answers = 0
for i, question in enumerate(questions):
article_no = question[0]
line_no = question[1]
question_phi = question[2]
correct = question[3]
question_seq = question[-1]
x = question_phi
phi_x = self.construct_phi(0, bow=question_phi)
statements = dataset_bow[article_no]
phi_m0 = None
phi_m1 = None
if len(statements) == 0:
print "Stupid question"
continue
elif len(statements) == 1:
print "Stupid question?"
phi_m0 = self.construct_phi(1, statements[0])
phi_m1 = self.construct_phi(1, statements[0])
else:
index_m0, m0 = self.find_m0(line_no, phi_x, statements[:line_no])
phi_m0 = self.construct_phi(1, m0)
index_m1, m1 = self.find_m0(index_m0, phi_x + phi_m0, statements[:line_no], ignore=index_m0)
correct_stmts = question[4].split(' ')
is_single_statement = len(correct_stmts) == 1
c1 = int(correct_stmts[0])
c2 = int(question[4].split(' ')[1]) if not is_single_statement else None
if (index_m0 == c1 or index_m0 == c2) and (index_m1 == c1 or index_m1 == c2):
fake_correct_answers += 1
predicted = self.find_word(
np.asarray(dataset_seq[article_no][index_m0] + dataset_seq[article_no][index_m1] + question_seq)
)
# print 'Correct: %s (%d), Guess: %s (%d)' % (self.id_to_word[correct], correct, self.id_to_word[predicted], predicted)
if predicted == correct:
correct_answers += 1
else:
wrong_answers += 1
print '%d correct, %d wrong, %d fake_correct' % (correct_answers, wrong_answers, fake_correct_answers)
def train_weak(self, dataset, questions, n_epochs=100, lr_schedule=None):
l_rate = self.lr
for epoch in xrange(n_epochs):
costs = []
if lr_schedule != None and epoch in lr_schedule:
l_rate = lr_schedule[epoch]
random.shuffle(questions)
for i, question in enumerate(questions):
article_no = question[0]
article = dataset[article_no]
line_no = question[1]
statements_seq = question[2][:-1]
question_seq = question[2][-1]
if line_no <= 1:
continue
# Correct word
correct_word = question[3]
cost = self.train_function(statements_seq, question_seq, correct_word)
#print "%d: %f" % (i, cost)
costs.append(cost)
print "Epoch %d: %f" % (epoch, np.mean(costs))
def predict_weak(self, dataset, questions):
correct_answers = 0
wrong_answers = 0
for i, question in enumerate(questions):
article_no = question[0]
article = dataset[article_no]
line_no = question[1]
statements_seq = question[2][:-1]
question_seq = question[2][-1]
correct = question[3]
predicted = self.predict_function(
np.asarray(statements_seq), np.asarray(question_seq)
)
# print 'Correct: %s (%d), Guess: %s (%d)' % (self.id_to_word[correct], correct, self.id_to_word[predicted], predicted)
if predicted == correct:
correct_answers += 1
else:
wrong_answers += 1
print '%d correct, %d wrong' % (correct_answers, wrong_answers)
if __name__ == "__main__":
train_file = sys.argv[1]
test_file = train_file.replace('train', 'test')
train_dataset_seq, train_dataset_bow, train_questions, word_to_id, num_words = parse_dataset(train_file)
test_dataset_seq, test_dataset_bow, test_questions, _, _ = parse_dataset(test_file, word_id=num_words, word_to_id=word_to_id, update_word_ids=False)
if len(sys.argv) > 2:
n_epochs = int(sys.argv[2])
else:
n_epochs = 10
memNN = MemNN(n_words=num_words, n_embedding=100, lr=0.01, margin=0.1, word_to_id=word_to_id)
#memNN.train(train_dataset_seq, train_dataset_bow, train_questions, n_epochs=n_epochs, lr_schedule=dict([(0, 0.02), (20, 0.01), (50, 0.005), (80, 0.002)]))
#memNN.train(train_dataset_seq, train_dataset_bow, train_questions, lr_schedule=dict([(0, 0.01), (15, 0.009), (30, 0.007), (50, 0.005), (60, 0.003), (85, 0.001)]))
#memNN.train(train_dataset_seq, train_dataset_bow, train_questions)
#memNN.predict(train_dataset, train_questions)
#memNN.predict(test_dataset_seq, test_dataset_bow, test_questions)
for i in xrange(20):
memNN.train(train_dataset_seq, train_dataset_bow, train_questions, n_epochs=5)
memNN.predict(test_dataset_seq, test_dataset_bow, test_questions)