-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrnn_construct_new_dimension.py
132 lines (106 loc) · 4.58 KB
/
rnn_construct_new_dimension.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from __future__ import print_function, division
from itertools import chain
from time import time
import theano
import theano.tensor as T
import lasagne
from sklearn.cross_validation import KFold
from common import *
SUBMISSION_FILE = '/home/marat/allen-ai-submission.csv'
QUESTION_LEN = 50
ANSWER_MAX_LEN = 20
BATCH_SIZE = 50
OUTPUT_DIM = 50
N_HIDDEN = 50
GRAD_CLIP = 10
NUM_EPOCHS = 100
MARGIN = np.exp(-1) + 0.1
# TODO: add L2
# TODO: add dropout
# TODO: add mask
all_data = read_data(TRAIN_FILE)
train_idx, valid_idx = KFold(len(all_data)).__iter__().next()
train_i, train_q, train_c, train_w = list(zip(*list(generate_batches([all_data[i] for i in train_idx]))))
train_i = np.array(train_i, dtype='int')
train_q = np.array(train_q, dtype='float32')
train_c = np.array(train_c, dtype='float32')
train_w = np.array(train_w, dtype='float32')
valid_i, valid_q, valid_c, valid_w = list(zip(*list(generate_batches([all_data[i] for i in valid_idx]))))
valid_i = np.array(valid_i, dtype='int')
valid_q = np.array(valid_q, dtype='float32')
valid_c = np.array(valid_c, dtype='float32')
valid_w = np.array(valid_w, dtype='float32')
print('Building question network ...')
q_in = lasagne.layers.InputLayer(shape=(None, QUESTION_LEN, W2V_DIM))
q_forward = lasagne.layers.RecurrentLayer(
q_in, N_HIDDEN,
# mask_input=l_mask,
grad_clipping=GRAD_CLIP,
W_in_to_hid=lasagne.init.HeUniform(),
W_hid_to_hid=lasagne.init.HeUniform(),
nonlinearity=lasagne.nonlinearities.rectify, only_return_final=True)
q_out = lasagne.layers.DenseLayer(q_forward, num_units=OUTPUT_DIM, nonlinearity=lasagne.nonlinearities.rectify)
target_values = T.vector('target_output')
question_vec = lasagne.layers.get_output(q_out)
question_nvec = question_vec / question_vec.norm(2, axis=1).dimshuffle(0, 'x')
q_params = lasagne.layers.get_all_params(q_out)
print('Building siamese answer network ...')
sa_in = lasagne.layers.InputLayer(shape=(None, QUESTION_LEN, W2V_DIM))
sa_forward = lasagne.layers.RecurrentLayer(
sa_in, N_HIDDEN,
# mask_input=l_mask,
grad_clipping=GRAD_CLIP,
W_in_to_hid=lasagne.init.HeUniform(),
W_hid_to_hid=lasagne.init.HeUniform(),
nonlinearity=lasagne.nonlinearities.rectify, only_return_final=True)
sa_out = lasagne.layers.DenseLayer(sa_forward, num_units=OUTPUT_DIM, nonlinearity=lasagne.nonlinearities.rectify)
answer_point = lasagne.layers.get_output(sa_out)
sa_params = lasagne.layers.get_all_params(q_out)
correct_vec = answer_point[0::2]
correct_nvec = correct_vec / correct_vec.norm(2, axis=1).dimshuffle((0, 'x'))
wrong_vec = answer_point[1::2]
wrong_nvec = wrong_vec / wrong_vec.norm(2, axis=1).dimshuffle((0, 'x'))
print('Computing cost ...')
correct_dist = T.exp(T.sum(question_nvec * correct_nvec, axis=1))
wrong_dist = T.exp(T.sum(question_nvec * wrong_nvec, axis=1))
cost = T.maximum(0, correct_dist - wrong_dist + MARGIN).mean()
print('Computing updates ...')
updates = lasagne.updates.adam(cost, q_params)
updates.update(lasagne.updates.adam(cost, sa_params))
print('Compiling functions ...')
train_fn = theano.function([q_in.input_var, sa_in.input_var], cost, updates=updates)
cost_fn = theano.function([q_in.input_var, sa_in.input_var], cost)
dist_fn = theano.function([q_in.input_var, sa_in.input_var], [correct_dist, wrong_dist])
print('Model compiled!')
print('Training ...')
indexi_train = np.arange(train_q.shape[0])
indexi_valid = np.arange(valid_q.shape[0])
for e in xrange(NUM_EPOCHS):
epoch_start = time()
np.random.shuffle(indexi_train)
train_costs = []
for i in xrange(0, indexi_train.shape[0], BATCH_SIZE):
keys = indexi_train[i:i + BATCH_SIZE]
cost = train_fn(train_q[keys], zip_arrays(train_c[keys], train_w[keys]))
train_costs.append(cost)
valid_costs = []
for i in xrange(0, indexi_valid.shape[0], BATCH_SIZE):
keys = indexi_valid[i:i + BATCH_SIZE]
cost = cost_fn(valid_q[keys], zip_arrays(valid_c[keys], valid_w[keys]))
valid_costs.append(cost)
time_passed = time() - epoch_start
print(e, np.mean(train_costs), np.mean(valid_costs), '%.0fsec' % time_passed)
# Forming submission file SUBMISSION_FILE
with open(SUBMISSION_FILE, 'w') as f:
f.write('id,correctAnswer\n')
for i, q, A, B, C, D in read_submission_sample(TEST_FILE):
q = np.repeat(q[np.newaxis], 2, axis=0)
a = np.vstack([A[np.newaxis], B[np.newaxis], C[np.newaxis], D[np.newaxis]])
d = list(chain(*dist_fn(q, a)))
answer_index = np.argmin(d)
answer = 'ABCD'[answer_index]
text = '%s,%s\n' % (i, answer)
f.write(text)
print(text[:-1], d)