forked from jstarc/nli_generation
/
adverse_alg.py
127 lines (100 loc) · 5.44 KB
/
adverse_alg.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
import numpy as np
import os
import load_data
from keras.callbacks import ModelCheckpoint, EarlyStopping
import adverse_models as am
def train_adverse_model(train, dev, adverse_model, generative_model, word_index, model_dir,
nb_epochs, batch_size, hypo_len):
if not os.path.exists(model_dir):
os.makedirs(model_dir)
train_gen = adverse_generator(train, generative_model, word_index, 0.8, batch_size,
hypo_len)
dev_gen = adverse_generator(dev, generative_model, word_index, 0.0, batch_size, hypo_len)
val_data = prepare_dev_data(dev_gen, len(dev) / batch_size)
saver = ModelCheckpoint(model_dir + '/adverse.weights', monitor = 'loss')
es = EarlyStopping(patience = 5)
return adverse_model.fit_generator(train_gen, samples_per_epoch = 64000,
nb_epoch = nb_epochs, callbacks = [saver, es], validation_data = val_data)
def prepare_dev_data(dev_gen, batches):
dev_data = [next(dev_gen) for _ in range(batches)]
trains = np.vstack([data[0][0] for data in dev_data])
gens = np.vstack([data[0][1] for data in dev_data])
ys = np.concatenate([data[1] for data in dev_data])
return [trains, gens], ys
def adverse_generator(train, gen_model, word_index, cache_prob, batch_size, hypo_len):
cache = []
while True:
mb = load_data.get_minibatches_idx(len(train), batch_size, shuffle=True)
for i, train_index in mb:
if len(train_index) != batch_size:
continue
orig_batch = [train[k] for k in train_index]
if np.random.random() > cache_prob or len(cache) < 100:
gen_batch, _ = make_gen_batch(orig_batch, gen_model, word_index, hypo_len)
cache.append(gen_batch)
else:
gen_batch = cache[np.random.random_integers(0, len(cache) - 1)]
train_batch = make_train_batch(orig_batch, word_index, hypo_len)
yield [train_batch, gen_batch], np.zeros(batch_size)
def make_train_batch(orig_batch, word_index, hypo_len):
_, X_hypo, _ = load_data.prepare_split_vec_dataset(orig_batch, word_index.index)
return load_data.pad_sequences(X_hypo, maxlen = hypo_len, dim = -1, padding = 'post')
def manual_tester(dev, aug_dev, discriminator, word_index, target_size, filename):
true_ex = 0
with open(filename+'.hidden', 'w') as h, open(filename+'.revealed', 'w') as r:
indices = np.random.random_integers(0, len(dev[0]) - 1, target_size)
for i in indices:
gen_ex = np.array([aug_dev[1][i]])
org_ex = np.array([dev[1][i]])
gen_pred = discriminator.predict(gen_ex)[0][0]
train_pred = discriminator.predict(org_ex)[0][0]
gen_hypo = word_index.print_seq(aug_dev[1][i])
train_hypo = word_index.print_seq(dev[1][i])
r_data = [gen_hypo, train_hypo, str(gen_pred), str(train_pred)]
r.write("\t".join(r_data) + '\n')
h_data = [gen_hypo, train_hypo] if np.random.rand() > 0.5 else [train_hypo, gen_hypo]
h.write("\t".join(h_data) + '\n')
if train_pred > gen_pred:
true_ex += 1
print true_ex / float(target_size)
def adverse_model_train(model_dir, train, aug_train, dev, aug_dev, dim, glove):
discriminator = am.discriminator(glove, dim)
ad_model = am.adverse_model(discriminator)
dev_len = len(aug_dev[1])
res = ad_model.fit([train[1], aug_train[1]], np.zeros(len(train[1])),
validation_data=([dev[1][:dev_len], aug_dev[1]], np.zeros(dev_len)),
verbose = 1, nb_epoch = 5)
discriminator.save_weights(model_dir + '/adverse.weights')
def adverse_model_validate(model_dir, dev, aug_dev, glove, dim):
discriminator = am.discriminator(glove, dim)
discriminator.load_weights(model_dir + '/adverse.weights')
dev_len = len(aug_dev[1])
preds_orig = discriminator.predict(dev[1][:dev_len])
preds_aug = discriminator.predict(aug_dev[1])
return np.mean(preds_orig > preds_aug)
def init_adverserial_threshold_function(target_dir, ad_threshold, glove, dim, wi, hypo_len):
### too slow to use
discriminator = am.discriminator(glove, dim)
discriminator.load_weights(target_dir + '/adverse.weights')
padding_example = (['a'] * 25, ['a'] * 15, 'entailment')
def t_func(example, loss, cpred, ctrue):
if not ctrue:
return False
vec_ex = load_data.prepare_split_vec_dataset([example, padding_example], wi.index, True)
pred = discriminator.predict(vec_ex[1][:1])[0][0]
return pred > ad_threshold
return t_func
def filter_adverserial(dataset, threshold, target_size, target_dir, glove, dim):
arg = threshold[1] == 'b'
num = float(threshold[2:])
label_size = target_size / 3
discriminator = am.discriminator(glove, dim)
discriminator.load_weights(target_dir + '/adverse.weights')
ad_preds = discriminator.predict(dataset[1], verbose = 1)[:,0]
indices = (ad_preds < num) if arg else (ad_preds > num)
final_indices = []
for l in range(3):
label_args = dataset[2][:, l] == 1
final_indices += list(np.where(indices * label_args)[0][:label_size])
ind = np.sort(final_indices)
return (dataset[0][ind], dataset[1][ind], dataset[2][ind])