forked from jstarc/deep_reasoning
-
Notifications
You must be signed in to change notification settings - Fork 0
/
experiment.py
204 lines (168 loc) · 6.76 KB
/
experiment.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
# -*- coding: utf-8 -*-
"""
Created on Mon Sep 28 13:43:55 2015
@author: Janez
"""
import sys
sys.path.append('../keras')
import load_data
import models
import misc
import paraphrase
import numpy as np
import itertools
import os
if __name__ == "__main__":
train, dev, test = load_data.load_all_snli_datasets('data/snli_1.0/')
glove = load_data.import_glove('data/snli_vectors.txt')
for ex in train+dev:
load_data.load_word_vecs(ex[0] + ex[1], glove)
load_data.load_word_vec('EOS', glove)
wi = load_data.WordIndex(glove)
def grid_experiments(train, dev, glove, embed_size = 300, hidden_size = 100):
lr_vec = [0.001, 0.0003, 0.0001]
dropout_vec = [0.0, 0.1, 0.2]
reg_vec = [0.0, 0.001, 0.0003, 0.0001]
for params in itertools.product(lr_vec, dropout_vec, reg_vec):
filename = 'lr' + str(params[0]).replace('.','') + '_drop' + str(params[1]).replace('.','') + '_reg' + str(params[2]).replace('.','')
print 'Model', filename
model = models.init_model(embed_size, hidden_size, params[0], params[1], params[2])
models.train_model(train, dev, glove, model, 'models/' + filename)
def test_model2(model, dev, glove):
from misc import predict_example
tp = 0
for ex in dev:
probs = predict_example(" ".join(ex[0]), " ".join(ex[1]), model, glove)
label = load_data.LABEL_LIST[np.argmax(probs)]
if label == ex[2]:
tp +=1
return tp / float(len(dev))
def test_all_models(dev, test, glove, folder = 'models/'):
files = os.listdir(folder)
extless = set([file.split('.')[0] for file in files if os.path.isfile(file)]) - set([''])
epoch_less = set([file.split('~')[0] for file in extless])
for model_short in epoch_less:
if model_short in extless:
modelname = model_short
else:
same_exper = [m for m in extless if m.startswith(model_short)]
epoch_max = max([int(file.split('~')[1]) for file in same_exper])
modelname = model_short + '~' + str(epoch_max)
print modelname
model = models.load_model(folder + modelname)
dev_acc = models.test_model(model, dev, glove)
test_acc = models.test_model(model, test, glove)
print "Dev:", '{0:.2f}'.format(dev_acc * 100), "Test_acc:", '{0:.2f}'.format(test_acc * 100)
print
def accuracy_for_subset(y_pred, y_gold, subset):
pred = y_pred[subset]
gold = y_gold[subset]
return np.sum(np.argmax(pred, axis=1) == np.argmax(gold, axis=1)) / float(len(gold))
def augmented_dataset(glove, dataset, ppdb):
new_examples = []
for ex in dataset:
new_examples += augment_example(glove, ex, ppdb)
return new_examples
def augment_example(glove, example, ppdb):
new_examples = []
for word in set(example[0] + example[1]):
if word in ppdb:
for rep in ppdb[word]:
if word in glove and rep in glove:
new_examples.append(make_new_ex(example, word, rep))
return new_examples
def make_new_ex(example, original, replacement):
premise = [replacement if word == original else word for word in example[0]]
hypo = [replacement if word == original else word for word in example[1]]
return (premise, hypo, example[2])
def test_augmentation(glove, dev, ppdb_file):
ppdb = paraphrase.load_parap(ppdb_file)
aug = augmented_dataset(glove, dev, ppdb)
return aug
def parapharse_models(glove, train, dev, ppdb_file):
ppdb = paraphrase.load_parap(ppdb_file)
aug = augmented_dataset(glove, train, ppdb)
train_aug = train + aug
models.train_model(train_aug, dev, glove, model_filename = 'models/train_aug')
models.train_model(train, dev, glove, model_filename = 'models/train_noaug')
def tune_model(observed_example, train_example, model, glove):
class_arg = load_data.LABEL_LIST.index(observed_example[2])
prem = " ".join(observed_example[0])
hypo = " ".join(observed_example[1])
print prem, hypo, observed_example[2], class_arg
for i in range(30):
probs = misc.predict_example(prem, hypo, model, glove)[0]
print i, probs
if probs.argmax() == class_arg:
break
models.update_model_once(model, glove, [train_example])
def generate_tautologies(dataset):
unique = set()
result = []
for ex in dataset:
premise = " ".join(ex[0])
if premise not in unique:
result.append((ex[0], ex[0], 'entailment'))
unique.add(premise)
return result
def generate_contradictions(dataset):
result = []
for ex in dataset:
if ex[2] == 'contradiction':
result.append((ex[1],ex[0],ex[2]))
return result
def generate_neutral(dataset):
result = []
for ex in dataset:
if ex[2] == 'entailment':
result.append((ex[1],ex[0],'neutral'))
return result
def generate_all(dataset):
return generate_tautologies(dataset) + generate_contradictions(dataset) + generate_neutral(dataset)
def unknown_words_analysis(train, dev):
train_words = set.union(*[set(ex[0]+ex[1]) for ex in train])
indices = [[],[]]
for i in range(len(dev)):
diff = len(set(dev[i][0] + dev[i][1]) - train_words)
if diff == 0:
indices[0].append(i)
else:
indices[1].append(i)
return indices
def color_analysis(dev):
COLORS = set(['black', 'blue', 'orange', 'white', 'yellow', 'green', 'pink', 'purple', 'red', 'brown', 'gray', 'grey'])
indices = [[],[]]
for i in range(len(dev)):
diff = len(set(dev[i][0] + dev[i][1]) & COLORS)
if diff == 0:
indices[0].append(i)
else:
indices[1].append(i)
return indices
def mixture_experiments(train, dev, glove, splits = 5):
for i in range(splits):
model_name = 'mixture' + str(i)
print 'Model', model_name
model = models.init_model()
div = len(train) / splits
models.train_model(train[:i*div] + train[(i+1)*div:splits*div], dev, glove, model, 'models/' + model_name)
def extended_tautologies(train, dev, glove):
augment_data = generate_all(train)
from random import shuffle
shuffle(augment_data)
augment_weight = [0, 0.05, 0.15, 0.5]
for w in augment_weight:
new_train = train + augment_data[:int(len(train)*w)]
str = str(w).replace('.','')
model = models.init_model()
models.train_model(new_train, dev, glove, model = model, model_dir = 'models/aug' + w_str)
def test_tautologies(train, dev, glove, paths = ['aug0','aug005','aug015','aug05']):
testsets = [dev, generate_tautologies(dev), generate_contradictions(dev), generate_neutral(dev)]
names = ['dev' , 'ent', 'contr' ,'neu']
for path in paths:
print path
model_path = misc.best_model_path('models/' + path)
model = models.load_model(model_path)
accs = [models.test_model(model, dataset, glove) for dataset in testsets]
for name, dataset, acc in zip (names, testsets, accs):
print name, acc, len(dataset)