-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathablation.py
379 lines (288 loc) · 12.8 KB
/
ablation.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
import numpy as np
import pandas as pd
import os
import sys
import pickle
import re
import json
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.python.client import device_lib
from keras import regularizers
from keras.backend import mean
from keras.models import Sequential, Model, load_model
from keras.layers import Dense, Conv2D, Flatten, MaxPooling2D, Activation, Input, Dropout, \
Add, add, LSTM, Bidirectional, Multiply, concatenate, Lambda
from keras.layers.normalization import BatchNormalization
from keras.layers.embeddings import Embedding
from keras.preprocessing import sequence
from keras.callbacks import ModelCheckpoint
from keras.optimizers import Adam
def load_json(json_file):
json_data = open(json_file)
data = json.load(json_data)
json_data.close()
return data
def load_pickle(pickle_file):
pickle_data = open(pickle_file, "rb")
data = pickle.load(pickle_data)
pickle_data.close()
return data
def load_glove(directory):
glove = {}
with open(directory) as file:
for line in file:
words = line.split(' ')
word, vector = words[0], np.asarray(words[1:], dtype='float32')
glove[word] = vector
return glove
def extract_questions(questions, answers):
data = {}
for q in questions:
data[q['question_id']] = {'q': q['question'], 'image_id': q['image_id']}
for a in answers:
data[a['question_id']]['a'] = a['multiple_choice_answer']
return data
def top_answers(data, k=1000):
answer_list = [v['a'] for k, v in data.items()]
a_set = set(answer_list)
w_to_i, i_to_w = {}, {}
for i, w in enumerate(a_set):
w_to_i[w] = i
i_to_w[i] = w
score = len(a_set) * [0]
for w in answer_list: score[w_to_i[w]] += 1
w_sorted = np.flip(np.argsort(score))
top_a = [i_to_w[i] for i in w_sorted[:k]]
proportion = np.sum([score[i] for i in w_sorted[:k]]) / len(answer_list)
return top_a, proportion
def clean_sentence(s, glove):
s_pre = s.lower().replace("\'s", " ").replace("\'", " ").replace("/", " ").replace("?", " ")
s_clean = s_pre.replace("-", " ").replace(".", " ").replace(",", " ").replace("\"", " ")
words = []
for w in s_clean.split(" "):
if w in glove: words.append(w)
if len(words) == 0: return None
return words
def minimize_vocab(train, test, glove):
vocabulary = set()
new_train, new_test = {}, {}
for k, v in train.items():
clean = clean_sentence(v['q'], glove)
if clean == None: continue
v['q'] = clean
new_train[k] = v
vocabulary.update(clean)
for k, v in test.items():
clean = clean_sentence(v['q'], glove)
if clean == None: continue
v['q'] = clean
new_test[k] = v
vocabulary.update(clean)
return new_train, new_test, vocabulary
def a_to_one_hot(a, a_to_i, answer_length):
b = np.zeros(answer_length)
b[a_to_i[a]] = 1
return b
def preprocessing(data, features, glove, answers, vocabulary):
answer_length = len(answers)
a_to_i = {w:i for i, w in enumerate(answers)}
w_to_i = {w:(i+1) for i, w in enumerate(vocabulary)}
embedding = np.zeros(shape=(len(vocabulary) + 1, glove_dimension))
for w, i in w_to_i.items():
embedding[i,:] = glove[w]
x_img, x_word, y = [], [], []
for k, v in data.items():
if v['a'] not in a_to_i: continue
x_img.append(features[v['image_id']])
x_word.append(np.array([w_to_i[w] for w in v['q']]))
y.append(a_to_one_hot(v['a'], a_to_i, answer_length))
x_word = sequence.pad_sequences(x_word, maxlen=25)
x = [np.array(x_img), np.array(x_word)] # vstack(np.expand_dims(x_word, axis=0))
return x, np.array(y), embedding, a_to_i
def model_one(x, y, embedding, filepath, epochs=100, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(2 * hidden_dim)(a1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
#b2 = Bidirectional(LSTM(hidden_dim, return_sequences=True, activation='relu'))(b2)
b2 = Bidirectional(LSTM(hidden_dim, activation='relu'))(b2)
b = concatenate([b1, b2])
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=[c])
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
checkpoint = ModelCheckpoint(filepath=filepath+"-{epoch:02d}.hdf5", period=5)
model.fit(x, y, callbacks=[checkpoint], epochs=epochs, batch_size=64)
return model
def model_add(x, y, embedding, filepath, epochs=100, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(2 * hidden_dim)(a1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
#b2 = Bidirectional(LSTM(hidden_dim, return_sequences=True, activation='relu'))(b2)
b2 = Bidirectional(LSTM(hidden_dim, activation='relu'))(b2)
b = Add()([b1, b2])
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=[c])
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
checkpoint = ModelCheckpoint(filepath=filepath+"-{epoch:02d}.hdf5", period=5)
model.fit(x, y, callbacks=[checkpoint], epochs=epochs, batch_size=64)
return model
def model_no_embedding(x, y, embedding, filepath, epochs=100, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
vec_dim = 50 # train vectors
a1 = Input(shape=(feat_dim,))
b1 = Dense(2 * hidden_dim)(a1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
#b2 = Bidirectional(LSTM(hidden_dim, return_sequences=True, activation='relu'))(b2)
b2 = LSTM(hidden_dim, activation='relu')(b2)
b = concatenate([b1, b2])
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=[c])
opt = Adam(lr=0.0005)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
checkpoint = ModelCheckpoint(filepath=filepath+"-{epoch:02d}.hdf5", period=1)
model.fit(x, y, callbacks=[checkpoint], epochs=epochs, batch_size=64)
return model
def model_final(x, y, embedding, filepath, epochs=100, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(2 * hidden_dim, activation='relu')(a1)
b1 = Dropout(0.2)(b1)
b1 = Dense(2 * hidden_dim, activation='tanh', activity_regularizer=regularizers.l2(0.01))(b1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
#b2 = LSTM(hidden_dim, activation='tanh', return_sequences=True)(b2)
b2 = LSTM(2 * hidden_dim, activation='tanh')(b2)
b = Multiply()([b1, b2])
c = Dense(2 * hidden_dim, activation='relu')(b)
c = Dropout(0.2)(c)
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=[c])
opt = Adam(lr=0.0005)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
checkpoint = ModelCheckpoint(filepath=filepath+"-{epoch:02d}.hdf5", period=1)
model.fit(x, y, callbacks=[checkpoint], epochs=epochs, batch_size=64)
return model
def model_new_final(x, y, embedding, filepath, epochs=100, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(hidden_dim, activation='tanh')(a1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
b2 = LSTM(hidden_dim, activation='tanh', return_sequences=True)(b2)
b2 = LSTM(hidden_dim, activation='tanh')(b2)
b = Multiply()([b1, b2])
c = Dense(2 * hidden_dim, activation='relu')(b)
c = Dropout(0.2)(c)
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=c)
opt = Adam()
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
checkpoint = ModelCheckpoint(filepath=filepath+"-{epoch:02d}.hdf5", period=1)
model.fit(x, y, callbacks=[checkpoint], epochs=epochs, batch_size=64, verbose=2)
return model
def test_preprocess(data, features, glove, vocabulary):
w_to_i = {w:i for i, w in enumerate(vocabulary)}
embedding = np.zeros(shape=(len(vocabulary), glove_dimension))
for w, i in w_to_i.items(): embedding[i,:] = glove[w]
x_img, x_word, y, q_id = [], [], [], []
for k, v in data.items():
x_img.append(features[v['image_id']])
x_word.append(np.array([w_to_i[w] for w in v['q']]))
y.append(v['a'])
q_id.append(k) # k is question id!
x_word = sequence.pad_sequences(x_word, maxlen=25)
x = [np.array(x_img), np.array(x_word)]
return x, y, q_id
def model_output(model, x_input, answer_key):
y_index = model.predict(x_input)
return [answer_key[np.argmax(pred)] for pred in y_index]
def evaluate(y_pred, y_test):
return np.mean([pred == truth for pred, truth in zip(y_pred, y_test)])
def write_results(results, file_name):
with open(file_name, 'w') as f:
for item in results:
f.write("%s\n" % item)
def ultimate_evaluate(file_name, model, x_test, y_test, q_ids, answer_key):
y_index = model.predict(x_test)
y_pred = [answer_key[np.argmax(pred)] for pred in y_index]
score = np.mean([pred == truth for pred, truth in zip(y_pred, y_test)])
d = [{'answer':p, 'question_id':i} for p, i in zip(y_pred, q_ids)]
with open(file_name, 'w') as f:
json.dump(d, f)
return score
def bow_model_1(x, y, embedding, epochs=10, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(hidden_dim, activation='tanh')(a1)
b1 = Dense(hidden_dim, activation='tanh')(b1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
b2 = Lambda(lambda x: mean(x, axis=1))(b2)
b2 = Dense(hidden_dim, activation='tanh')(b2)
b2 = Dense(hidden_dim, activation='tanh')(b2)
b = Multiply()([b1, b2])
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=c)
opt = Adam()
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
model.fit(x, y, epochs=epochs, batch_size=64, verbose=2)
return model
def bow_model_2(x, y, embedding, epochs=10, hidden_dim=512):
(vocab_size, vec_dim) = embedding.shape
feat_dim, seq_length, output_dim = len(x[0][0]), len(x[1][0]), len(y[0])
a1 = Input(shape=(feat_dim,))
b1 = Dense(hidden_dim, activation='tanh')(a1)
a2 = Input(shape=(seq_length,))
b2 = Embedding(vocab_size, vec_dim, weights=[embedding], input_length=seq_length, trainable=False)(a2)
b2 = Lambda(lambda x: mean(x, axis=1))(b2)
b2 = Dense(hidden_dim, activation='tanh')(b2)
b = Multiply()([b1, b2])
c = Dense(output_dim, activation='tanh')(b)
c = Dense(output_dim, activation='tanh')(b)
c = Dense(output_dim, activation='softmax')(b)
model = Model(inputs=[a1, a2], outputs=c)
opt = Adam()
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
model.fit(x, y, epochs=epochs, batch_size=64, verbose=2)
return model
if __name__ == "__main__":
print("\n---- Starting Training ----\n")
train_features = load_pickle("data/train_features.pickle")
val_features = load_pickle("data/val_features.pickle")
train_questions = load_json("data/train_questions.json")['questions']
val_questions = load_json("data/val_questions.json")['questions']
train_answers = load_json("data/train_annotations.json")['annotations']
val_answers = load_json("data/val_annotations.json")['annotations']
print(" - Raw data finished loading")
glove_dimension = 100
glove_directory = "glove.6B/glove.6B." + str(glove_dimension) + "d.txt"
glove = load_glove(glove_directory)
print(" - Glove finished loading")
train_raw = extract_questions(train_questions, train_answers)
test_raw = extract_questions(val_questions, val_answers)
answer_list, _ = top_answers(train_raw, k=500)
train_data, test_data, vocabulary = minimize_vocab(train_raw, test_raw, glove)
x_train, y_train, embedding, a_to_i = preprocessing(train_data, train_features, glove, answer_list, vocabulary)
i_to_a = {i:a for a,i in a_to_i.items()}
x_test, y_test, q_test = test_preprocess(test_data, val_features, glove, vocabulary)
print(" - Data preprocessing completed")
# TRAIN AND TEST
bow_1 = bow_model_1(x_train, y_train, embedding)
bow_2 = bow_model_2(x_train, y_train, embedding)
score_1 = ultimate_evaluate("bow-1-preds.json", bow_1, x_test, y_test, q_test, i_to_a)
score_2 = ultimate_evaluate("bow-2-preds.json", bow_2, x_test, y_test, q_test, i_to_a)
print("Bow 1 score: {}".format(score_1))
print("Bow 2 score: {}".format(score_2))
print(" - Model finished training and evaluation")
print("\n---- Program Complete ----\n")