-
Notifications
You must be signed in to change notification settings - Fork 1
/
MisRoBÆRTa.py
259 lines (219 loc) · 11.8 KB
/
MisRoBÆRTa.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
# coding: utf-8
__author__ = "Ciprian-Octavian Truică"
__copyright__ = "Copyright 2021, University Politehnica of Bucharest"
__license__ = "GNU GPL"
__version__ = "0.1"
__email__ = "ciprian.truica@upb.ro"
__status__ = "Production"
import os
import sys
# helpers
import time
# classification
from sklearn.metrics import accuracy_score, classification_report
from sklearn.feature_extraction.text import TfidfVectorizer
import matplotlib.pyplot as plt
from sklearn.metrics import accuracy_score, recall_score, precision_score, classification_report, plot_confusion_matrix, confusion_matrix
from sklearn.model_selection import train_test_split
import numpy as np
import matplotlib.pyplot as plt
import pickle
import re
import pandas as pd
import sys
import os
import os.path
# split data set
import multiprocessing as mp
# import keras
from keras.models import Sequential,load_model
from keras.layers import Dense, Embedding, GRU, Dropout, LSTM, Bidirectional, SimpleRNN, Input, Concatenate
# from keras.layers import Attention, Concatenate, Input
from keras.layers.embeddings import Embedding
from keras.initializers import Constant
from keras.optimizers import Adam
from keras.utils import plot_model, np_utils
from keras.callbacks import EarlyStopping, ModelCheckpoint
from tensorflow.python.keras.preprocessing.text import Tokenizer
from tensorflow.python.keras.preprocessing.sequence import pad_sequences
import math
from keras import Model
import tensorflow as tf
from keras.layers import Conv1D, Flatten, MaxPooling1D, Reshape
import random as rnd
from simpletransformers.language_representation import RepresentationModel
from sentence_transformers import SentenceTransformer
TFIDF_FOUT = "./models/tfidf_features"
MAX_FEATURES = 5000
MIN_DF = 5
# MAX_DF = 1
MAX_FEATURES=5000
LOWER = True
USE_IDF = True
SMOOTH_IDF =True
ANALYZER = 'word'
accuracies= []
precisions_micro = []
precisions_macro = []
recalls_micro = []
recalls_macro = []
execution_time = []
def evaluate(y_test, y_pred, modelName='GRU', iters=0):
y_pred_norm = []
for elem in y_pred:
line = [ 0 ] * len(elem)
try:
# if an error appears here
# get a random class
elem[np.isnan(elem)] = 0
line[elem.tolist().index(max(elem.tolist()))] = 1
except:
print("Error for getting predicted class")
print(elem.tolist())
line[rnd.randint(0, len(elem)-1)] = 1
y_pred_norm.append(line)
y_p = np.argmax(np.array(y_pred_norm), 1)
y_t = np.argmax(np.array(y_test), 1)
accuracy = accuracy_score(y_t, y_p)
accuracies.append(accuracy)
precision_micro = precision_score(y_t, y_p, average='micro')
precisions_micro.append(precision_micro)
precision_macro = precision_score(y_t, y_p, average='macro')
precisions_macro.append(precision_macro)
recall_micro = recall_score(y_t, y_p, average='micro')
recalls_micro.append(recall_micro)
recall_macro = recall_score(y_t, y_p, average='macro')
recalls_macro.append(recall_macro)
print("Accuracy", accuracy)
print("Precision Micro", precision_micro)
print("Precision Macro", precision_macro)
print("Recall Micro", recall_micro)
print("Recall Macro", recall_macro)
print("Report", classification_report(y_t, y_p))
print(confusion_matrix(y_t, y_p))
return y_p, y_t
def splitDataset(X, y):
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30, shuffle = True, stratify = y)
return X_train, X_test, np.array(y_train), np.array(y_test)
def getBERTEncoding(X_train, X_test, use_cuda=False):
model = RepresentationModel(model_type="bert", model_name="bert-large-uncased", use_cuda=use_cuda)
X_train_BERT = model.encode_sentences(X_train, combine_strategy="mean")
X_test_BERT = model.encode_sentences(X_test, combine_strategy="mean")
X_train_BERT = X_train_BERT.reshape((X_train_BERT.shape[0], 1, X_train_BERT.shape[1]))
X_test_BERT = X_test_BERT.reshape((X_test_BERT.shape[0], 1, X_test_BERT.shape[1]))
return X_train_BERT, X_test_BERT
def getRoBERTaEncoding(X_train, X_test, use_cuda=False):
model = RepresentationModel(model_type="roberta", model_name="roberta-base", use_cuda=use_cuda)
X_train_RoBERTa = model.encode_sentences(X_train, combine_strategy="mean")
X_test_RoBERTa = model.encode_sentences(X_test, combine_strategy="mean")
X_train_RoBERTa = X_train_RoBERTa.reshape((X_train_RoBERTa.shape[0], 1, X_train_RoBERTa.shape[1]))
X_test_RoBERTa = X_test_RoBERTa.reshape((X_test_RoBERTa.shape[0], 1, X_test_RoBERTa.shape[1]))
return X_train_RoBERTa, X_test_RoBERTa
def getBARTEncoding(X_train, X_test, use_cuda=False):
model = SentenceTransformer('facebook/bart-large')
# model = RepresentationModel(model_type="bart", model_name="facebook/bart-large", use_cuda=use_cuda)
# X_train_BART = model.encode_sentences(X_train, combine_strategy=None)
X_train_BART = model.encode(X_train)
# X_test_BART = model.encode_sentences(X_test, combine_strategy=None)
X_test_BART = model.encode(X_test)
print(X_train_BART)
X_train_BART = X_train_BART.reshape((X_train_BART.shape[0], 1, X_train_BART.shape[1]))
X_test_BART = X_test_BART.reshape((X_test_BART.shape[0], 1, X_test_BART.shape[1]))
return X_train_BART, X_test_BART
def prepareYTrainTestNN(y_train, y_test, num_classes):
y_train = np_utils.to_categorical(y_train, num_classes=num_classes)
y_test = np_utils.to_categorical(y_test, num_classes=num_classes)
return y_train, y_test
if __name__ =="__main__":
FIN = sys.argv[1]
USE_CUDA = bool(int(sys.argv[2])) # 0 - False, 1 - True
print(USE_CUDA)
NUM_ITER = int(sys.argv[3])
dataSet = pd.read_csv(FIN, encoding = "utf-8")
labels = dataSet['label'].unique()
num_classes = len(labels)
id2label = {}
idx = 0
for label in labels:
id2label[idx] = label
dataSet.loc[dataSet['label'] == label, 'label'] = idx
idx += 1
for key in id2label:
print(key, id2label[key])
X = dataSet['content'].astype(str).to_list()
X = X[ : 10] + X[-10:]
y = dataSet['label'].astype(int).to_list()
y = y[ : 10] + y[-10:]
epochs_n = 100
filters = 64
units = 128
no_attributes = int(units * 2) # no units
# no_attributes = units # no units
kernel_size = int(no_attributes/2)
# randomizing features
for idx in range(0, NUM_ITER):
start_time = time.time()
# Sampling examples
X_train, X_test, y_train, y_test = splitDataset(X, y)
x_vec_train = [ [], [], [], [] ]
x_vec_test = [ [], [], [], [] ]
print(y_train.shape, y_test.shape)
y_vec_train, y_vec_test = prepareYTrainTestNN(y_train, y_test, num_classes)
print(y_vec_train.shape, y_vec_test.shape)
print("Bart")
x_vec_train[0], x_vec_test[0] = getBARTEncoding(X_train, X_test, USE_CUDA)
print(x_vec_train[0].shape)
print(x_vec_test[0].shape)
# RoBERTa encodings
print("RoBERTa")
x_vec_train[1], x_vec_test[1] = getRoBERTaEncoding(X_train, X_test, USE_CUDA)
print(x_vec_train[1].shape)
print(x_vec_test[1].shape)
input_bart = Input(shape=(x_vec_train[0].shape[1], x_vec_train[0].shape[2]), name = 'BART_Input')
model_bart = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'BART_BiLSTM_1')(input_bart)
model_bart = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'BART_BiLSTM_2')(input_bart)
model_bart = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'BART_BiLSTM_3')(input_bart)
model_bart = Reshape((no_attributes, 1), name = 'BART_Reshape_1')(model_bart) # reshape to number of units
model_bart = Conv1D(filters = filters, kernel_size=kernel_size, activation='relu', name = 'BART_CNN_1')(model_bart)
model_bart = MaxPooling1D(name='BART_MaxPolling_1')(model_bart)
model_bart = Flatten(name='BART_Flatten_1')(model_bart)
input_roberta = Input(shape=(x_vec_train[1].shape[1], x_vec_train[1].shape[2]), name = 'BERT_Input')
model_roberta = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'RoBERTa_BiLSTM_1')(input_roberta)
model_roberta = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'RoBERTa_BiLSTM_2')(input_roberta)
model_roberta = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'RoBERTa_BiLSTM_3')(input_roberta)
model_roberta = Reshape((no_attributes, 1), name = 'RoBERTa_Reshape_1')(model_roberta) # reshape to number of units
model_roberta = Conv1D(filters = filters, kernel_size=kernel_size, activation='relu', name = 'RoBERTa_CNN_1')(model_roberta)
model_roberta = MaxPooling1D(name='RoBERTa_MaxPolling_1')(model_roberta)
model_roberta = Flatten(name='RoBERTa_Flatten_1')(model_roberta)
# model_unigram = Model(inputs=input_unigram, outputs=model_unigram)
combined = Concatenate(name='Model_Concat')([model_bart, model_roberta])
combined = Reshape((1, combined.shape[1]), name = 'COMB_Reshape_1')(combined) # reshape to number of units
combined = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'COMB_BiLSTM_1')(combined)
combined = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'COMB_BiLSTM_2')(combined)
combined = Bidirectional(LSTM(units = units, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True), name = 'COMB_BiLSTM_3')(combined)
combined = Reshape((no_attributes, 1), name = 'COMB_Reshape_2')(combined) # reshape to number of units
combined = Conv1D(filters = filters, kernel_size=kernel_size, activation='relu', name = 'COMB_CNN_1')(combined)
combined = MaxPooling1D(name='COMB_MaxPolling_1')(combined)
combined = Flatten(name='COMB_Flatten_1')(combined)
output = Dense(units=num_classes, activation = 'softmax', name = 'Model_Output')(combined) #sigmoid #softmax
# model = Model(inputs=[model_unigram.input, model_bigram.input, model_trigram.input], outputs=output)
model = Model(inputs=[input_bart, input_roberta], outputs=output, name="BART-RoBERTa-BiLSTM-CNN")
model.compile(loss = 'categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])
print(model.summary())
# plot_model(model, show_shapes=True, dpi=90)
es = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=20)
# mc = ModelCheckpoint(fileName, monitor='val_accuracy', mode='max', verbose=1, save_best_only=True)
history = model.fit(x = [x_vec_train[0], x_vec_train[1]], y = y_vec_train, epochs=epochs_n, verbose=True, validation_data=([x_vec_test[0], x_vec_test[1]], y_vec_test), batch_size=1000, callbacks=[es])
y_pred = model.predict([x_vec_test[0], x_vec_test[1]], verbose=False)
y_p, y_t = evaluate(y_vec_test, y_pred, modelName=model.name, iters=idx)
end_time = time.time()
execution_time.append(end_time - start_time)
print("Time taken to train: ", end_time - start_time)
print("\n\n========================================\n\n")
print("BART-RoBERTa-3LSTM-CNN")
print("Accuracy",np.mean(accuracies), np.std(accuracies))
print("Precision Micro",np.mean(precisions_micro), np.std(precisions_micro))
print("Precision Macro",np.mean(precisions_macro), np.std(precisions_macro))
print("Recall Micro",np.mean(recalls_micro), np.std(recalls_micro))
print("Recall Macro",np.mean(recalls_macro), np.std(recalls_macro))
print("Execution Time",np.mean(execution_time), np.std(execution_time))