forked from gallipoligiuseppe/SentiModel
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
256 lines (231 loc) · 13 KB
/
main.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
import logging
import os
import argparse
from datetime import datetime
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
from datasets import load_dataset
from SentiModel.dataset import *
from SentiModel.model import *
from SentiModel.model_utils import *
from SentiModel.utils import *
from tensorflow.keras.optimizers import Adam
from sklearn.svm import SVC
from sklearn.naive_bayes import MultinomialNB
SEED = 31
np.random.seed(SEED)
tf.random.set_seed(SEED)
tf.keras.utils.set_random_seed(SEED)
def main(args):
try:
if args.mode == 'train' or args.mode == 'test':
bilstm(args)
elif args.baseline == 'lstm':
lstm(args)
elif args.baseline == 'svm' or args.baseline == 'nb':
baseline(args)
except (FileNotFoundError, OSError):
logging.error('File not found - Execution stopped')
return
def bilstm(args):
if args.dataset == 'music':
df_reviews = pd.read_csv('./data/musical_reviews.csv', usecols=['reviewText', 'overall', 'summary'])
dataset = AmazonDataset(df_reviews, SEED)
elif args.dataset == 'imdb':
df_reviews = pd.read_csv('./data/IMDb_movie_reviews.csv', names=['reviewText', 'label'], header=0)
dataset = IMDbDataset(df_reviews, SEED)
elif args.dataset == 'home_multi':
df_reviews = load_dataset('amazon_reviews_multi', split='train').to_pandas()
dataset = AmazonMultiDataset(df_reviews, SEED)
else:
logging.info('Not supported yet')
return
logging.info('Dataset preprocessing...')
dataset.process()
if args.load_data:
dataset.load(ASPECT_PATH, SENTIM_PATH, LEXICON_PATH)
else:
dataset.create_aspect_sentim()
dataset.create_lexicon()
dataset.save(ASPECT_PATH, SENTIM_PATH, LEXICON_PATH)
model = SentiModel(args.encoder, args.max_seq_len, args.n_hidden_layers, True, 256, args.dropout_rate, (dataset.aspect_terms,
dataset.sentim_words), dataset.sentiment_lexicon, MODEL_LOAD_PATH)
opt = Adam(learning_rate=1e-3)
model.compile(optimizer=opt, loss=senti_loss, metrics=[accuracy, precision, recall, f1], run_eagerly=True)
X_train, X_eval, X_test, y_train, y_eval, y_test = dataset.create_splits()
if args.mode == 'train':
callbacks = [TrainingCallback(MODEL_SAVE_PATH, args.validation_step)]
logging.info(f'Training set size: {len(X_train)}')
logging.info(f'Evaluation set size: {len(X_eval)}\n')
logging.info('Start training...')
history = model.fit(x=X_train, y=y_train, epochs=args.n_epochs, batch_size=args.batch_size, validation_data=(X_eval, y_eval),
validation_freq=args.validation_step, shuffle=True, callbacks=callbacks, verbose=1)
logging.info('End training...')
fig, axs = plt.subplots(1, 3, figsize=(20, 4))
plot_metrics(history, 'loss', axs[0], 'Loss per epoch')
plot_metrics(history, 'accuracy', axs[1], 'Accuracy per epoch')
plot_metrics(history, 'precision', axs[2], 'Precision per epoch')
filename = PLOTS_PATH+f'/loss_acc_prec.pdf'
fig.savefig(filename, format='pdf')
fig, axs = plt.subplots(1, 2, figsize=(13, 4))
plot_metrics(history, 'recall', axs[0], 'Recall per epoch')
plot_metrics(history, 'f1', axs[1], 'F1 score per epoch')
filename = PLOTS_PATH+f'/rec_f1.pdf'
fig.savefig(filename, format='pdf')
aspect_terms_att, sentim_words_att = model.return_attention_weights()
topK = 10
fig, ax = plt.subplots(figsize=(10, 5))
title = f'Aspect terms – top {topK} attention weights'
plot_attention_weights(aspect_terms_att, topK, 'minmax', ax, title)
filename = PLOTS_PATH+f'/att_weights_aspects.pdf'
fig.savefig(filename, format='pdf')
fig, ax = plt.subplots(figsize=(10, 5))
title = f'Sentimental words – top {topK} attention weights'
plot_attention_weights(sentim_words_att, topK, 'minmax', ax, title)
filename = PLOTS_PATH+f'/att_weights_senti_words.pdf'
fig.savefig(filename, format='pdf')
elif args.from_pretrained is None:
logging.info(f'No pretrained model was given, so initial weights will be used')
logging.info(f'Test set size: {len(X_test)}')
y_pred, y_pred_prob = evaluate_model(model, 'SentiModel', X_test, y_test, 'test')
fig, axs = plt.subplots(1, 2, figsize=(14, 6))
plot_confusion(y_test, y_pred, axs[0])
plot_ROC(y_test, y_pred_prob, axs[1])
filename = PLOTS_PATH+f'/confusion_roc_test.pdf'
fig.savefig(filename, format='pdf')
get_samples(model, args.max_seq_len, args.n_samples, X_test, SAMPLES_PATH)
review_sample = X_test[31]
input = tf.constant([review_sample])
_, att_weights = model(input, training=False, return_attention=True)
sentence_attention(review_sample, args.max_seq_len, att_weights, 'exp')
def lstm(args):
if args.dataset == 'music':
df_reviews = pd.read_csv('./data/musical_reviews.csv', usecols=['reviewText', 'overall', 'summary'])
dataset = AmazonDataset(df_reviews, SEED)
elif args.dataset == 'imdb':
df_reviews = pd.read_csv('./data/IMDb_movie_reviews.csv', names=['reviewText', 'label'], header=0)
dataset = IMDbDataset(df_reviews, SEED)
elif args.dataset == 'home_multi':
df_reviews = load_dataset('amazon_reviews_multi', split='train').to_pandas()
dataset = AmazonMultiDataset(df_reviews, SEED)
else:
logging.info('Not supported yet')
return
logging.info('Dataset preprocessing...')
dataset.process()
if args.load_data:
dataset.load(ASPECT_PATH, SENTIM_PATH, LEXICON_PATH)
else:
dataset.create_aspect_sentim()
dataset.create_lexicon()
dataset.save(ASPECT_PATH, SENTIM_PATH, LEXICON_PATH)
model = SentiModel(args.encoder, args.max_seq_len, args.n_hidden_layers, False, 256, args.dropout_rate, (dataset.aspect_terms,
dataset.sentim_words), dataset.sentiment_lexicon, MODEL_LOAD_PATH)
opt = Adam(learning_rate=1e-3)
model.compile(optimizer=opt, loss=senti_loss, metrics=[accuracy, precision, recall, f1], run_eagerly=True)
X_train, X_eval, X_test, y_train, y_eval, y_test = dataset.create_splits()
if args.lstm_mode == 'train':
callbacks = [TrainingCallback(MODEL_SAVE_PATH, args.validation_step)]
logging.info(f'Training set size: {len(X_train)}')
logging.info(f'Evaluation set size: {len(X_eval)}\n')
logging.info('Start training...')
history = model.fit(x=X_train, y=y_train, epochs=args.n_epochs, batch_size=args.batch_size, validation_data=(X_eval, y_eval),
validation_freq=args.validation_step, shuffle=True, callbacks=callbacks, verbose=1)
logging.info('End training...')
elif args.from_pretrained is None:
logging.info(f'No pretrained model was given, so initial weights will be used')
logging.info(f'Test set size: {len(X_test)}')
y_pred, y_pred_prob = evaluate_model(model, 'SentiModel_LSTM', X_test, y_test, 'test')
def baseline(args):
if args.dataset == 'music':
df_reviews = pd.read_csv('./data/musical_reviews.csv', usecols=['reviewText', 'overall', 'summary'])
dataset = AmazonDataset(df_reviews, SEED)
elif args.dataset == 'imdb':
df_reviews = pd.read_csv('./data/IMDb_movie_reviews.csv', names=['reviewText', 'label'], header=0)
dataset = IMDbDataset(df_reviews, SEED)
elif args.dataset == 'home_multi':
df_reviews = load_dataset('amazon_reviews_multi', split='train').to_pandas()
dataset = AmazonMultiDataset(df_reviews, SEED)
else:
logging.info('Not supported yet')
return
logging.info('Dataset preprocessing...')
dataset.process()
X_train, X_eval, X_test, y_train, y_eval, y_test = dataset.create_splits()
X_train_tfidf = dataset.vectorizer.fit_transform(X_train).toarray()
X_eval_tfidf = dataset.vectorizer.transform(X_eval).toarray()
X_test_tfidf = dataset.vectorizer.transform(X_test).toarray()
if args.baseline == 'svm':
model_SVM = SVC(C=1, kernel='rbf', degree=3, gamma='scale', random_state=SEED)
logging.info('Start fitting...')
model_SVM.fit(X_train_tfidf, y_train)
logging.info('End fitting...')
logging.info(f'Evaluation set size: {len(X_test)}')
_ = evaluate_model(model_SVM, 'SVM', X_eval_tfidf, y_eval, 'evaluation')
logging.info('\n#######################################################\n')
logging.info(f'Test set size: {len(X_test)}')
_ = evaluate_model(model_SVM, 'SVM', X_test_tfidf, y_test, 'test')
elif args.baseline == 'nb':
model_NB = MultinomialNB(alpha=1)
logging.info('Start fitting...')
model_NB.fit(X_train_tfidf, y_train)
logging.info('End fitting...')
logging.info(f'Evaluation set size: {len(X_test)}')
_ = evaluate_model(model_NB, 'Naive Bayes', X_eval_tfidf, y_eval, 'evaluation')
logging.info('\n#######################################################\n')
logging.info(f'Test set size: {len(X_test)}')
_ = evaluate_model(model_NB, 'Naive Bayes', X_test_tfidf, y_test, 'test')
def arg_parser(args=None):
parser = argparse.ArgumentParser(description='BiLSTM + Attention for Sentiment Analysis')
subparser = parser.add_subparsers(required=True, dest='mode')
parser_aux = argparse.ArgumentParser(add_help=False)
parser_aux_dataset = argparse.ArgumentParser(add_help=False)
parser_aux_dataset.add_argument('--dataset', type=str, choices=['music', 'imdb', 'home_multi'], required=True, help='dataset (Amazon Music Reviews or IMDb Movie Reviews or Amazon Multilingual Home Reviews)')
parser_aux.add_argument('--encoder', type=str, choices=['bert', 'roberta'], required=True, help='embedding model (BERT or RoBERTa)')
parser_aux.add_argument('--n_epochs', type=int, default=5, help='number of epochs')
parser_aux.add_argument('--load_data', action='store_true', default=False, help='load precomputed aspect/sentim/lexicon')
parser_aux.add_argument('--from_pretrained', type=str, default=None, metavar='PRETRAINED PATH', help='path of pretrained model')
parser_aux.add_argument('--max_seq_len', type=int, default=64, help='maximum sequence length to process')
parser_aux.add_argument('--n_hidden_layers', type=int, default=1, help='number of hidden layers')
parser_aux.add_argument('--dropout_rate', type=float, default=0.2, help='dropout rate')
parser_aux.add_argument('--lr', type=float, default=1e-3, help='learning rate')
parser_aux.add_argument('--batch_size', type=int, default=32, help='batch size')
parser_aux.add_argument('--validation_step', type=int, default=1, help='to perform validation after each validation_step epoch(s)')
parser_aux.add_argument('--n_samples', type=int, default=5, help='number of sentence samples to save')
subparser.add_parser('train', help='BiLSTM training', parents=[parser_aux_dataset, parser_aux])
subparser.add_parser('test', help='BiLSTM test', parents=[parser_aux_dataset, parser_aux])
parser_baseline = subparser.add_parser('baseline', help='LSTM, SVM, Naive Bayes')
subparser_baseline = parser_baseline.add_subparsers(required=True, dest='baseline')
parser_baseline_lstm = subparser_baseline.add_parser('lstm', help='LSTM')
subparser_lstm = parser_baseline_lstm.add_subparsers(required=True, dest='lstm_mode')
subparser_lstm.add_parser('train', help='LSTM training', parents=[parser_aux_dataset, parser_aux])
subparser_lstm.add_parser('test', help='LSTM test', parents=[parser_aux_dataset, parser_aux])
subparser_baseline.add_parser('svm', help='SVM', parents=[parser_aux_dataset])
subparser_baseline.add_parser('nb', help='Naive Bayes', parents=[parser_aux_dataset])
return parser.parse_args(args=args)
def init_path(args):
timestamp = datetime.now().strftime('%d%H%M')
global ROOT_PATH, OUTPUT_PATH, PLOTS_PATH, SAMPLES_PATH, ASPECT_PATH, SENTIM_PATH, LEXICON_PATH, MODEL_LOAD_PATH, MODEL_SAVE_PATH
ROOT_PATH = './executions/'
OUTPUT_PATH = ROOT_PATH+timestamp+'/'
PLOTS_PATH = OUTPUT_PATH+'plots/'
SAMPLES_PATH = OUTPUT_PATH+'/samples/'
ASPECT_PATH = ROOT_PATH+f'aspect_set_{args.dataset}.pickle'
SENTIM_PATH = ROOT_PATH+f'sentim_set_{args.dataset}.pickle'
LEXICON_PATH = ROOT_PATH+f'sentim_lexicon_{args.dataset}.pickle'
if (args.mode == 'train' or args.mode == 'test' or args.baseline == 'lstm') and args.from_pretrained is not None:
MODEL_LOAD_PATH = ROOT_PATH+args.from_pretrained+'/model.h5'
else:
MODEL_LOAD_PATH = None
MODEL_SAVE_PATH = OUTPUT_PATH+'model.h5'
os.makedirs(OUTPUT_PATH, exist_ok=True)
os.makedirs(PLOTS_PATH, exist_ok=True)
os.makedirs(SAMPLES_PATH, exist_ok=True)
if __name__ == '__main__':
args = arg_parser()
init_path(args)
logging.basicConfig(filename=OUTPUT_PATH+'log.txt', filemode='w', format='%(message)s', level=logging.INFO)
logging.info(f'Run with parameters:\n{args}\n')
main(args)