-
Notifications
You must be signed in to change notification settings - Fork 0
/
toxic-xlm-roberta.py
183 lines (155 loc) · 5.51 KB
/
toxic-xlm-roberta.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
import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import roc_auc_score, roc_curve
import tensorflow as tf
import tensorflow_addons as tfa
from transformers import TFXLMRobertaModel, XLMRobertaConfig
from transformers import AutoTokenizer, XLMRobertaTokenizer
from utils import generate_random_seed
from utils import regular_encode
from utils import load_train_set
from utils import load_test_set
from utils import build_dataset
from utils import build_classifier
from utils import show_training_process
from utils import train_classifier
from utils import predict_with_classifier
from utils import show_roc_auc
# Checking for TPU
try:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
print("Running on TPU ", tpu.master())
except ValueError:
tpu = None
if tpu:
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu)
model_name = "jplu/tf-xlm-roberta-large"
max_seq_len = 256
batch_size_for_xlmr = 8 * strategy.num_replicas_in_sync
else:
strategy = tf.distribute.get_strategy()
physical_devices = tf.config.list_physical_devices("GPU")
for device_idx in range(strategy.num_replicas_in_sync):
tf.config.experimental.set_memory_growth(physical_devices[device_idx], True)
max_seq_len = 256
model_name = "jplu/tf-xlm-roberta-base"
batch_size_for_xlmr = 4 * strategy.num_replicas_in_sync
random.seed(42)
np.random.seed(42)
tf.random.set_seed(42)
tfa.register_all()
xlmr_learning_rate = 1e-5
dataset_dir = "jigsaw-multilingual-toxic-comment-classification"
final_classifier_name = "xlmr_for_toxicity.h5"
xlmroberta_tokenizer = AutoTokenizer.from_pretrained(model_name)
xlmroberta_config = XLMRobertaConfig.from_pretrained(model_name)
sentence_embedding_size = xlmroberta_config.hidden_size
assert max_seq_len <= xlmroberta_config.max_position_embeddings
corpus_for_training = load_train_set(
os.path.join(dataset_dir, "jigsaw-toxic-comment-train.csv"),
text_field="comment_text",
lang_field="lang",
sentiment_fields=[
"toxic",
"severe_toxic",
"obscene",
"threat",
"insult",
"identity_hate",
],
)
assert "en" in corpus_for_training
multilingual_corpus = load_train_set(
os.path.join(dataset_dir, "validation.csv"),
text_field="comment_text",
lang_field="lang",
sentiment_fields=[
"toxic",
],
)
assert "en" not in multilingual_corpus
max_size = 0
print("Multilingual data:")
for language in sorted(list(multilingual_corpus.keys())):
print(" {0}\t\t{1} samples".format(language, len(multilingual_corpus[language])))
assert set(map(lambda cur: cur[1], multilingual_corpus[language])) == {0, 1}
if len(multilingual_corpus[language]) > max_size:
max_size = len(multilingual_corpus[language])
nonenglish_languages = sorted(list(multilingual_corpus.keys()))
corpus_for_validation = dict()
for lang in nonenglish_languages:
random.shuffle(multilingual_corpus[lang])
n = len(multilingual_corpus[lang]) // 2
corpus_for_validation[lang] = multilingual_corpus[lang][0:n]
corpus_for_training[lang] = multilingual_corpus[lang][n:]
del multilingual_corpus[lang]
texts_for_submission = load_test_set(
os.path.join(dataset_dir, "test.csv"),
text_field="content",
lang_field="lang",
id_field="id",
)
for language in sorted(list(texts_for_submission.keys())):
print(" {0}\t\t{1} samples".format(language, len(texts_for_submission[language])))
dataset_for_training, n_batches_per_data = build_dataset(
texts=corpus_for_training,
dataset_size=150000,
tokenizer=xlmroberta_tokenizer,
maxlen=max_seq_len,
batch_size=batch_size_for_xlmr,
shuffle=True,
)
dataset_for_validation, n_batches_per_epoch = build_dataset(
texts=corpus_for_validation,
dataset_size=6000,
tokenizer=xlmroberta_tokenizer,
maxlen=max_seq_len,
batch_size=batch_size_for_xlmr,
shuffle=False,
)
preparing_duration = int(round(time.time() - experiment_start_time))
print(
"Duration of data loading and preparing to the Siamese NN training is "
"{0} seconds.".format(preparing_duration)
)
with strategy.scope():
xlmr_based_classifier = build_classifier(
transformer_name=model_name,
hidden_state_size=sentence_embedding_size,
max_len=max_seq_len,
lr=xlmr_learning_rate,
)
train_classifier(
nn=xlmr_based_classifier,
trainset=dataset_for_training,
steps_per_trainset=n_batches_per_data,
steps_per_epoch=min(5 * n_batches_per_epoch, n_batches_per_data),
validset=dataset_for_validation,
max_duration=int(round(2.0 * 3600.0 - preparing_duration)),
classifier_file_name=final_classifier_name,
)
val_predictions = predict_with_classifier(
texts=corpus_for_validation,
tokenizer=xlmroberta_tokenizer,
maxlen=max_seq_len,
classifier=xlmr_based_classifier,
batch_size=batch_size_for_xlmr,
)
calculated_probas = []
true_labels = []
for lang in val_predictions:
probabilities_, true_labels_ = val_predictions[lang]
calculated_probas.append(probabilities_)
true_labels.append(true_labels_)
calculated_probas = np.concatenate(calculated_probas)
true_labels = np.concatenate(true_labels)
show_roc_auc(y_true=true_labels, probabilities=calculated_probas, label="multi")
final_predictions = predict_with_classifier(
texts=texts_for_submission,
tokenizer=xlmroberta_tokenizer,
maxlen=max_seq_len,
classifier=xlmr_based_classifier,
batch_size=batch_size_for_xlmr,
)