File size: 6,258 Bytes
6f70897 |
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 |
# IMPORTS
import pandas as pd
import nltk
nltk.download("punkt")
from nltk import tokenize
import time
from sklearn.model_selection import train_test_split
from transformers import BertConfig, BertTokenizer, TFBertModel
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow import convert_to_tensor
from tensorflow.keras.layers import Input, Dense
from tensorflow.keras.initializers import TruncatedNormal
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.metrics import CategoricalAccuracy, Precision, Recall
# SET PARAMETERS
DATA_PATH="..."
SAVE_MODELS_TO=".../"
# READ DATA
tab=pd.read_hdf(DATA_PATH)
# PREPARE DATA FOR BERT
def data_to_values(dataframe):
"""Converts data to values.
"""
abstracts=dataframe.Abstract.values
labels=dataframe.Label.values
return abstracts, labels
def tokenize_abstracts(abstracts):
"""For given texts, adds '[CLS]' and '[SEP]' tokens
at the beginning and the end of each sentence, respectively.
"""
t_abstracts=[]
for abstract in abstracts:
t_abstract="[CLS] "
for sentence in tokenize.sent_tokenize(abstract):
t_abstract=t_abstract + sentence + " [SEP] "
t_abstracts.append(t_abstract)
return t_abstracts
tokenizer=BertTokenizer.from_pretrained('bert-base-multilingual-uncased')
def b_tokenize_abstracts(t_abstracts, max_len=512):
"""Tokenizes sentences with the help
of a 'bert-base-multilingual-uncased' tokenizer.
"""
b_t_abstracts=[tokenizer.tokenize(_)[:max_len] for _ in t_abstracts]
return b_t_abstracts
def convert_to_ids(b_t_abstracts):
"""Converts tokens to its specific
IDs in a bert vocabulary.
"""
input_ids=[tokenizer.convert_tokens_to_ids(_) for _ in b_t_abstracts]
return input_ids
def abstracts_to_ids(abstracts):
"""Tokenizes abstracts and converts
tokens to their specific IDs
in a bert vocabulary.
"""
tokenized_abstracts=tokenize_abstracts(abstracts)
b_tokenized_abstracts=b_tokenize_abstracts(tokenized_abstracts)
ids=convert_to_ids(b_tokenized_abstracts)
return ids
def pad_ids(input_ids, max_len=512):
"""Padds sequences of a given IDs.
"""
p_input_ids=pad_sequences(input_ids,
maxlen=max_len,
dtype="long",
truncating="post",
padding="post")
return p_input_ids
def create_attention_masks(inputs):
"""Creates attention masks
for a given seuquences.
"""
masks=[]
for sequence in inputs:
sequence_mask=[float(_>0) for _ in sequence]
masks.append(sequence_mask)
return masks
# CREATE MODEL
def create_model():
config=BertConfig.from_pretrained(
"bert-base-multilingual-uncased",
num_labels=17,
hidden_dropout_prob=0.2,
attention_probs_dropout_prob=0.2)
bert=TFBertModel.from_pretrained(
"bert-base-multilingual-uncased",
config=config)
bert_layer=bert.layers[0]
input_ids_layer=Input(
shape=(512),
name="input_ids",
dtype="int32")
input_attention_masks_layer=Input(
shape=(512),
name="attention_masks",
dtype="int32")
bert_model=bert_layer(
input_ids_layer,
input_attention_masks_layer)
target_layer=Dense(
units=17,
kernel_initializer=TruncatedNormal(stddev=config.initializer_range),
name="target_layer",
activation="sigmoid")(bert_model[1])
model=Model(
inputs=[input_ids_layer, input_attention_masks_layer],
outputs=target_layer,
name="aurora_sdg_mbert_multilabel")
optimizer=Adam(
learning_rate=5e-05,
epsilon=1e-08,
decay=0.01,
clipnorm=1.0)
model.compile(
optimizer=optimizer,
loss="binary_crossentropy",
metrics=[Precision(), Recall()])
return model
abstracts, labels=data_to_values(tab)
ids=abstracts_to_ids(abstracts)
print("Abstracts tokenized, tokens converted to ids.")
padded_ids=pad_ids(ids)
print("Sequences padded.")
train_inputs, temp_inputs, train_labels, temp_labels=train_test_split(padded_ids, labels, random_state=1993, test_size=0.3)
validation_inputs, test_inputs, validation_labels, test_labels=train_test_split(temp_inputs, temp_labels, random_state=1993, test_size=0.5)
print("Data splited into train, validation, test sets.")
train_masks, validation_masks, test_masks=[create_attention_masks(_) for _ in [train_inputs, validation_inputs, test_inputs]]
print("Attention masks created.")
train_inputs, validation_inputs, test_inputs=[convert_to_tensor(_) for _ in [train_inputs, validation_inputs, test_inputs]]
print("Inputs converted to tensors.")
train_labels, validation_labels, test_labels=[convert_to_tensor(_) for _ in [train_labels, validation_labels, test_labels]]
print("Labels converted to tensors.")
train_masks, validation_masks, test_masks=[convert_to_tensor(_) for _ in [train_masks, validation_masks, test_masks]]
print("Masks converted to tensors.")
model=create_model()
print("Model initialized.")
history=model.fit([train_inputs, train_masks], train_labels,
batch_size=16,
epochs=4,
validation_data=([validation_inputs, validation_masks], validation_labels))
model.save(SAVE_MODEL_TO+"mbert_multilabel.h5")
print("Model saved.")
test_score=model.evaluate([test_inputs, test_masks], test_labels,
batch_size=8)
print("Model tested.")
stats=pd.DataFrame(test_score)
stats.to_excel(SAVE_MODEL_TO+"mbert_multilabel_stats.xlsx", index=False)
print("Stats saved.")
|