Commit
·
24f90ea
1
Parent(s):
7f341be
Upload data.py
Browse files
data.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from datasets import load_dataset, Dataset
|
3 |
+
from transformers import BertTokenizerFast
|
4 |
+
import pandas as pd
|
5 |
+
from imblearn.under_sampling import RandomUnderSampler
|
6 |
+
import logging
|
7 |
+
import os
|
8 |
+
|
9 |
+
|
10 |
+
def balance_data(dataset):
|
11 |
+
df = dataset.to_pandas()
|
12 |
+
|
13 |
+
logging.info(f"Balancing {df['label'].value_counts()}")
|
14 |
+
|
15 |
+
rus = RandomUnderSampler(random_state=42, replacement=True)
|
16 |
+
|
17 |
+
X_resampled, y_resampled = rus.fit_resample(
|
18 |
+
df['text'].to_numpy().reshape(-1, 1), df['label'].to_numpy())
|
19 |
+
|
20 |
+
df = pd.DataFrame(
|
21 |
+
{'text': X_resampled.flatten(), 'label': y_resampled})
|
22 |
+
|
23 |
+
logging.info(f"After balancing: {df['label'].value_counts()}")
|
24 |
+
|
25 |
+
return Dataset.from_pandas(df)
|
26 |
+
|
27 |
+
|
28 |
+
def tokenize(dataset):
|
29 |
+
tokenizer = BertTokenizerFast.from_pretrained("neuralmind/bert-large-portuguese-cased")
|
30 |
+
|
31 |
+
dataset = dataset.map(lambda example: tokenizer(
|
32 |
+
example["text"], truncation=True, padding="max_length", max_length=512))
|
33 |
+
|
34 |
+
return dataset
|
35 |
+
|
36 |
+
# This function supports the Notebook version of LID. No usage elsewhere.
|
37 |
+
def tokenize_single_document(text):
|
38 |
+
tokenizer = BertTokenizerFast.from_pretrained("neuralmind/bert-large-portuguese-cased")
|
39 |
+
|
40 |
+
return tokenizer(text, truncation=True, padding="max_length", max_length=512)
|
41 |
+
|
42 |
+
def load_dataloader(domain):
|
43 |
+
|
44 |
+
logging.info(f"Loading {domain} dataset...")
|
45 |
+
|
46 |
+
if domain == 'dslcc':
|
47 |
+
dataset = load_dataset("arubenruben/portuguese_dslcc")
|
48 |
+
else:
|
49 |
+
dataset = load_dataset("Random-Mary-Smith/port_data_random", domain)
|
50 |
+
|
51 |
+
DEBUG = (os.getenv('DEBUG', 'False') == 'True')
|
52 |
+
|
53 |
+
dataset['train'] = balance_data(dataset['train'])
|
54 |
+
|
55 |
+
dataset['test'] = dataset['test'].select(range(min(len(dataset['test']), 10_000)))
|
56 |
+
|
57 |
+
for split in ['train', 'test']:
|
58 |
+
if DEBUG:
|
59 |
+
logging.info("DEBUG MODE: Loading only 100 samples")
|
60 |
+
dataset[split] = dataset[split].select(range(min(len(dataset[split]), 50)))
|
61 |
+
|
62 |
+
dataset = tokenize(dataset)
|
63 |
+
dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'label'])
|
64 |
+
|
65 |
+
# Create Dataloaders
|
66 |
+
train_dataloader = torch.utils.data.DataLoader(dataset['train'], batch_size=int(os.getenv('BATCH_SIZE')), shuffle=True)
|
67 |
+
test_dataloader = torch.utils.data.DataLoader(dataset['test'], batch_size=int(os.getenv('BATCH_SIZE')), shuffle=False)
|
68 |
+
|
69 |
+
return train_dataloader, test_dataloader
|