Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- ko
|
4 |
+
tags:
|
5 |
+
- roberta
|
6 |
+
license:
|
7 |
+
- mit
|
8 |
+
---
|
9 |
+
|
10 |
+
## ํ๋ จ ์ฝ๋
|
11 |
+
|
12 |
+
```python
|
13 |
+
from datasets import load_dataset
|
14 |
+
from tokenizers import ByteLevelBPETokenizer
|
15 |
+
|
16 |
+
tokenizer = ByteLevelBPETokenizer(unicode_normalizer="nfkc", trim_offsets=True)
|
17 |
+
ds = load_dataset("Bingsu/my-korean-training-corpus", use_auth_token=True)
|
18 |
+
# ๊ณต๊ฐ๋ ๋ฐ์ดํฐ๋ฅผ ์ฌ์ฉํ ๊ฒฝ์ฐ
|
19 |
+
# ds = load_dataset("cc100", lang="ko") # 50GB
|
20 |
+
|
21 |
+
|
22 |
+
# ์ด ๋ฐ์ดํฐ๋ 35GB์ด๊ณ , ๋ฐ์ดํฐ๊ฐ ๋๋ฌด ๋ง์ผ๋ฉด ์ปดํจํฐ๊ฐ ํฐ์ ธ์ ์ผ๋ถ๋ง ์ฌ์ฉํ์ต๋๋ค.
|
23 |
+
ds_sample = ds["train"].train_test_split(0.35, seed=20220819)["test"]
|
24 |
+
|
25 |
+
|
26 |
+
def gen_text(batch_size: int = 5000):
|
27 |
+
for i in range(0, len(ds_sample), batch_size):
|
28 |
+
yield ds_sample[i : i + batch_size]["text"]
|
29 |
+
|
30 |
+
|
31 |
+
tokenizer.train_from_iterator(
|
32 |
+
gen_text(),
|
33 |
+
vocab_size=50265,
|
34 |
+
min_frequency=2,
|
35 |
+
special_tokens=[
|
36 |
+
"<s>",
|
37 |
+
"<pad>",
|
38 |
+
"</s>",
|
39 |
+
"<unk>",
|
40 |
+
"<mask>",
|
41 |
+
],
|
42 |
+
)
|
43 |
+
tokenizer.save("my_tokenizer.json")
|
44 |
+
```
|
45 |
+
|
46 |
+
์ฝ 7์๊ฐ ์๋ชจ (i5-12600 non-k)
|
47 |
+
![image](https://i.imgur.com/LNNbtGH.png)
|
48 |
+
|
49 |
+
## ์ฌ์ฉ๋ฒ
|
50 |
+
|
51 |
+
#### 1.
|
52 |
+
|
53 |
+
```python
|
54 |
+
tokenizer = AutoTokenizer.from_pretrained("Bingsu/BBPE_tokenizer_test")
|
55 |
+
|
56 |
+
# tokenizer๋ RobertaTokenizerFast ํด๋์ค๊ฐ ๋ฉ๋๋ค.
|
57 |
+
```
|
58 |
+
|
59 |
+
#### 2.
|
60 |
+
|
61 |
+
`tokenizer.json`ํ์ผ์ ๋จผ์ ๋ค์ด๋ฐ์ต๋๋ค.
|
62 |
+
|
63 |
+
```python
|
64 |
+
from transformers import BartTokenizerFast, BertTokenizerFast
|
65 |
+
|
66 |
+
bart_tokenizer = BartTokenizerFast(tokenizer_file="tokenizer.json")
|
67 |
+
bert_tokenizer = BertTokenizerFast(tokenizer_file="tokenizer.json")
|
68 |
+
```
|
69 |
+
|
70 |
+
roberta์ ๊ฐ์ด BBPE๋ฅผ ์ฌ์ฉํ bart๋ ๋ฌผ๋ก ์ด๊ณ bert์๋ ๋ถ๋ฌ์ฌ ์ ์์ต๋๋ค.
|
71 |
+
๋ค๋ง ์ด๋ ๊ฒ ๋ถ๋ฌ์์ ๊ฒฝ์ฐ, model_max_len์ด ์ง์ ์ด ๋์ด์์ง ์์ผ๋ ์ง์ ํด์ผ ํฉ๋๋ค.
|