lexiconium's picture
Update README.md
b9af64e
|
raw
history blame
1.6 kB
metadata
language:
  - ko
tags:
  - gpt2
license: cc-by-nc-sa-4.0

Example

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM 

tokenizer = AutoTokenizer.from_pretrained(
    "CheonggyeMountain-Sherpa/kogpt-trinity-punct-wrapper",
    revision="punct_wrapper-related_words-overfit",  # or punct_wrapper-related_words-minevalloss
    bos_token="<s>",
    eos_token="</s>",
    unk_token="<unk>",
    pad_token="<pad>",
    mask_token="<mask>",
)
model = AutoModelForCausalLM.from_pretrained(
    "CheonggyeMountain-Sherpa/kogpt-trinity-punct-wrapper",
    revision="punct_wrapper-related_words-overfit",  # or punct_wrapper-related_words-minevalloss
    pad_token_id=tokenizer.eos_token_id,
).to(device="cuda")
model.eval()

prompt = "์„์–‘์ด ๋ณด์ด๋Š” ๊ฒฝ์น˜"
wrapped_prompt = f"@{prompt}@<usr>\n"
with torch.no_grad():
    tokens = tokenizer.encode(wrapped_prompt, return_tensors="pt").to(device="cuda")
    gen_tokens = model.generate(
        tokens,
        max_length=64,
        repetition_penalty=2.0,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
        bos_token_id=tokenizer.bos_token_id,
        top_k=16,
        top_p=0.8,
    )
    generated = tokenizer.decode(gen_tokens[0][len(tokens[0]):])
 
print(generated)
# ํ•ด๊ฐ€ ์ง€๊ณ  ์žˆ์„ ๋ฌด๋ ต
# ๋‚˜๋Š” ์„์–‘์„ ๋ณด๋Ÿฌ ๊ฐ„๋‹ค
# ๋ถ‰์€ ํ•˜๋Š˜๊ณผ ํ•˜์–€ ๊ตฌ๋ฆ„์ด ๋‚˜๋ฅผ ๋ฐ˜๊ฒจ์ค„ ๊ฒƒ ๊ฐ™์•„์„œ๋ฆฌ
# ํ•˜์ง€๋งŒ ๋‚ด๊ฐ€ ๋ณธ ํ•ด๋Š” ์ €๋ฌผ์–ด๋งŒ ๊ฐ€๊ณ 
# ๊ตฌ๋ฆ„๋งˆ์ € ์ž์ทจ๋ฅผ ๊ฐ์ถ˜ ์–ด๋‘ ๋งŒ์ด ๋‚จ์•„์žˆ์„ ๋ฟ์ด๋„ค
# ๋‚ด๊ฐ€ ํƒ„ ๋ฐฐ๋Š” ๋ณด์ด์ง€๋„ ์•Š๊ณ