koelectra small v3 nsmc

提供商daekeun-ml
分类text-classification
许可证mit
下载量2.6M
星标0

简介

Koelectra-small-v3-nsmc 是一款针对韩语文本分类优化的轻量化模型。它基于 ELECTRA 架构，并在 NSMC（Naver 电影评论数据集）上进行了精调，擅长处理韩语的情感分析和文本分类任务。对于中国开发者而言，该模型最大的优势在于其“小而美”的特性，推理速度快且资源占用低，非常适合部署在端侧设备或作为韩语预处理流水线的一部分。如果你需要快速构建一个韩语评论分析工具，而不想部署庞大的通用大模型，这是一个极佳的替代方案。

核心亮点

  • 专为韩语情感分析优化，识别评论情绪精准
  • 轻量化架构，推理延迟低，部署成本极低
  • 基于 NSMC 数据集精调，在电影评论场景表现强
  • MIT 协议开源，方便商业化集成与二次开发

使用方法

安装依赖
# 安装 Hugging Face transformers
pip install transformers torch
SDK 使用
# 使用 transformers 加载模型
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("daekeun-ml/koelectra-small-v3-nsmc")
tokenizer = AutoTokenizer.from_pretrained("daekeun-ml/koelectra-small-v3-nsmc")

Hugging Face 下载

我们推荐使用命令行或者 Hugging Face Hub SDK 来进行模型的下载。

操作指引：在下载前，请先通过如下命令安装 huggingface_hub：

操作指引 
pip install -U huggingface_hub

命令行下载

下载完整模型库

下载完整模型库 
huggingface-cli download daekeun-ml/koelectra-small-v3-nsmc

下载单个文件到指定本地文件夹（以下载 config.json 到当前路径下 ./dir 目录为例）

下载单个文件到指定本地文件夹（以下载 config.json 到当前路径下 ./dir 目录为例） 
huggingface-cli download daekeun-ml/koelectra-small-v3-nsmc config.json --local-dir ./dir

更多命令行下载选项，可参见官方文档

SDK 下载

SDK 下载 
# 模型下载
from huggingface_hub import snapshot_download
model_dir = snapshot_download('daekeun-ml/koelectra-small-v3-nsmc')

Git 下载

请确保 lfs 已经被正确安装

Git 下载 
git lfs install
git clone https://huggingface.co/daekeun-ml/koelectra-small-v3-nsmc

如果您希望跳过 lfs 大文件下载，可以使用如下命令

跳过 LFS 
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/daekeun-ml/koelectra-small-v3-nsmc

模型文件托管在 Hugging Face Hub，使用 HF CLI / SDK / Git 直接下载，不经过本站。

PyTorch / Transformers 使用

安装 Transformers

安装 Transformers 
pip install -U transformers torch

模型加载和推理

模型加载和推理 
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('daekeun-ml/koelectra-small-v3-nsmc')
tokenizer = AutoTokenizer.from_pretrained('daekeun-ml/koelectra-small-v3-nsmc')

完整文档

来源: HuggingFace

---
language:
- ko
tags:

  • classification

license: mit
datasets:
  • nsmc

widget:
  • text: "불후의 명작입니다! 이렇게 감동적인 내용은 처음이에요"

example_title: "Positive"
  • text: "시간이 정말 아깝습니다. 10점 만점에 1점도 아까워요.."

example_title: "Negative"
metrics:
  • accuracy

  • f1

  • precision

  • recall- accuracy

---

Sentiment Binary Classification (fine-tuning with KoELECTRA-Small-v3 model and Naver Sentiment Movie Corpus dataset)

Usage (Amazon SageMaker inference applicable)

It uses the interface of the SageMaker Inference Toolkit as is, so it can be easily deployed to SageMaker Endpoint.

inference_nsmc.py

python
import json
import sys
import logging
import torch
from torch import nn
from transformers import ElectraConfig
from transformers import ElectraModel, AutoTokenizer, ElectraTokenizer, ElectraForSequenceClassification

logging.basicConfig(
    level=logging.INFO, 
    format='[{%(filename)s:%(lineno)d} %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler(filename='tmp.log'),
        logging.StreamHandler(sys.stdout)
    ]
)
logger = logging.getLogger(__name__)


max_seq_length = 128
classes = ['Neg', 'Pos']


tokenizer = AutoTokenizer.from_pretrained("daekeun-ml/koelectra-small-v3-nsmc")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


def model_fn(model_path=None):
    ####
    # If you have your own trained model
    # Huggingface pre-trained model: 'monologg/koelectra-small-v3-discriminator'
    ####
    #config = ElectraConfig.from_json_file(f'{model_path}/config.json')
    #model = ElectraForSequenceClassification.from_pretrained(f'{model_path}/model.pth', config=config)
    
    # Download model from the Huggingface hub
    model = ElectraForSequenceClassification.from_pretrained('daekeun-ml/koelectra-small-v3-nsmc')   
    model.to(device)
    return model


def input_fn(input_data, content_type="application/jsonlines"): 
    data_str = input_data.decode("utf-8")
    jsonlines = data_str.split("\n")
    transformed_inputs = []


for jsonline in jsonlines:
        text = json.loads(jsonline)["text"][0]
        logger.info("input text: {}".format(text))          
        encode_plus_token = tokenizer.encode_plus(
            text,
            max_length=max_seq_length,
            add_special_tokens=True,
            return_token_type_ids=False,
            padding="max_length",
            return_attention_mask=True,
            return_tensors="pt",
            truncation=True,
        )
        transformed_inputs.append(encode_plus_token)
        
    return transformed_inputs


def predict_fn(transformed_inputs, model):
    predicted_classes = []
    
    for data in transformed_inputs:
        data = data.to(device)
        output = model(**data)


softmax_fn = nn.Softmax(dim=1)
        softmax_output = softmax_fn(output[0])
        _, prediction = torch.max(softmax_output, dim=1)


predicted_class_idx = prediction.item()
        predicted_class = classes[predicted_class_idx]
        score = softmax_output[0][predicted_class_idx]
        logger.info("predicted_class: {}".format(predicted_class))


prediction_dict = {}
        prediction_dict["predicted_label"] = predicted_class
        prediction_dict['score'] = score.cpu().detach().numpy().tolist()


jsonline = json.dumps(prediction_dict)
        logger.info("jsonline: {}".format(jsonline))        
        predicted_classes.append(jsonline)


predicted_classes_jsonlines = "\n".join(predicted_classes)
    return predicted_classes_jsonlines


def output_fn(outputs, accept="application/jsonlines"):
    return outputs, accept

test.py

python
>>> from inference_nsmc import model_fn, input_fn, predict_fn, output_fn
>>> with open('samples/nsmc.txt', mode='rb') as file:
>>>     model_input_data = file.read()
>>> model = model_fn()
>>> transformed_inputs = input_fn(model_input_data)
>>> predicted_classes_jsonlines = predict_fn(transformed_inputs, model)
>>> model_outputs = output_fn(predicted_classes_jsonlines)
>>> print(model_outputs[0])    
   
[{inference_nsmc.py:47} INFO - input text: 이 영화는 최고의 영화입니다
[{inference_nsmc.py:47} INFO - input text: 최악이에요. 배우의 연기력도 좋지 않고 내용도 너무 허접합니다
[{inference_nsmc.py:77} INFO - predicted_class: Pos
[{inference_nsmc.py:84} INFO - jsonline: {"predicted_label": "Pos", "score": 0.9619030952453613}
[{inference_nsmc.py:77} INFO - predicted_class: Neg
[{inference_nsmc.py:84} INFO - jsonline: {"predicted_label": "Neg", "score": 0.9994170665740967}
{"predicted_label": "Pos", "score": 0.9619030952453613}
{"predicted_label": "Neg", "score": 0.9994170665740967}

Sample data (samples/nsmc.txt)

code
{"text": ["이 영화는 최고의 영화입니다"]}
{"text": ["최악이에요. 배우의 연기력도 좋지 않고 내용도 너무 허접합니다"]}

References

  • KoELECTRA: https://github.com/monologg/KoELECTRA
  • Naver Sentiment Movie Corpus Dataset: https://github.com/e9t/nsmc
查看官方来源