1、适配昇腾处理器：Ascend310、Ascend910系列 2、开发环境：Ascend-cann-toolkit_xxx、Ascend-cann-kernels-xxx(可选)、python3.8 3、下载代码：git clone https://modelers.cn/ShanXi/ko-reranker.git 4、安装依赖：pip install -r examples/requirements.txt 5、推理测试：python examples/inference.py 6、推理脚本：
import argparse
import torch
from openmind import pipeline, is_torch_npu_available
from transformers import AutoModelForCausalLM,AutoModelForSequenceClassification, AutoTokenizer
import numpy as np
from openmind_hub import snapshot_download
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--model_name_or_path",type=str,help="模型路径",default="./",)
args = parser.parse_args()
return args
if is_torch_npu_available():
device = "npu:0"
else:
device = "cpu"
args = parse_args()
if args.model_name_or_path:
model_path = args.model_name_or_path
else:
model_path = snapshot_download('ShanXi/ko-reranker',revision='main',resume_donwload=True,ignore_patterns=['*.h5','*.ot','*.msgpack'])
def exp_normalize(x):
b = x.max()
y = np.exp(x - b)
return y / y.sum()
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSequenceClassification.from_pretrained(model_path)
model.eval()
pairs = [["나는 너를 싫어해", "나는 너를 사랑해"], ["나는 너를 좋아해", "너에 대한 나의 감정은 사랑 일 수도 있어"]]
with torch.no_grad():
inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
scores = exp_normalize(scores.numpy())
print (f'第一个: {scores[0]}, 第二个: {scores[1]}')