在昇腾 Ascend 910 (torch_npu, npu:0) 上运行 LayoutLM 表单实体识别 (token classification) 模型 philschmid/layoutlm-funsd。
LayoutLMForTokenClassification(hidden_size=768, 12 层, 12 头, vocab=30522, model_type=layoutlm)npu:0,代码无任何 CUDA 调用requirements.txtJSON 中给出每行/每词的文本 tokens 与 0~1000 归一化坐标框 boxes([x0, y0, x1, y1] 整数):
{
"tokens": ["INVOICE", "No.", "1029", "DATE", "2023-06-15"],
"boxes": [[400, 40, 620, 100], [60, 150, 140, 190], [150, 150, 250, 190], [300, 150, 390, 190], [400, 150, 560, 190]]
}LayoutLMTokenizer(BERT WordPiece)会把 word 拆成 subword,并把该 word 的 box 复制到每个 subword,再在首尾加 [CLS]/[SEP]。
权重缓存目录:./model_weights/philschmid__layoutlm-funsd/。程序启动时优先加载本地缓存,未命中才从镜像下载:
export HF_ENDPOINT=https://hf-mirror.com
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download('philschmid/layoutlm-funsd', local_dir='./model_weights/philschmid__layoutlm-funsd', max_workers=2)
"python3 inference.py --json demo.json
# 或
python3 inference.py --text "Invoice No. 1029"输出逐 token 预测(标签 + 置信度 + bbox)以及按 BIO 聚合的 HEADER/QUESTION/ANSWER 实体列表。
uvicorn inference:app --host 0.0.0.0 --port 8102POST /predict:Content-Type: application/json,body 为上述 {"tokens": [...], "boxes": [...]}。POST /predict/upload:multipart 上传 .json 文件(字段名 file)。http://127.0.0.1:8102/docscurl -s -X POST http://127.0.0.1:8102/predict -H "Content-Type: application/json" -d @demo.json{
"model": "philschmid/layoutlm-funsd",
"device": "npu:0",
"task": "form_entity_recognition",
"num_tokens": 11,
"labels": {"0": "O", "1": "B-HEADER", "2": "I-HEADER", "3": "B-QUESTION", "4": "I-QUESTION", "5": "B-ANSWER", "6": "I-ANSWER"},
"tokens": [
{"token": "[CLS]", "label": "O", "score": 1.0, "bbox": [0, 0, 0, 0]},
{"token": "invoice", "label": "B-HEADER", "score": 0.998, "bbox": [400, 40, 620, 100]}
],
"entities": [
{"type": "HEADER", "text": "INVOICE", "score": 0.998, "token_indices": [1]},
{"type": "QUESTION", "text": "No.", "score": 0.993, "token_indices": [2]}
]
}layoutlm-funsd-npu/
├── inference.py # CLI + FastAPI 双模式入口
├── demo.json # 模拟发票表单 demo 输入
├── requirements.txt
├── .gitignore # 排除 model_weights/ logs/ __pycache__/
├── assets/ # 截图
└── model_weights/ # 权重缓存(不入库)