RapidOCR - NPU 深度测评与优化报告

pip install modelscope

from modelscope import snapshot_download
model_dir = snapshot_download('RapidAI/RapidOCR')

git clone https://www.modelscope.cn/RapidAI/RapidOCR.git

import torch
import torch_npu
import onnx
from onnx import helper, shape_inference
from onnx2torch import convert

# 1. 修复 ONNX (Constant→initializer)
model = onnx.load("onnx/PP-OCRv4/det/ch_PP-OCRv4_det_mobile.onnx")
inferred = shape_inference.infer_shapes(model)
# ... (详见 scripts/convert_to_npu.py)

# 2. 转换为 PyTorch 并移至 NPU
torch_model = convert(fixed_path).npu().eval()

# 3. 保存 TorchScript 模型 (下次无需重新转换)
traced = torch.jit.trace(torch_model.half(), dummy_input)
traced.save("scripts/traced_models/det.pt")

图片                    预测文字               Ground Truth        结果
──────────────────────────────────────────────────────────────────
SALUTES  → "SALUTES"      (GT: SALUTES)     ✅ 完全正确
KINGFISHER  → "KINGFISHER" (GT: KINGFISHER)  ✅ 完全正确
HOLLYWOOD → "HOLLYWOOD."  (GT: HOLLYWOOD)    ⚠️ 多了一个点号
GYPSY     → "Gypsy"       (GT: GYPSY)        ⚠️ 大小写不一致
MTV       → "MTV"         (GT: MTV)          ✅ 完全正确

📷 图片: ch_en_num.jpg (41.5KB)
  尺寸: 323x430
  检测到 6 个文本区域 (det: 18.0ms)
    [0] "Hello"  (rec: 9.2ms)
    [1] "World"  (rec: 8.5ms)
    [2] "12345"  (rec: 8.1ms)

{
    "env": { "torch": "2.9.0", "npu": "Ascend910_9362", ... },
    "models": { "cls": { "params": 130000, ... }, ... },
    "accuracy": { ... },
    "perf": { ... },
    "pipeline": { "det_rec_2_crops_ms": 52.797, "fps": 18.9 }
}

RapidOCR/
├── onnx/
│   ├── PP-OCRv4/
│   │   ├── cls/    # 文本方向分类
│   │   ├── det/    # 文本检测 (DBNet)
│   │   └── rec/    # 文本识别 (CRNN)
│   └── PP-OCRv5/   # 新版模型
├── scripts/
│   ├── convert_to_npu.py    # ONNX→PyTorch→NPU 转换脚本
│   ├── quantize_int8.py     # INT8 量化脚本
│   └── traced_models/       # TorchScript 编译后的模型
│       ├── cls.pt
│       ├── det.pt
│       └── rec.pt
├── inference.py             # 推理引擎
├── VALIDATION_RESULT.json   # 验证报告
└── VALIDATION_SUMMARY.md    # 验证摘要

pip install torch torch_npu onnx onnxruntime onnx2torch

python scripts/convert_to_npu.py --model all

import torch, torch_npu

det = torch.jit.load("scripts/traced_models/det.pt")
img = torch.randn(1, 3, 640, 640, dtype=torch.float16).npu()
with torch.no_grad():
    output = det(img)  # (1, 1, 640, 640)