RapidOCR Ascend 910 推理性能优化实战（Autoresearch 版）

git checkout -b exp/rapidocr_{optimization}_$(date +%s)
git add -A && git commit -m "exp(...): ...
- metric_primary: XX → XX
- metric_cer_diff: XX
- status: PASS/FAIL
"

{
  "experiment_id": "exp_001_rapidocr_env",
  "timestamp": "2026-05-14T15:00:00Z",
  "hypothesis": "TASK_QUEUE_ENABLE=2 + CPU_AFFINITY_CONF=2 可提升推理吞吐",
  "changes": [],
  "metrics": {
    "throughput_images_per_sec": 22.0,
    "latency_p50_ms": 44.0,
    "latency_p99_ms": 58.0,
    "cer_diff": 0.001,
    "memory_mb": 2800
  },
  "status": "PASS",
  "commit_hash": "a1b2c3d"
}

if current.throughput > best.throughput * 1.05 and current.cer_diff < 0.01:
    git_merge_to_main()
    best = current
else:
    git_checkout_main()

goal: "RapidOCR Ascend 910 推理性能达标且精度误差<1%"
termination:
  success: throughput >= baseline_target AND cer_diff < 0.01
  failure: time_budget_exceeded (5h) OR consecutive_rollbacks >= 5
metrics:
  primary: throughput
  secondary: latency_p50_ms, latency_p99_ms
  guardrail: cer_diff, memory_mb
loop:
  interval: "每个实验完成后立即评估"
  max_iterations: 50
actions:
  on_pass: git merge → update best → generate_next_hypothesis
  on_fail: git rollback → log_failure → try_alternative_path
  on_timeout: package_current_best → submit_deliverables

npu-smi info
python -c "import torch; import torch_npu; print(torch_npu.__version__)"

cd /workspace
git clone https://ai.gitcode.com/hf_mirrors/pitapo/rapidocr rapidocr-ascend
cd rapidocr-ascend
pip install -r requirements.txt  # torch, torchvision, onnxruntime, opencv-python
git init
git add -A && git commit -m "init: clone rapidocr upstream"

python eval/baseline_run.py --device npu --batch 1,4,8,16

git branch main-best

git checkout -b exp/env_optimization
git checkout main-best -- .
export TASK_QUEUE_ENABLE=2
export CPU_AFFINITY_CONF=2
python eval/perf_eval.py --device npu

git checkout -b exp/det_backbone_optimization
# ResNet/MobileNet backbone 的 BN + ReLU
# DB head 的卷积和 sigmoid
python eval/perf_eval.py

git checkout -b exp/rec_model_optimization
# CRNN: CNN backbone + BiLSTM + CTC
# SVTR/Transformer: Attention → npu_fusion_attention (bf16)
python eval/perf_eval.py --dtype bf16

git checkout -b exp/preprocess_optimization
# resize、normalize、padding 优化
# 尝试用 torchvision.transforms 或 NPU 预处理
python eval/perf_eval.py

git checkout -b exp/contiguous_layout
python eval/perf_eval.py

git checkout -b exp/om_conversion
git checkout main-best -- .
# 检测和识别模型分别导出
python om_model/export_det_onnx.py --model ch_PP-OCRv4_det
python om_model/export_rec_onnx.py --model ch_PP-OCRv4_rec
python eval/compare_onnx_torch.py

bash om_model/atc_convert_det.sh
bash om_model/atc_convert_rec.sh

python inference_acl.py --det_model om_model/det.om --rec_model om_model/rec.om --device npu
python eval/perf_eval.py --backend acl

import torch_npu
prof = torch_npu.profiler.profile(...)

python eval/final_compare.py

git tag v1.0-optimized
git checkout v1.0-optimized

mkdir -p deliverables
cp inference.py readme.md eval/ logs/ SKILL.md deliverables/

REPO_URL="https://oauth2:${ATOMGIT_USER_TOKEN}@gitcode.com/yourname/rapidocr-ascend.git"
git remote add origin "$REPO_URL"
git push -u origin main

if experiment.cer_diff > 0.01:
    rollback("精度超标")
elif experiment.throughput < best.throughput * 0.95:
    rollback("性能倒退 > 5%")
elif experiment.memory_mb > best.memory_mb * 1.3:
    rollback("显存暴涨 > 30%")
elif experiment.latency_p99_ms > best.latency_p99_ms * 1.2:
    rollback("尾部延迟恶化 > 20%")
else:
    merge_to_main_best()