PatchCore: Towards Total Recall in Industrial Anomaly Detection — NPU 适配版

export PIP_INDEX_URL=https://repo.huaweicloud.com/repository/pypi/simple/
pip install pretrainedmodels faiss-cpu timm torch_npu

# NPU 运行时优化（自动设置）
export TASK_QUEUE_ENABLE=2
export CPU_AFFINITY_CONF=1
export LD_PRELOAD="${LD_PRELOAD:-}/opt/atomgit/tcmalloc-install/lib/libtcmalloc.so"

# 一键验证：环境预检 + 精度对齐 + 延迟基准 + Batch 扩展 + 多卡并行 + L2 Profiling
chmod +x quick_verify.sh
./quick_verify.sh

# → 输出详细日志到 ./results/verify_YYYYMMDD_HHMMSS.log
# → 自动判断 PASS/WARN，提取关键指标

export PYTHONPATH=src

# Step A: CPU vs NPU 精度对齐 + L1 Profiling
python3 profile_inference.py

# Step B: Batch 推理扩展性摸高 (bs=1~16)
python3 batch_benchmark.py

# Step C: 2 卡并行吞吐测试
python3 multi_npu_benchmark.py

# Step D: L2 Profiling（全栈调用栈 + 内存）
python3 profile_l2.py

source /usr/local/Ascend/ascend-toolkit/set_env.sh
export ASCEND_RT_VISIBLE_DEVICES=0
export PYTHONPATH=src

# 训练（以 bottle 类别为例）
python bin/run_patchcore.py --gpu 0 --seed 0 --save_patchcore_model \
  --log_group IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0 \
  patch_core -b wideresnet50 -le layer2 -le layer3 \
  --pretrain_embed_dimension 1024 --target_embed_dimension 1024 \
  --anomaly_scorer_num_nn 1 --patchsize 3 \
  sampler -p 0.1 approx_greedy_coreset \
  dataset --resize 256 --imagesize 224 -d bottle /path_to/mvtec

# 加载已有模型推理
python bin/load_and_evaluate_patchcore.py --gpu 0 --seed 0 $savefolder \
  patch_core_loader -p /path/to/model \
  dataset --resize 366 --imagesize 320 -d bottle /path_to/mvtec

# NPU 推理验证 + CPU 对齐 + L1 Profiling
python profile_inference.py

import torch
import patchcore.backbones
import patchcore.common
import patchcore.patchcore

# 设备选择
device = torch.device('npu:0')

# 构建 backbone
backbone = patchcore.backbones.load('wideresnet50')
backbone.name = 'wideresnet50'

# 构建 PatchCore
nn_method = patchcore.common.FaissNN(on_gpu=False, num_workers=4)
pc = patchcore.patchcore.PatchCore(device)
pc.load(
    backbone=backbone,
    layers_to_extract_from=['layer2', 'layer3'],
    device=device,
    input_shape=[3, 224, 224],
    pretrain_embed_dimension=1024,
    target_embed_dimension=1024,
    patchsize=3,
    featuresampler=patchcore.sampler.IdentitySampler(),
    anomaly_scorer_num_nn=1,
    nn_method=nn_method,
)

# 图像编码与异常检测
img = torch.randn(1, 3, 224, 224).to(device)
with torch.no_grad():
    features = pc._embed(img)
# features 为 patch-level embedding，供后续最近邻搜索与异常评分

┌────────────────────────────────────────────┐
│  Backbone Forward (WideResNet50): ~30 ms   │
│  Patchify + Preprocessing:        ~4 ms    │
│  Feature Aggregation:             ~3 ms    │
│  ───────────────────────────────────────   │
│  总延迟 (embedding):              ~37 ms   │
│  对应吞吐量:                      ~26.8 FPS│
└────────────────────────────────────────────┘

$ python profile_inference.py
============================================================
PatchCore NPU Inference Validation & L1 Profiling
============================================================

NPU: npu:0, CPU: cpu

[1] Building PatchCore models...
  NPU...
  CPU...

[2] Building memory bank on NPU...

[3] Embedding comparison (NPU vs CPU)...
  Cosine similarity: 0.854682
  Relative L2 error: 2.305088e-04

[4] Latency & Throughput...
  NPU:
    Latency: 37.26 ± 0.21 ms, Throughput: 26.84 fps
  CPU:
    Latency: 360.55 ± 2.39 ms, Throughput: 2.77 fps

  NPU vs CPU speedup: 9.68x

[5] L1 Profiling on NPU...
  L1 profiling -> ./npu-profiling-L1
  Profiling done!

============================================================
SUMMARY
  Cosine similarity (NPU vs CPU): 0.854682
  NPU latency: 37.26 ms, CPU latency: 360.55 ms
  Speedup: 9.68x
  L1 profiling: SUCCESS
============================================================

patchcore-inspection/
├── bin/
│   ├── run_patchcore.py              # 训练 + 评估入口
│   └── load_and_evaluate_patchcore.py # 加载模型并评估
├── src/
│   └── patchcore/
│       ├── backbones.py              # Backbone 加载 (torchvision/timm)
│       ├── common.py                 # FaissNN, NetworkFeatureAggregator, Preprocessing
│       ├── patchcore.py              # PatchCore 核心类
│       ├── sampler.py                # Coreset 采样器
│       ├── utils.py                  # 工具函数 (device, seed, plot)
│       ├── metrics.py                # AUROC/PRO 计算
│       └── datasets/
│           └── mvtec.py              # MVTec-AD 数据集加载
├── raw_weights/
│   └── wide_resnet50_2-95faca4d.pth  # ImageNet1K_V1 预训练权重 (git-lfs)
├── profile_inference.py              # NPU 推理验证 + L1 Profiling 脚本
├── alignment_results.json            # CPU vs NPU 精度/性能对齐结果
├── README.md                         # 本文档
└── .gitattributes                    # git-lfs 配置

Torchvision WideResNet50 (ImageNet1K_V1 预训练权重)
         │
         ├──→ [替换] 硬编码 .cuda() → .to(device) 动态绑定
         │
         ├──→ [替换] torch.cuda.device → torch.npu.device
         │
         ├──→ [替换] torch.cuda.empty_cache → torch.npu.empty_cache
         │
         ▼
   torch_npu.transfer_to_npu 设备注入
         │
         ├──→ [适配] FaissNN (faiss-cpu, on_gpu=False)
         │
         ├──→ [适配] set_torch_device() 返回 npu:X
         │
         ▼
   昇腾 NPU 在线推理 (embedding 37.26ms, 精度 rel_diff=2.30e-04)

@inproceedings{roth2021patchcore,
  title={Towards total recall in industrial anomaly detection},
  author={Roth, Karsten and Pemula, Latha and Zepeda, Joaquin and Sch{\"o}lkopf, Bernhard and Brox, Thomas and Gehler, Peter},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}