cutpaste是一种简单有效的自监督学习方法，其目标是构建一个高性能的两阶段缺陷检测模型，在没有异常数据的情况下检测图像的未知异常模式。首先通过cutpaste数据增强方法学习自监督深度表示，然后在学习的表示上构建生成的单类分类器，从而实现自监督的异常检测。
论文: CutPaste: Self-Supervised Learning for Anomaly Detection and Localization
参考repo: pytorch-cutpaste
在此非常感谢Runinho等人贡献的pytorch-cutpaste ，提高了本repo复现论文的效率。
aistudio体验教程: CutPaste_paddle
| defect_type | CutPaste (3-way) | Runinho. CutPaste (3-way) | Li et al. CutPaste (3-way) |
|---|---|---|---|
| bottle | 90.7 | 99.6 | 98.3 |
| cable | 91.7 | 77.2 | 80.6 |
| capsule | 93.2 | 92.4 | 96.2 |
| carpet | 97.3 | 60.1 | 93.1 |
| grid | 100.0 | 100.0 | 99.9 |
| hazelnut | 99.0 | 86.8 | 97.3 |
| leather | 100.0 | 100.0 | 100.0 |
| metal_nut | 98.6 | 87.8 | 99.3 |
| pill | 97.3 | 91.7 | 92.4 |
| screw | 88.4 | 86.8 | 86.3 |
| tile | 98.8 | 97.2 | 93.4 |
| toothbrush | 98.6 | 94.7 | 98.3 |
| transistor | 98.7 | 93.0 | 95.5 |
| wood | 99.3 | 99.4 | 98.6 |
| zipper | 100.0 | 98.8 | 99.4 |
| average | 96.7 | 91.0 | 95.2 |
首先介绍下支持的硬件和框架版本等环境的要求：
- 硬件：GPU显存建议在6G以上
- 框架：
- PaddlePaddle >= 2.2.0
- 环境配置：直接使用
pip install -r requirements.txt安装依赖即可。
- 全量数据训练：
- 下载好 metec-ad 数据集
- 将其解压到 Data 文件夹下
- 少量数据训练：
- 无需下载数据集，使用lite_data里的数据即可
- 默认不使用resnet18预训练模型进行训练，如想使用,需要预先下载预训练权重 (提取码：l7c3)至项目根目录下并传入参数：
python train.py --pretrained True
- 全量数据训练：
- 下载好 metec-ad 数据集后，将其解压到 ./Data 文件夹下
- 运行指令
python tools/train.py --epochs 7000 --batch_size 32 --workers 4 --log_interval 10 --pretrained True
- 少量数据训练：
- 运行指令
python tools/train.py --data_dir lite_data --type lite --epochs 5 --batch_size 4
- 运行指令
- 部分训练日志如下所示：
> python tools/train.py --data_dir lite_data --type lite --epochs 5 --batch_size 4 --cuda False
Namespace(batch_size=4, cuda='False', data_dir='lite_data', epochs=5, freeze_resnet=20, head_layer=1, lr=0.03, model_dir='logs', optim='sgd', pretrained=False, save_interval=500, test_epochs=-1, type='l
ite', variant='3way', workers=0)
using device: cpu
training bottle
loading images
loaded 209 images
epoch:1/5 loss:1.2578 avg_reader_cost:0.05 avg_batch_cost:3.01 avg_ips:0.75
epoch:2/5 loss:1.6850 avg_reader_cost:0.02 avg_batch_cost:2.81 avg_ips:0.70
epoch:3/5 loss:1.5016 avg_reader_cost:0.02 avg_batch_cost:2.75 avg_ips:0.69
...
- 全量数据模型评估：
python eval.py - 少量数据模型评估：
python tools/eval.py --data_dir lite_data --type lite
> python tools/eval.py --data_dir lite_data --type lite --cuda False
Namespace(cuda='False', data_dir='lite_data', density='sklearn', head_layer=1, model_dir='logs', save_plots=True, type='lite')
evaluating bottle
loading model logs/bottle/final.pdparams
loading images
loaded 8 images
using density estimation GaussianDensitySklearn
bottle AUC: 0.875
average auroc:0.8750
- 基于原始代码的模型预测：
python tools/predict.py --data_type bottle --img-path images/demo0.png --dist_th 1.0
- 基于推理引擎的模型预测：
python deploy/export_model.py
python deploy/infer.py --data_type bottle --img-path images/demo0.png --dist_th 1.0
部分结果如下：
> python deploy/export_model.py
inference model has been saved into deploy
> python deploy/infer.py --data_type bottle --img-path images/demo0.png --dist_th 1.0
image_name: images/demo0.png, class_id: 0, prob: 0.07689752858017344
- tipc创建指南请见tipc创建及基本使用方法。
- 本项目TIPC脚本测试命令详见Linux GPU/CPU 基础训练推理测试
#测试环境准备脚本
bash test_tipc/prepare.sh test_tipc/configs/resnet18/train_infer_python.txt lite_train_lite_infer#测试训练验证推理一体化脚本
bash test_tipc/test_train_inference_python.sh test_tipc/configs/resnet18/train_infer_python.txt lite_train_lite_infer输出结果如下，表示命令运行成功。
Run successfully with command - python3.7 tools/train.py --type lite --model_dir logs --output=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0 --epochs=2 --batch_size=1!
......
Run successfully with command - python3.7 tools/eval.py --type lite --pretrained=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0/final.pdparams!
......
Run successfully with command - python3.7 deploy/export_model.py --pretrained=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0/final.pdparams --save-inference-dir=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0!
......
Run successfully with command - python3.7 deploy/infer.py --use-gpu=True --model-dir=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0 --batch-size=1 --benchmark=False > ./log/resnet18/lite_train_lite_infer/python_infer_gpu_batchsize_1.log 2>&1 !
......
Run successfully with command - python3.7 deploy/infer.py --use-gpu=False --model-dir=./log/resnet18/lite_train_lite_infer/norm_train_gpus_0 --batch-size=1 --benchmark=False > ./log/resnet18/lite_train_lite_infer/python_infer_cpu_batchsize_1.log 2>&1 ! 本项目的发布受Apache 2.0 license许可认证。
参考论文: CutPaste: Self-Supervised Learning for Anomaly Detection and Localization
参考repo: pytorch-cutpaste