funz-0 / promptad Goto Github PK

Official implementation of CVPR 2024 PromptAD: Learning Prompts with Only Normal Samples for Few-Shot Anomaly Detection

License: The Unlicense

Python 99.74% Shell 0.26%

anomaly-detection cvpr2024 few-shot-learning

promptad's Introduction

PromptAD Few-Shot Anomaly Detection

Official implementation of PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection (CVPR2024)

Install

conda create -n prompt_ad python==3.10
conda activate prompt_ad
bash install.sh

Data

Download the dataset from MvTec.

Download the dataset from VisA.

VisA preprocessing

Modify the source and target paths for the VisA dataset in ./dataset/prepare_visa_public.py

python ./dataset/prepare_visa_public.py

Modify the source paths for MvTec and VisA in ./dataset/mvtec.py and ./dataset/visa.py

Run

python run_cls.py  # image-level
python run_seg.py  # pixel-level

Citation

Please cite the following paper if this work helps your project:

@article{li2024promptad,
  title={PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection},
  author={Li, Xiaofan and Zhang, Zhizhong and Tan, Xin and Chen, Chengwei and Qu, Yanyun and Xie, Yuan and Ma, Lizhuang},
  journal={arXiv preprint arXiv:2404.05231},
  year={2024}
}

Acknowledge

We thank the great works WinCLIP and CoOp for assisting with our work.

promptad's People

Contributors

Stargazers

Watchers

Forkers

yoojlee yqlzbqc

promptad's Issues

can you provide a download link for this CVPR2024 paper?

CVPR2024 paper "PromptAD: Learning Prompts with Only Normal Samples for Few-Shot Anomaly Detection".

Screw class seems fail to reach the reported detetcion AUROC

There is no denying that promptAD is a very impressive work and thanks for the open resoursing for the knowledge broadcast in the field of AD.
But for those who is familiar with the mvtec dataset, there is no dout that the screw class is the most challenging one because of the none-allignment(basicly the rotation). In your paper , you report that in one-shot the auroc is blow 70,and then in two-shot setting the metric reach 98,and in four shot the metric just rocket to 100? Bro, u got be kidding me, since u even miswrite the carpet class with Garpet in your supplymentary table. I don't want to be mean, but you know, the data i see is just unrealist to me. And I run your code, and in each shot ,the detection auroc of screw is around 70.
And I think maybe the blame is not on you, so I check your baseline code the repo(WinClip) that be reimplemented by Mr.caoyunkang. In Cao's reimplentation, WinCLIP cannaot reach the reported detection auroc of amazon reported in class screw too，with a large margin of 20. Which makes me concused aboat which data should i really quote, the ones written in the paper, or the one i saw with my own eyes.

run_seg.py shots有什么含义？

从代码看就是单纯的控制循环次数，每个类别在不同的循环里跑出来的结果都是一样的，所以设计成[1,2,4]有什么特殊含义吗？它并没有按照论文所说的那样起作用
1-shot
Object:carpet =========================== Pixel-AUROC:99.53
Object:grid =========================== Pixel-AUROC:98.14
Object:leather =========================== Pixel-AUROC:99.39
Object:tile =========================== Pixel-AUROC:96.05
Object:wood =========================== Pixel-AUROC:95.82
Object:bottle =========================== Pixel-AUROC:97.31
Object:cable =========================== Pixel-AUROC:94.3
Object:capsule =========================== Pixel-AUROC:95.35
Object:hazelnut =========================== Pixel-AUROC:98.28
Object:metal_nut =========================== Pixel-AUROC:91.6
Object:pill =========================== Pixel-AUROC:93.61
Object:screw =========================== Pixel-AUROC:93.63
Object:toothbrush =========================== Pixel-AUROC:98.71
Object:transistor =========================== Pixel-AUROC:85.68
Object:zipper =========================== Pixel-AUROC:94.05

2-shot
Object:carpet =========================== Pixel-AUROC:99.53
Object:grid =========================== Pixel-AUROC:98.14
Object:leather =========================== Pixel-AUROC:99.39
Object:tile =========================== Pixel-AUROC:96.05
Object:wood =========================== Pixel-AUROC:95.82
Object:bottle =========================== Pixel-AUROC:97.31
Object:cable =========================== Pixel-AUROC:94.3
Object:capsule =========================== Pixel-AUROC:95.35
Object:hazelnut =========================== Pixel-AUROC:98.28
Object:metal_nut =========================== Pixel-AUROC:91.6
Object:pill =========================== Pixel-AUROC:93.61
Object:screw =========================== Pixel-AUROC:93.63
Object:toothbrush =========================== Pixel-AUROC:98.71
Object:transistor =========================== Pixel-AUROC:85.68
Object:zipper =========================== Pixel-AUROC:94.05

4-shot
Object:carpet =========================== Pixel-AUROC:99.53
Object:grid =========================== Pixel-AUROC:98.14
Object:leather =========================== Pixel-AUROC:99.39
Object:tile =========================== Pixel-AUROC:96.05
Object:wood =========================== Pixel-AUROC:95.82
Object:bottle =========================== Pixel-AUROC:97.31
Object:cable =========================== Pixel-AUROC:94.3
Object:capsule =========================== Pixel-AUROC:95.35
Object:hazelnut =========================== Pixel-AUROC:98.28
Object:metal_nut =========================== Pixel-AUROC:91.6
Object:pill =========================== Pixel-AUROC:93.61
Object:screw =========================== Pixel-AUROC:93.63
Object:toothbrush =========================== Pixel-AUROC:98.71
Object:transistor =========================== Pixel-AUROC:85.68
Object:zipper =========================== Pixel-AUROC:94.05

另外，以默认参数运行run_seg.py，得到的pixel-level的AUROC达不到论文精度，相差较远，是否有什么特别的设定需要更改

Code

Dear authors, thank you for your excellent work, and when will the code be released?

The results of few-shot anomaly detection will be different on different devices.

test on one new image

Thank you for your great work!
I'm very interested in trying to test this model on some new "industrial" images but I'm wondering how we can test the model on one (or more) new image?

code

Dear authors, thank you for your excellent work, and when will the code be released?

Pre-trained Weights.

Your work is awesome. Do you have pre-trained weights that can be used to reproduce the reported result on paper?

experiments requirements

What hardware configuration is needed to run your work? How much memory is needed for the graphics card? Thank you

datasets/seeds_mvtec/bottle/selected_samples_per_run.txt

selected_samples_per_run.txt这个文件中的内容
#1: 131
#2: 101 048
#4: 029 045 192 127
代表什么含义，我如何将数据集修改为我自己的数据

Can not reproduce the results

Hi, thanks for your research effort!

Actually, there is no way i can reproduce the results by really far margin. Can you please put the selected hyperparameters that are used for reporting results?

Thanks
Regards.

如何在自建数据集训练并测试？

首先感谢开源你的项目代码。
我想问问我该如何在自建数据集运行代码呢？
我在尝试后报错如下

May I know if the epoch is 1 or not?

Execution error with run_cls.py

I executed the run_cls.py file according to the readme and encountered the following error (my env: win10 with rtx3060):
Traceback (most recent call last): File "C:\Users\35314\Desktop\PromptAD\PromptAD\CLIPAD\model.py", line 236, in encode_text_embeddings x = x + self.positional_embedding.to(cast_dtype) RuntimeError: The size of tensor a (105) must match the size of tensor b (77) at non-singleton dimension 1

可能是显卡不支持？

script run_cls.py has error

Hi,
Thanks for sharing your code.
I get an issue when I run the run_cls.py script. Here is the error:

  File "/home/alexandre/projects/PromptAD/train_cls.py", line 257, in <module>
    main(args)
  File "/home/alexandre/projects/PromptAD/train_cls.py", line 189, in main
    metrics = fit(model, args, test_dataloader, device, check_path=check_path, train_data=train_dataloader)
  File "/home/alexandre/projects/PromptAD/train_cls.py", line 76, in fit
    abnormal_text_features_learned = model.encode_text_embedding(abnormal_text_prompt_learned, model.tokenized_abnormal_prompts_learned)
  File "/home/alexandre/projects/PromptAD/PromptAD/model.py", line 252, in encode_text_embedding
    text_features = self.model.encode_text_embeddings(text_embedding, original_tokens)
  File "/home/alexandre/projects/PromptAD/PromptAD/CLIPAD/model.py", line 236, in encode_text_embeddings
    x = x + self.positional_embedding.to(cast_dtype)
RuntimeError: The size of tensor a (105) must match the size of tensor b (77) at non-singleton dimension 1

I followed the procedure that is proposed in the readme.
How can I solve the problem ?
thanks in advance

ps: run_seg.py seems working like a charm

模型能否导出来onnx格式？

我尝试在 test_seg.py中断点，将模型导出为onnx，发现有以下报错：

torch.onnx.export(model, data, "test.onnx") // 这是我临时的导出代码，以下为执行结果消息
Traceback (most recent call last):
File "", line 1, in
File "f:\pyenv\Torch118gpu\lib\site-packages\torch\onnx\utils.py", line 516, in export
_export(
File "f:\pyenv\Torch118gpu\lib\site-packages\torch\onnx\utils.py", line 1612, in _export
graph, params_dict, torch_out = _model_to_graph(
File "f:\pyenv\Torch118gpu\lib\site-packages\torch\onnx\utils.py", line 1134, in _model_to_graph
graph, params, torch_out, module = _create_jit_graph(model, args)
File "f:\pyenv\Torch118gpu\lib\site-packages\torch\onnx\utils.py", line 1010, in _create_jit_graph
graph, torch_out = _trace_and_get_graph_from_model(model, args)
File "f:\pyenv\Torch118gpu\lib\site-packages\torch\onnx\utils.py", line 926, in _trace_and_get_graph_from_model
raise RuntimeError(
RuntimeError: state_dict changed after running the tracer; something weird is happening in your model!

请问是什么原因呀？如何解决呢？

感谢你们的开源和回复！