search

xrli-U / MuSc

This is an official PyTorch implementation for "MuSc : Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images" (MuSc ICLR2024).
MIT License
204 stars 12 forks source link

feature extractors #10

Open 7HHHHH opened 1 month ago

7HHHHH commented 1 month ago

Very good work! Do you support other feature extractors, or can you only choose from these three?

xrli-U commented 1 month ago

These three are the most classic ViT pre-training models. You can replace it with any pre-trained vision transformer such as MAE or SimCLR. It can also be replaced with swin transformer, but the AC and AS results are not good, please refer to Table 8 in our paper.

giraffeGYH commented 1 day ago

该方法能够使用CNN网络进行特征提取？比如resnet50等，效果如何？

xrli-U commented 1 day ago

该方法能够使用CNN网络进行特征提取？比如resnet50等，效果如何？

可以将特征提取器换成CNN，相比于ViT指标会有所下降。通常来讲特征提取器提取的特征越好，最终的分类和分割的指标会越高

giraffeGYH commented 21 hours ago

我尝试将特征提取的骨干网络修改为ResNet50，可是出现了错误，请问如何修改呢？（直接替换clip？）

xrli-U commented 20 hours ago

我尝试将特征提取的骨干网络修改为ResNet50，可是出现了错误，请问如何修改呢？（直接替换clip？） image

当前程序并没有编写兼容resent的代码，请使用torchvision.models.resnet50自定义resnet网络

xrli-U commented 20 hours ago

我尝试将特征提取的骨干网络修改为ResNet50，可是出现了错误，请问如何修改呢？（直接替换clip？） image

或者可以等待近两个月，我们会更新新版本的musc，其中会加入对更多feature extractor的兼容

giraffeGYH commented 19 hours ago

感谢您的回复。我想复现full-shto的实验结果，请问musc的full-shot条件如何设置？

xrli-U commented 13 hours ago

感谢您的回复。我想复现full-shto的实验结果，请问musc的full-shot条件如何设置？

官方代码会在近两个月进行更新，如果您想要自己修改的话，feature extract和LNAMD模块保持不变，mutual scoring模块中需要将打分的对象进行替换

giraffeGYH commented 10 hours ago

好的，谢谢您。