search

marco-rudolph / cs-flow

This is the official repository to the WACV 2022 paper "Fully Convolutional Cross-Scale-Flows for Image-based Defect Detection" by Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn and Bastian Wandt.
111 stars 30 forks source link

Try other classification networks as feature extractors #20

Closed CODEC-ZLY closed 1 year ago

CODEC-ZLY commented 2 years ago

Dear Marco-Rudolph, we are currently working on your research. We explored that when other classification networks were used as feature extractors, the losses sent into CS-flow training became extremely large and could not be well converged. Even when it is used in EfficientNet_V2 - m, also cannot achieve EfficientNet_b5 as feature extractor, network achieved better effect. This question has puzzled me for a long time, and I want to know whether you have made a special optimization for EfficientNet_b5. Theoretically, switching to another pre-trained classification network can also achieve a good effect. If you are busy reading this information,I am particularly looking forward to your reply to my questions, which will benefit me a lot. Thank you so much！！！

marco-rudolph commented 2 years ago

Hello, there was not made a special optimization for B5. I use the 'original' pretrained weights from ImageNet provided by pytorch. It's also not clear to me in part why there are such big differences between the networks. However, I could not observe this effect of a very large loss when using 'use_gamma=True'.