xcyao00 / fod Goto Github PK

Hi, thanks for your great work!
It seems that you directly use the test set as the validation set in code:

您好，我跑您的代码精度没有论文里高，大概是这样的

Mean Image-AUC: 0.974334492926727
Mean Pixel-AUC: 0.9817032875158126

想请教下是什么原因呢，因为 backbone 不同吗？

Hello, after I replaced the dataset, there were abnormal results. For example, the Pixel AUC increased from 0.72 to 0.8, but the Image AUC value decreased from 0.9 to 0.8. Can you help me answer my questions? thanks

我只有6G的显存，我使用resnet18替换wd_resnet50，并把batch_size调为4，输入是256，但是内存占用仍然过高，有什么办法可以训练吗

Hello, I would like to ask why I don't find the decoder section inside the mod of the code, is there no decoder? And I see that the paper is written with self-supervision inside, may I ask if this paper is based on reconstructed transformer or non-reconstructed transformer? Looking forward to your answer, thank you!

I found a bug in the project when I wanted to run the project in cuda:2. “RuntimeError ：Expected all tensors to be on the same device, but found at least two devices, cuda:2 and cuda:0!” occurred. On closer inspection, I found that the problem is at pos_embed.cuda() of PositionEmbedding.

您好，我在阅读您的论文中，对于这句话没法准确翻译：In this paper, we propose a novel AD framework: FOcusthe-Discrepancy (FOD), which can simultaneously spot the patch-wise, intra- and inter-discrepancies of anomalies.尤其是patch-wise这个名词，我想了很久，也没明白什么意思，您能解释下它的正确含义吗？非常感谢