Open lorenzofilizola opened 1 year ago
I have been studying the paper all these while. Teacher training on ImageNet and removal of dropout is absolutely right. This I have confirmed from the author. It may be due to fdfe, but u may try with patch size 17,33 or multi-scale also. I will also check the fdfe implementation. What performance did u get ?
Patch Size | AUPRO | AUROC |
---|---|---|
65 | 0.8203277491069684 | 0.9873015873015872 |
Patch Size | AUPRO | AUROC |
---|---|---|
65 | 0.588493822754899 | 0.8150299850074961 |
Patch Size | AUPRO | AUROC |
---|---|---|
65 | 0.6944213879887304 | 0.6445951336258476 |
33 | 0.60722074480315 | 0.6039090546469884 |
17 | 0.6617633433551405 | 0.5967291583566016 |
These are the results I got up to now. Before pre-training on ImageNet, results were much lower (mainly the AUPRO).
Can we discuss over mail. Give ur mail id or mail me at sushovanjena@gmail.com
Sure, I mailed you.
Hello, I am noticing that small anomalies are often imprecisely localized with square shapes, mainly using a patch size of 65.
Does anyone have an idea of what could cause this kind of problem? I suspect it could be a problem with FDFE, since it's pretty much the only part of the code I have not inspected myself.
In addition to the training implemented in the repo, I am doing a pre-training of the teacher network on ImageNet, as explained in the original paper. I also removed the dropout layers from the teacher/student architecture, since I wanted to reproduce the paper results and those are not mentioned in Bergmann's work.