How to run SuperSimpleNet in unsupervised manner? #2776
-
|
Hello, the SuperSimpleNet SOTA model was added in the latest version. Thank you for your team effort :). I have datasets such as normal and abnormal images without annotation boxes or segmentations. When I'm running SuperSimpleNet on my dataset the error given was an Attribute error on "the mask function" which is desired for segmentation handling. I have found the source repository GitHub of SuperSimpleNet. There are datasets such as KSDD2 similar to my dataset, such as not containing segmentation annotations and structured in normal and abnormal folders. Therefore, how to run SuperSimpleNet in this unsupervised manner without annotation? Thank you. |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments
-
|
Hi, for the unsupervised case you should only use normal samples in the training set. This is the only setting currently supported within the anomalib. If you want to also use anomalous samples, that would be a supervised setting that is only supported by the official repo. But mind that it's fully supervised setting, where segmentations are available (KSDD2 is one such case). In case that you do have anomalous images and don't have segmentations, you at least need image-level labels, that is then weakly supervised setting. This is currently supported by neither the anomalib or official repo. |
Beta Was this translation helpful? Give feedback.
-
|
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. |
Beta Was this translation helpful? Give feedback.
-
|
Note that we are working on enabling supervised setting in anomalib. In addition to supersimplenet, there are other new models that can do both unsupervised and supervised, so it would be good to have it in anomalib |
Beta Was this translation helpful? Give feedback.
Hi, for the unsupervised case you should only use normal samples in the training set. This is the only setting currently supported within the anomalib.
If you want to also use anomalous samples, that would be a supervised setting that is only supported by the official repo. But mind that it's fully supervised setting, where segmentations are available (KSDD2 is one such case).
In case that you do have anomalous images and don't have segmentations, you at least need image-level labels, that is then weakly supervised setting. This is currently supported by neither the anomalib or official repo.
The code however is ready for weak supervision, but it's not public yet. I think I'll release it …