• I fixed all models to run smoothly, however, I am unsure if models other than OCGNN require updating the hypersphere radius since I was focussing mainly on the OCGNN model. Theoretically, if any model depends on hypersphere learning it should require similar handling.

• For enabling "active_mask" to filter train/valid/test data, I have enabled it in OCGNN. Similarly, I enabled it in models that utilize the DeepDetector function, namely AdONE, AnomalyDAE, CONAD, DOMINANT, GUIDE, GAE, and DONE. I enabled it by filtering all vectors passed to the loss_func function with "active_mask". (Please confirm that logic)

• The other models need more study to their algorithms to enable "active_mask".

Hi Ahmed, thanks a lot for your contributions. However, I have some concerns.

• active_mask: as discussed in How do you use different masks in PyGOD datasets. #50, we are focusing on unsupervised settings so far. Thus, there is no need to mask any node in the training data, and we won't encourage users to do so.
• is_train: I am a little confused by "training only." For the transductive setting, the score is ready after training. For the inductive setting, we won't use warmup if predict is called.
• c: we are happy to take this change if it helps.

It's my pleasure to contribute, regarding your concerns.

• For active_mask: I understand some models do not require masking training nodes in unsupervised settings. However, since OCGNN is a one-class classifier model, it requires training on only one class of data, as mentioned in the original paper, and testing on both classes. To achieve that in a transductive setting I utilized masks.
• I used is_train to avoid updating the hypersphere radius or centre when predict is called. If we initialize warmup=math.inf, the condition self.warmup > 0 will still be true even when predict is called. I think it's ok if you see a better way to assert that without using is_train flag.

Thanks for your reply. That makes sense. I am still considering whether we should add this mask for the only one class model in our library. Actually, this one class setting is very strict. The training data cannot contain any outliers. It might be too ideal for real-world applications.

For the is_train flag, it can be more elegant by setting the warmup to the number of training steps, which can be done by passing a parameter. So no need to change this part.

I am just confirming, Is there any tasks you need from my side for this pull request?

I kept the part you marked and reversed the rest of the changes.

Thanks for reversing. But I still see several other files was changed, including the cache files.

Yes it was cache files and empty spaces, It should be clean now.

Pull Request Test Coverage Report for Build 7393446504

• 6 of 6 (100.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.001%) to 99.573%

Totals
Change from base Build 7283849275:	0.001%
Covered Lines:	3029
Relevant Lines:	3042

💛 - Coveralls

Thanks for your effort. PR Merged.