Marco Rudolph

Thank you! Evaluation code can be found in "evaluate.py".

As the loss reflects the negative log likelihood, it can be negative as the log likelihood can be positive.

Thank you :) The NF model does not return the log-likelihood, it is just the output z with same dimension as the input. Since the negative log likelihood of a...

I cannot reproduce your error. Did you change config.py or anything else?

Hello, thank you. In principle, this can work, since the feature extractors pre-trained on ImageNet should have seen similar structures. The performance depends on how big the differences in shape...

There is no way for this method to use labeled anomalies in training. I am not exactly sure how this is meant. So there are different object categories within the...

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...

Hi, unfortunately, there is currently no option to configure the number of scales. n_scales is used for many parts of the code for readability, but not for the NF architecture...

I tried this, but got a slightly worse performance for MVTec. It might be beneficial for other datasets.

We use pretrained weights from ImageNet and do not train them further. These are suitable for detecting anomalies. It does not really make sense to keep training these weights, because...