search

sungikchoi / NVB

Code of Novelty Detection Via Blurring
MIT License
0 stars 2 forks source link

Question about blurred image sets #1

Open sesmae opened 3 years ago

sesmae commented 3 years ago

Thanks for sharing the code for your amazing paper. I went over baseline_main_nobar.py . I have a question about it. it seems that in each call to baseline_main_nobar.py you are initializing three networks. f, g0 and g1. does this mean that b_train = 1 in all of the experiments in the script? Because based on your paper there need to be multiple randomly initialized target networks for each set of blurred images.But it seems that in each experiment only one set of blurred CIFAR20 images are generated by zeroing the bottom Ki non-zero singular values for each image. Am I missing or misunderstood something here?

sungikchoi commented 3 years ago

Yup. In the CIFAR-10 dataset, we fixed the number of the blurred datasets (b_train) to 1. However, in LSUN or TinyImageNet dataset, or replication of experiment where we used uniform log effective rank criteria, the number of the blurred datasets becomes larger, which means you need more randomly initialized g's.

sesmae commented 3 years ago

Yup. In the CIFAR-10 dataset, we fixed the number of the blurred datasets (b_train) to 1. However, in LSUN or TinyImageNet dataset, or replication of experiment where we used uniform log effective rank criteria, the number of the blurred datasets becomes larger, which means you need more randomly initialized g's.

Great! Thank you!