megvii-research / fssd_ood_detection Goto Github PK

Hi. I am wondering how to get the initial concentration as it is shown in Figure 3 in the paper. Could the default pytorch initialization method for ResNet 34 get the same result?

I am trying to reproduce the concentration result using ResNet-34 for CIFAR-10. While the penultimate feature of OoD data (uniformly generated data ranging form [0,1]) seems to concentrate to a fix point, the feature representation of CIFAR-10 images got from penultimate layer seems quite sparse, as you can see in the following figure where 10 stands for OoD data.

Could you kindly provide the code generating Figure 3 in the paper? or please enlighten me if there is anything i am missing.

Hi. I am trying to reproduce and use your dogs vs non-dogs dataset. I followed the link to download it but the archive file seems to be broken.
I have read this issue, where you explain how you created the dataset.
in it, you declare using 50K images :

• Dogs-50A-train: 50K images 1000 images from ImageNet Training Set for each 50 classes (I guess, because you wrote 100). Another problem is that for certain classes, you only have ~700 images.
• Dogs-50A-val: 10K images (so I guess 200 images for each 50 classes). You write once that you take it from ImageNet validation (in the table) and then from the training set. The validation set only contain 50 images per class. I thus think that you might have taken these images from the training set. Am I right?

Would you have a link to a non-broken archive ? Or could you please clear my above concerns such that I can generate an equivalent one ?

Hi I tried to click on the link in pretrained link in google cloud, but it returns 404 now
老哥，来个百度网盘也行

Thanks for your excellent work.

I'm trying to reproduce the experiment result of ImageNet (dogs) vs ImageNet (non-dogs). Could you please describe how do you construct the dataset in detail or post related code?

Hi,

The paper presents results with the Bacteria Genome dataset using LSTM networks, but the repository does not have the implementation to reproduce the results. Could you provide the codes used for this task? I tried to implement it but without success, AUROC is at 50% in all algorithms