haoqiwang / vim Goto Github PK

I am not able to reproduce the imagenet validation set. The labels in your list do not match the predictions of my model (model = torch.hub.load('pytorch/vision:v0.10.0', 'wide_resnet50_2', pretrained=True)) at all:
https://github.com/haoqiwang/vim/blob/master/datalists/imagenet2012_val_list.txt

Hello dear author, thank you very much for your work, very efficient. I applied your method on another pre-trained model, but I found that the effect was very poor. I know this is not your responsibility, but I still hope you can help me out.
My current situation is as follows, I still use a pre-trained Vit model, but texture, SVHN, LSUN and places365 all have AUROC close to 1 with cifar100 as ID, but only about 0.91 on cifar10 as OOD detection dataset . We ablated the Residual module in your method and found that this was mainly caused by the very large change in this score.
Since our current method does not involve bias b, I directly set b = np.ones(class_num). In addition, I also have a scal parameter that acts on logit. We use this parameter while calculating energy, and energy is normal, but in the Residual part we found It is better not to use this parameter. My question is: Is the current poor performance due to problems with b and scal parameters, or is it caused by different pre-trained models?
Looking forward to your reply, thank you very much

I‘m very interested in your paper. I would like to ask whether your method can be applied to the field of natural language processing to detect out-of-distribution sentences. The sentence features and the picture features are different. Look forward to your reply.