Open SissiW opened 4 months ago
Thank you very much for your interest in our work. You can obtain pre-trained models at https://github.com/facebookresearch/deit/blob/main/README_deit.md or directly download them from https://dl.fbaipublicfiles.com/dino/dino_deitsmall16_pretrain/dino_deitsmall16_pretrain.pth and https://dl.fbaipublicfiles.com/dino/dino_vitbase16_pretrain/dino_vitbase16_pretrain.pth.
By the way, I have read your work QSFormer before, it is a great work and has inspired me a lot. If there is an opportunity, we can carry out further cooperation.
Thank you very much for your help and for recognizing our work. It would be my honor to cooperate with you if given the opportunity.
-阳光与茉莉?? @.***
------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2024年5月9日(星期四) 上午10:54 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [codeshop715/SPM] FileNotFoundError: [Errno 2] No such file or directory: 'pretraind model' (Issue #2)
Thank you very much for your interest in our work. You can obtain pre-trained models at https://github.com/facebookresearch/deit/blob/main/README_deit.md or directly download them from https://dl.fbaipublicfiles.com/dino/dino_deitsmall16_pretrain/dino_deitsmall16_pretrain.pth and https://dl.fbaipublicfiles.com/dino/dino_vitbase16_pretrain/dino_vitbase16_pretrain.pth.
By the way, I have read your work QSFormer before, it is a great work and has inspired me a lot. If there is an opportunity, we can carry out further cooperation.
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>
To the best of my knowledge, the pre-trained weights were obtained by training the model on ImageNet2012. Inevitably, ImageNet2012 contains unseen test samples. The use of test samples is prohibited when the model is trained in a few-shot setting, Could you please explain this? Thank you!
When I read this paper, I was surprised by the base result of the model, and I wondered how the pre-training result could reach 93.09. If your work uses pre-trained models on ImageNet2012, I think it is unreasonable because your model has already acquired knowledge on the test set during pre-training. Could you clarify this question?
Thank you very much for your @Cris07lab @peacelwh questions. We use pre-trained models on ImageNet2012 because we followed the work P>M>F [1]. In order to ensure a fair comparison, The same experimental settings are used with him (use pre-trained models on ImageNet2012). Please refer to the Class overlap between pre-training and meta-testing section of this paper for an explanation: Class overlap between pre-training and meta-testing. Although unsupervised pre-training does not utilize labels, it is very likely that some classes used by pre-training also appear in meta-testing. Does this class overlap go against the very definition of few-shot learning? From a meta-learning point of view, the answer is yes. But we argue that class overlap is almost unavoidable unless a careful data split is simulated. For example, in the case of Meta-Dataset, the CUB dataset, the Aircraft dataset and the COCO dataset have a class overlap with ImageNet, but they are still used in meta-testing. As we consider more practical large-scale experiments, the class overlap issue becomes ubiquitous. For more details, please refer to [1].
[1] Hu, Shell Xu, et al. "Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.
The pre-trained model was not provided, and execution to the model/init.py file failed. Could you please share the vit-small and vit-base pre-training models with me? Or provide a download link? Your reply will be highly appreciated. Looking forward to your reply. Thank you very much.