Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

在自己的领域的数据集上按照你的conflict-robust cluster discrimination再微调如何? #89

Open
tengshaofeng opened this issue Jan 16, 2025 · 5 comments

Comments

@tengshaofeng
Copy link

真的是很棒的工作。我想请教一下,我想加载你的模型作为预训练初始化,然后我用我自己的数据集(如电商场景),大概100万张图片,用你的conflict-robust cluster discrimination方式再微调一下,你觉得会有大的提高吗。因为我看你的结果里在In-shop数据集上表现是86.7,但是目前最好的已经达到91.9了,所以是不是按照我说的用领域内的图片按照你的方式微调一下会更好

@anxiangsir
Copy link
Collaborator

当然可以,我已经有 inshop 微调的结果了

image

出处之这里:
https://github.com/deepglint/unicom/tree/main/unicom#result-of-supervised-image-retrieval

@tengshaofeng
Copy link
Author

@anxiangsir 这么强的吗,我看paperwithcode的最好效果在in-shop上才91.9

Image
你这个也是Recall@1的结果吗,我记得in-shop的数据集是这样描述的:
7,982 number of clothing items;
52,712 number of in-shop clothes images, and ~200,000 cross-pose/scale pairs;
Each image is annotated by bounding box, clothing type and pose type.
对于in-shop而言,训练集是啥,测试集是啥啊,怎么分怎么测的呀。
期待大佬的回复

@tengshaofeng
Copy link
Author

@anxiangsir 另外问一下Result of Zero-Shot Image Retrieval里面最好的结果是86.7是代表只在laion400M做预训练直接作用于in-shop数据集来测, 而Result of Supervised Image Retrieval里面最好的结果是96.7是在laion400M做预训练后拿in-shop数据集进一步用conflict-robust cluster discrimination方式再微调,对吗?

@anxiangsir
Copy link
Collaborator

@anxiangsir 另外问一下Result of Zero-Shot Image Retrieval里面最好的结果是86.7是代表只在laion400M做预训练直接作用于in-shop数据集来测, 而Result of Supervised Image Retrieval里面最好的结果是96.7是在laion400M做预训练后拿in-shop数据集进一步用conflict-robust cluster discrimination方式再微调,对吗?

你理解的没有错,是可以在 paper with code 上面更新结果的

@tengshaofeng
Copy link
Author

tengshaofeng commented Feb 12, 2025

@anxiangsir ,大佬,不好意思再来打扰到您,您之前说你已经有了微调in-shop的模型,

unicom.available_models()
['ViT-B/32', 'ViT-B/16', 'ViT-L/14', 'ViT-L/14@336px']
unicom.load('ViT-B/32')
1%|▍ | 4.53M/385M [00:27<50:34, 132kiB/s]

通过这种方式只能加载到zero-shot方式的模型吧,那[result-of-supervised-image-retrieval]中对应监督微调in-shop的模型在哪里呀?
(https://github.com/deepglint/unicom/tree/main/unicom#result-of-supervised-image-retrieval)

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants