Why not provide 22k-supervised finetuning model??? I am really shocked by that every available ConvNeXt-V2 pre-training weights has been finetuned on imagenet-1k. Please make 22k-supervised ConvNeXt-V2 open just like ConvNeXt-V1 !!!!!! 🙏🙏🙏

[yan-hao-tian]

Hi, I am looking for the 22k-supervised fine-tuning ConvNeXt-V2-H model without 1k-supervised fine-tuning. I want to use it to fine-tune on ade20k, reproducing the result in Table 7 of the paper.

[blackpearl1022]

@yan-hao-tian
You can see the pretrained ConvNeXt-V2 models on hugging face which were already fine tuned with ImageNet-22k.
For example, convnextv2_base.fcmae_ft_in22k_in1k_384 was pre-trained with ImageNet-22k and you can see other models as well on hugging face.

[yan-hao-tian]

Thanks for the reply.

What I need is the ConvNeXt-v2-Huge weights ended with '22k', which means the model's last pre-training step is fine-tuning on ImageNet22k without then finetuning on ImageNet1k. Because the last line of table 7 in the ConvNeXt-v2 paper achieves 57.0 mIoU with it on ade20k dataset.

It seems nowhere I can find this model.