Question about training efficientvit classification model on imagenet-22k dataset

[lufanma]

Thanks for your great work!

I plan to train efficientvit classification model on imagenet-22k dataset for a better backbone initialization for open vocabulary object detection.

I know that the released classification model now are all pretrained on imagenet-1k dataset.

But I'm wondering how the training hyperparams should be set for the imagenet-22k pretraining. And should I use the 1k pretrained model to initialize or just train from scratch, which one is prefered ?

Hope your reply @han-cai ! Thanks ！！！

[han-cai]

Thank you for the question. I have not trained efficientvit classification models on the imagenet-22k dataset. I am not sure what the best practice is for setting training hyperparameters for this case. Regarding training from scratch or using the 1k pretrained weight, I would prefer using the 1k pretrained weight, as it can reduce the training cost, according to https://github.com/Alibaba-MIIL/ImageNet21K.