A question about ckpt

[Sample-design-alt]

A nice work! I want to finetune visionTS for classification task, but i have some question for the weights.

What is the difference between https://huggingface.co/facebook/vit-mae-base and the weights your proposed （visiontspp_model.ckpt）in the code?

[Keytoyze]

Thanks for your support! 😊

There are three available checkpoints:

• https://dl.fbaipublicfiles.com/mae/visualize/mae_visualize_vit_base.pth: original MAE model without time series CPT, released by Kaiming.
• https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_base.pth: also without time series CPT, but disgarding the decoder weight.
• visiontspp_model.ckpt: our released ckpt, by continual pre-training mae_visualize_vit_base.pth on time series data (See VisionTS++ paper)

We did not test vit-mae-base on the huggingface so I have no idea what it is.