one trivial question about DMAE code

[go-ahead-maker]

Hi Yutong and Cihang,

Thank you for contributing such a great work. Recently, I have been studying the code of DMAE, and I wanna ask a stupid question.

In file models_mae_distill.py, the intermediate features collected (i.e., forward_encoder_customized func.) from ViT encoder are transformed into fp32 by using torch.float32, is this because the teacher model's tensor type is by default fp32 and it requires that both student and teacher's features are alignment w.r.t data type?

Thank a lot.