How can I use other pretrained ViTs or CNNs?

[lianzheng-research]

Thanks for your excellent work!
For the plain ViT, I want to use its tiny version to support mobile applications. How can I load the multi-modal pretrained parameters for ViT-T model? Pretrained parameters of DeiT-T seem to perform a little bit weak on semantic segmentation task.
For SPM, I want to use a pretrained resnet-50 or resnet-18 to replace it. Do you think this will benefit the model performance? What should I do for this?
Thanks you very much!