Is the ViT-L used here pre-trained on imagenet, or is it trained from scratch? about multi-task-transformer

[prismformore]

@tb2-sy Hi, all encoders are pre-trained on ImageNet as stated in the implementation details section of the paper. We directly used the weights provided by timm repository.

Thank you for your interest in our work!

[tb2-sy]

Thanks for the reply, this is a great piece of work!
Will the Vit pre-trained model be automatically loaded when the program is running? I didn't see the process of downloading the pre-trained model. I wonder if I have used timm before and loaded it directly from the default location.

[prismformore]

Thanks for the reply, this is a great piece of work!
Will the Vit pre-trained model be automatically loaded when the program is running? I didn't see the process of downloading the pre-trained model. I wonder if I have used timm before and loaded it directly from the default location.

Timm will load the pre-trained models by setting pretrained=True here: https://github.com/prismformore/InvPT/blob/3b70fcc5a4f7053a7e32a9f85da5dda670c18fba/utils/common_config.py#L17

It will download the model to your $TORCH_HOME/hub if my memory is correct.

[tb2-sy]

Thanks for the reply, this is a great piece of work!
Will the Vit pre-trained model be automatically loaded when the program is running? I didn't see the process of downloading the pre-trained model. I wonder if I have used timm before and loaded it directly from the default location.

Timm will load the pre-trained models by setting pretrained=True here:

https://github.com/prismformore/InvPT/blob/3b70fcc5a4f7053a7e32a9f85da5dda670c18fba/utils/common_config.py#L17

It will download the model to your $TORCH_HOME/hub if my memory is correct.

I have understood the problem, your reply solved my doubts, thank you very much！

[prismformore]

Issue solved.