search

ryongithub / GatedPromptTuning

MIT License
48 stars 4 forks source link

DIP #1

Open xiaoyaod opened 1 year ago

xiaoyaod commented 1 year ago

how DIP work

ryongithub commented 1 year ago

Deep image prior trains a lightweight reconstruction network to restore the original image from fixed random noise. The reconstruction network is trained to generate restored image to be closer to the original image in the representation space of a neural network, rather than in the input space.

Thank you.

xiaoyaod commented 1 year ago

Do you have plans to release DIP training code

Whether the input of DIP is the characteristic output of backbone network image

ryongithub commented 1 year ago

We utilized the implementation of DIP available at https://github.com/nanxuanzhao/Good_transfer. (repository for "What Makes Instance Discrimination Good For Transfer Learning?", Zhao et al., ICLR 2021) In DIP, the input for the reconstruction network is a fixed random noise. But the training loss is computed in the representation space of a neural network. In our work, we used patch representation from the ViT block for computing loss between the original image and the reconstructed image.

Thank you