How do I use this project on my own dataset?

[luoling1993]

I have a image2image tasks,and I have the image pairs, nne has a watermark, the other one doesn't. I want to use this project to train a model to remove the watermark.How do I use this project?

[yuval-alaluf]

Please see the README section on preparing your own data: https://github.com/eladrich/pixel2style2pixel#preparing-your-data After defining your data paths, transforms, and experiment type, you can train your image to image task. We also provide example train commands which you can rely on.

[luoling1993]

When I read the README, I know how to preparing my data. But there is some confusion about pretrained model.

1. Because of my task is image2image, so id_lambda=0
2. pSp：encoder and decoder, I found that ir_se50 was used by default.
   • Should I use this pretrained model in my task?
   • If not, how do I pretrained my model?

[yuval-alaluf]

Regarding the id_lambda, if you're not working on facial images, then you can try setting id_lambda=0. Regarding the pre-trained model, please see Issue #76 .

[luoling1993]

When I read the issue#76 and the code, do I understand you correctly.

1. encoder, default ir_se50, trained by facical datasets. Because the params will be updated while training, I can use this pretrained model in my tasks, or replaced with a Resnet-IR trained based on imagenet.
2. decoder, default styleGan2, trained by FFHQ datasets, and NOT update default. I can set train_decoder=True to fine-tune the decoder

[yuval-alaluf]

Yes you understood correctly

[luoling1993]

Thank you very much for your answers！