Closed zsMoonshine closed 1 year ago
Hi!
You are correct, StyleCLIP is implemented from scratch. The already existing baseline would not work on fashion data since it is trained on another dataset.
Since I don't think this is a popular request I was not intending to note the StyleCLIP-related instructions in README. I would instead suggest to email me and I will send instructions associated with its implementation.
Best, Martin
@zsMoonshine I am facing some problems in training FICE from scratch, can you share the code you used for training. I am pretty new and am unable to analyse this whole codebase
@zsMoonshine I am facing some problems in training FICE from scratch, can you share the code you used for training. I am pretty new and am unable to analyse this whole codebase
Your email address please. And you can get help to train the model from scratch refering to this Commit[https://github.com/MartinPernus/FICE/commit/7354216f462edc71854bfdfe4fa617bf5531b10f adds additional instructions about training from scratch!]
@zsMoonshine arorasagar912@gmail.com , thanks I'll refer to the commit too
Hi, Martin. With your help I have successfully trained FICE's models from scratch. And I noticed that in the FICE paper, a global direction has been used on several GAN inversion encoders(pSp, E4e, ReStyle, HyperStyle).
I've tried global direction using code provided by styleclip's global_pytorch. And I noticed that they use e4e. But an error happened, which indicates that the keys in weight dict are not right.
I guess that it's because the stylegan2 checkpoint is based on stylegan2-ada and the e4e in FICE is trained on stylegan2-ada too. I knew that this is slight modification in e4e's psp.py, but I don't know if it's related to the error, or it's just for the reason of the diffrence between stylegan2-ada and stylegan2.
So how did you implement the global direction? To train a new model using the official stylegan2 or do something to change weight dict? Looking forward to your reply. Thanks a lot.