Train on my own dataset

[49xxy]

Thank you for a job well done! Your idea is very good to alleviate my problem of trying to conditional generation on my own restricted dataset (only 7K)！

I would like to know about implementing your work on my own non-conditional model (based on Stylegan2-ADA) and dataset.
The following is my idea, I do not know whether it is correct or not:
First, the PKL model should convert the weights into PT format, then form my own classified dataset according to your instructions, and then train.

Looking forward to your reply!

[hecoding]

Hi, thanks! Yeah that sounds right. You can directly pick a pretrained model from StyleGAN2(-ADA), modify the code for the hyper-modulation, compile the dataset the way they instruct and then train. You may find easier training if you also do self-alignment. Good luck! On 1 Aug 2022, at 10:38, 49xxy ***@***.******@***.***>> wrote: Thank you for a job well done! Your idea is very good to alleviate my problem of trying to conditional generation on my own restricted dataset (only 7K)！ I would like to know about implementing your work on my own non-conditional model (based on Stylegan2-ADA) and dataset. The following is my idea, I do not know whether it is correct or not: First, the PKL model should convert the weights into PT format, then form my own classified dataset according to your instructions, and then train. Looking forward to your reply! — Reply to this email directly, view it on GitHub<#1>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ABWK3U6RJCAWPWWNGKCEJATVW6EGXANCNFSM55GWJEFA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[49xxy]

Hello, I ‘d like to know about the AFHQ(3 classes) dataset. Your method transfers unconditional GANs to conditional GANs. Does the final condition generation need not specify the class generation, is it the effect of directly generating three animals together?

[hecoding]

Hi, sorry I missed this comment. The final conditional generator would need the class you want to generate, like any other conditional generator. In this case, the generator needs to be given a class vector to modulate the weights to direct the generation towards that specified class. No three animals are generated together, in my opinion.