How to finetune BigGAN on custom datasets?

[Yearsflow]

Hi,

Sorry to bother you again. Last time I wondered how to get pretrained BigGAN parameters on custom datasets, you told me that I could finetune BigGAN. Should I just load pretrained Image-Net parameters and freeze other layers except the last fully-connected layers? I tried to train BigGAN but I found that I didn't have enough GPU memory to train it. Could you please give me some advice? Thank you very much. ^_^

[PatrickZH]

Hi, I followed BigGAN official repo and trained models for CIFAR datasets from scratch. I don't know how many data do you have in your custom datasets. So, finetuning the ImageNet-version BigGAN can be better choice. It does need a lot of GPU memory due to the image size. Maybe you can try some tricks like model quantization or reduce the batchsize.

[Yearsflow]

Thanks for your advice. My custom dataset includes more than 20,000 images with 224*224 size. I can't afford the cost of GPU memory. By the way, if I want to finetune ImageNet-version BigGAN, should I just freeze the most of layers except the last fully-connected layers? Is it right to finetune? I don't know much about finetuning model, so I appreciate it if you can answer me.

[PatrickZH]

You may need to find some tutorial for GAN finetuning. Maybe you can set a very small learning rate and tune all parameters.

[Yearsflow]

Thank you very much. It helps me a lot.