Fine-tuning FFHQ curriculum

PeterouZh / CIPS-3D

3D-aware GANs based on NeRF (arXiv).

MIT License

610 stars 60 forks source link

Fine-tuning FFHQ curriculum #24

Open longnhatne opened 2 years ago

longnhatne commented 2 years ago

Can I have .yaml config file for fine-tuning FFHQ? I'm using your exp/cips3d/configs/finetune_exp.yaml (for AFHQ) and after 7 hours, FID ~ 145 :(((.

PeterouZh commented 2 years ago

OK, I have updated the repo.

Preparing AFHQ dataset:

# Prepare training dataset.
python scripts/dataset_tool.py \
    --source=datasets/AFHQv2/AFHQv2 \
    --dest=datasets/AFHQv2/AFHQv2_stylegan2.zip

The pre-trained FFHQ checkpoints: https://github.com/PeterouZh/CIPS-3D#pre-trained-checkpoints-including-generator-and-discriminator. The start script: exp/cips3d/bash/finetuning_exp/finetune_afhq.sh. The yaml file: exp/cips3d/configs/finetune_afhq.yaml.