[AAAI '23] Progressive Few-Shot Adaptation of Generative Model with Align-Free Spatial Correlation

Official PyTorch implementation of ASC adaptation.

Paper

Overall framework

Note: The base model is taken from Few-shot-gan-adaptation

• Linux
• NVIDIA GPU + CUDA CuDNN 10.2
• PyTorch 1.7.0
• Python 3.6.9
• Install all the libraries through pip install -r requirements.txt

Training

Pretrained Source domain model

• FFHQ Note: Please download it into your training path.

Dataset for training

10 shot

• Sketches | Amedeo Modigliani | Babies | Sunglasses | Animation faces

5 shot

• Sketches | Amedeo Modigliani | Babies | Sunglasses | Animation faces

Note: Please download it into your training <root/raw_data> path.

python prepare_data.py --out ./processed_data/<dataset_name> --size 256 ./raw_data/<dataset_name>

CUDA_VISIBLE_DEVICES=0 python train.py --ckpt_source /path/to/source_model --data_path /path/to/target_data --exp <exp_name> \
    --use_ema --use_rel_ema --k 1.0 --extra_ema \
    --use_flow --winsize_inc 1.0002 --winsize 16 --cutoff 0.6 \
    --use_pred --pred_noiseW 2.0 --use_flow_pred --blur 7 \
    --LFC --m 2 --LFCw 0.5 --extra_step 0

Evaluating

Dataset for FID evaluating

• Sketches test set from Few-shot-gan-adaptation --> link
• Babies test set from Few-shot-gan-adaptation --> link
• Sunglassies test set from Few-shot-gan-adaptation --> link
• Animation Faces test set from AniGAN --> link

# if there is a large dataset of the target, use --test_imgs for FID evaluation
# if not, please remove this option. 
CUDA_VISIBLE_DEVICES=0 python evals.py --ckpt_source /path/to/source_model --model_ckpt /path/to/target_model \
    --out /path/to/out --train_imgs /path/to/processed_data/domain --source_key <domain_key> \
    --batch 64 --n_sample <test_sample_count> --seed <number> --test_imgs /path/raw_data/domain/images/

Dataset for Reconstruction evaluating

Note: To evaluate identity, please training source to target first, then try to re-train it to FFHQ again with this following ffhq-10s images:

• ffhq-10s

After finish the reconstruction (source -> target -> source), evaluate it as this following step:

# if there is a large dataset of the target, use --test_imgs for FID evaluation
# if not, please remove this option. 
CUDA_VISIBLE_DEVICES=0 python evals.py --ckpt_source /path/to/source_model --model_ckpt /path/to/target_model \
    --out /path/to/out --train_imgs /path/to/processed_data/domain --source_key <domain_key> \
    --batch 64 --n_sample <test_sample_count> --seed <number> --test_imgs /path/raw_data/domain/images/
    
CUDA_VISIBLE_DEVICES=0 python recon_evals.py --source_ckpt /path/to/source_model --model_ckpt /path/to/target_model \
   --out /path/to/out --source_key <domain_key> \
   --batch 4 --n_sample <test_sample_count> --seed <number>