softlabel-gan

The repository for "Label Augmentation as Inter-class Data Augmentation for Conditional Image Synthesis with Imbalanced Data," WACV 2024
Paper

This repo is implemented upon the BigGAN-PyTorch repo and DiffAugment repo. The main dependencies are:

Python 3.7.9 or later
PyTorch 1.10 or later
TensorFlow 1.15 with GPU support

Setup enviroment

conda env create tf_env.yaml
conda activate tf

Training classifier

python3 SpinalNet/Transfer_Learning_AnimeFace.py

Training

python train.py --experiment_name animeface --DiffAugment color,translation,cutout --mirror_augment --which_best FID --num_inception_images 5000 --shuffle --batch_size 64 --parallel --num_G_accumulations 1 --num_D_accumulations 1 --num_epochs 1000 --num_D_steps 1 --G_lr 1e-4 --D_lr 4e-4 --dataset T128 --G_ch 80 --D_ch 80 --G_depth 1 --D_depth 1 --G_shared --shared_dim 128 --dim_z 120 --hier --ema --use_ema --ema_start 20000 --test_every 4000 --save_every 2000 --target_type softmax --ann_file ~/tinyimagenet128_softmax.npz  --adam_eps 1e-6 --SN_eps 1e-6 --BN_eps 1e-4 --num_worker 32  --load_in_mem --G_eval_mode --num_samples 50000

Evaluation

python eval.py --experiment_name animeface --network weights/animeface/G_ema_best.pth --num_inception_images 5000 --batch_size 32 --parallel --dataset T128 --G_ch 80 --D_ch 80 --G_depth 1 --D_depth 1 --G_shared --shared_dim 128 --dim_z 120 --hier --ema --use_ema --ema_start 20000 --target_type softmax --ann_file ~/tinyimagenet128_softmax.npz  --adam_eps 1e-6 --SN_eps 1e-6 --BN_eps 1e-4 --num_worker 32  --load_in_mem --G_eval_mode

Acknowledgements