dclm / faceswap

just in case

This repo is for the final project of Machine Learning in 2023 based on Simswap

We provide several models for face swap, including Deform, dancer, shifter

Run dancer:

python train.py --model_name dancer --dataset /content/TrainingData/vggface2_crop_arcfacealign_224 --lambda_cycle 0 --lambda_gp 0 --n_blocks 3 --n_layers 3 --model_freq 3000 --total_step 100000

Run shifter:

python train.py --model_name simplified --dataset /kaggle/working/TrainingData/vggface2_crop_arcfacealign_224 --lambda_gp 0 --n_blocks 6 --n_layers 5 --model_freq 3000 --total_step 100000 --kernel_type deform

Run deform:

python train.py --model_name simswap+=+ --name simswap --dataset /kaggle/working/TrainingData/vggface2_crop_arcfacealign_224 --lambda_gp 0 --n_blocks 6 --n_layers 3 --model_freq 20000 --total_step 100000 --kernel_type deform

In the wandb page, there are different running environments. You can find it by group the runs by batchSize. 4 is on Colab, with this repo, and 8 is on another group member's server, with the original simswap repo.