Skip to content

Latest commit

 

History

History

training

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
readme.md
readme.md
 
 

readme.md

Please following eg3d and stylegan2-ada-pytorch for environment installation respectively.

Code

Clone our eg3d and stylegan2-ada-pytorch repos.

cd training
git clone [email protected]:oneThousand1000/LPFF-eg3d.git
git clone [email protected]:oneThousand1000/LPFF-stylegan2-ada-pytorch.git

Training

EG3D

cd eg3d/eg3d

New options:

--camera_sample_mode

camera_sample_mode is the dataset used to sample the camera for training.

--gen_pose_cond_avg

When set --gen_pose_cond_avg=True and --gen_pose_cond=False, the generator will be conditioned on the average camera.

var1-64.pkl

python train.py --outdir=./training-runs/var1-64 \
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 --gen_pose_cond=True  \
    --camera_sample_mode=FFHQ_LPFF

var1-128.pkl

python train.py --outdir=./training-runs/var1-128\
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 --gen_pose_cond=True --neural_rendering_resolution_final=128 --kimg=20000 \
    --resume=var1-64.pkl  \
    --camera_sample_mode=FFHQ_LPFF

var2-64.pkl

python train.py --outdir=./training-runs/var2-64 \
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 --gen_pose_cond=True  --gpc_reg_prob=0.8 --kimg=20000  \
    --resume=var1-64.pkl \
    --camera_sample_mode=FFHQ_LPFF_rebalanced

var2-128.pkl

python train.py --outdir=./training-runs/var2-128 \
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 --gen_pose_cond=True  --gpc_reg_prob=0.8 --neural_rendering_resolution_final=128 --kimg=20000\
    --resume=var2-64.pkl \
    --camera_sample_mode=FFHQ_LPFF_rebalanced

var3-64.pkl

python train.py --outdir=./training-runs/var3-64 \
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 \
    --gen_pose_cond=False --gen_pose_cond_avg=True \ 
    --kimg=20000  \
    --resume=var1-64.pkl \
    --camera_sample_mode=FFHQ_LPFF

var3-128.pkl

python train.py --outdir=./training-runs/var3-128 \
    --cfg=ffhq --data=./dataset/eg3d_dataset.zip  \
    --gpus=8 --batch=32 --gamma=1 \
    --gen_pose_cond=False --gen_pose_cond_avg=True \
    --neural_rendering_resolution_final=128 --kimg=20000  \
    --resume=var3-64.pkl \
    --rebalance=FFHQ_LPFF

StyleGAN

cd stylegan2-ada-pytorch

FFHQ_LPFF.pkl

python train.py --outdir=./training-runs/FFHQ_LPFF \
	--data=./dataset/stylegan_dataset.zip \
	--gpus=8 --cfg=stylegan2 --mirror=1 --kimg=35000 \
	--camera_sample_mode=FFHQ_LPFF

FFHQ_LPFF_rebalanced.pkl

python train.py --outdir=./training-runs/FFHQ_LPFF_rebalanced \
	--data=./dataset/stylegan_dataset.zip \
	--gpus=8 --cfg=stylegan2 --mirror=1 --kimg=35000\
	--resume=FFHQ_LPFF.pkl\
	--camera_sample_mode=FFHQ_LPFF_rebalanced

EG3D FID Evaluation

cd eg3d
cd eg3d

--camera_sample_mode

camera_sample_mode is the dataset used to sample rendering camera ( $c_r$ in our main paper). Notice the feature stats for the dataset are computed using the same dataset as camera_sample_mode (e.g., when set --camera_sample_mode=FFHQ, we will use the FFHQ dataset to compute the dataset feature).

--conditional_camera_sample_mode

conditional_camera_sample_mode is the dataset used to sample the conditional camera ( $c_g$ in our main paper). If you want to fix $c_g$ as the average camera, please use --conditional_camera_sample_mode=avg. If you want to set $c_g = c_r$, please use --conditional_camera_sample_mode=None. If you want to sample $c_g$ from a certain dataset, please use --conditional_camera_sample_mode=LPFF or --conditional_camera_sample_mode=FFHQ.

Usage

$c_g =c_{avg}, \quad c_r \sim FFHQ$

python calc_metrics.py  --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \  
	--network=network.pkl \  
	--gpus=8  \
	--conditional_camera_sample_mode=avg --camera_sample_mode=FFHQ

$c_g =c_{avg}, \quad c_r \sim LPFF$

python calc_metrics.py  --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \  
	--network=network.pkl \  
	--gpus=8  \
	--conditional_camera_sample_mode=avg --camera_sample_mode=LPFF

$c_g \sim FFHQ, \quad c_r \sim FFHQ$

python calc_metrics.py --metrics=fid50k_full\
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--conditional_camera_sample_mode=FFHQ --camera_sample_mode=FFHQ

$c_g \sim FFHQ, \quad c_r \sim LPFF$

python calc_metrics.py --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--conditional_camera_sample_mode=FFHQ --camera_sample_mode=LPFF

$c_g \sim LPFF, \quad c_r \sim FFHQ$

python calc_metrics.py --metrics=fid50k_full  \
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--conditional_camera_sample_mode=LPFF --camera_sample_mode=FFHQ

$c_g \sim LPFF, \quad c_r \sim LPFF$

python calc_metrics.py --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--conditional_camera_sample_mode=LPFF --camera_sample_mode=LPFF

$c_g \sim FFHQ, \quad c_r =c_g$

python calc_metrics.py --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--camera_sample_mode=FFHQ

$c_g \sim LPFF, \quad c_r =c_g$

python calc_metrics.py --metrics=fid50k_full \
	--data=./dataset/eg3d_dataset.zip \
	--network=network.pkl \ 
	--gpus=8 \
	--camera_sample_mode=LPFF