Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization

Unofficial pytorch-lightning implementation of a paper, "Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization"

Requirements

• torch==1.10.1
• torchvision==0.11.2
• pytorch-lightning==1.5.7
• tqdm==4.62.3
• wandb==0.12.10
• opencv-python-headless==4.5.5.62
• easydict==1.9
• onnxruntime==1.10.0

Install requirements by pip

pip install -r requirements.txt 

Install dependency by docker

docker-compose up -d 

Pretrained weights / models download

• Encoder(VGG) weights download by "pytorch-AdaIN" repo
• Decoder weights download by link
• Net Torchscript model download by link
• Net onnx model download by link

Dataset

• content : COCO DATASET
• style : WikiArt

Training

1. download dataset (content, style)
2. download Encoder weights and mv to weights dir
3. check dataset path
4. run script

python main.py \
--content_root_dir="{YOUR CONTENT DATASET PATH}" \
--style_root_dir="{YOUR STYLE DATASET PATH}" \
--num_workers=8 \
--batch_size=8 \
--log_every_n_steps=5 \
--gpus=1 \
--max_steps=16000

Training result

result plot, images

loss plot

sample images

• video

APP

Streamlit app source code referenced "Neural-Style-Transfer-Streamlit "

Menual

1. Download "Torchscript" model to "app/src" directory

2. Install APP dependency by pip

cd app
pip install -r requirements.txt 

• torch==1.10.1
• torchvision==0.11.2
• pytorch-lightning==1.5.7
• streamlit==1.7.0
• opencv-python-headless==4.5.5.62

2. Install AdaIN app by docker

2.1. docker env

• MODEL_PATH : torchscript file path

2.2. Start container

cd app
docker-compose up -d 

References

• Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization
• pytorch-AdaIN
• Neural-Style-Transfer-Streamlit