ternerss / mobile-style-transfer

Mobile Style Transfer in PyTorch

Mobile Style Transfer

Implementation of Neural Style Transfer based on MobileNet architecture with Sobel loss idea.

Installation

1. Clone this repository:

git clone https://github.com/ternerss/mobile-style-transfer.git
cd mobile-style-transfer

2. install requirements:

pip install -r requirements.txt

Usage

Dataset: COCO 2017 Val images

• Train:

python train.py --dataroot data/ --style data/styles/wave.jpeg --batch_size 16 --lr 0.001 --epochs 30

• Test:

python test.py --input data/val/building.jpeg --weights weights mobile_style_transfer.pth --out out/ 

Main idea

References

• A Neural Algorithm of Artistic Style
• Perceptual Losses for Real-Time Style Transfer and Super-Resolution
• Instance Normalization
• MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications