cryer/simple_style_transfer

simple style transfer with a very clear code

simple_style_transfer

simple style transfer with a very clear code

Note

I have created a repo about style transfer several months ago.It is implemented referring to official tutorial. And its code may be a little hard to read and understand.However,this code is very simple and clear,and has a lot of tricks.You may understand them and use them in your own porjects.

Usage

git clone https://github.com/cryer/simple_style_transfer.git
cd simple_style_transfer
pip install -r requirements.txt
python train.py train

Add additional configs:

    --content = './image/gakki.jpeg'
    --style = './image/in1.png'
    --max_size = 400
    --total_step = 5000
    --log_step = 10
    --sample_step = 1000
    --style_weight = 100
    --lr = 0.003
    --use_gpu = True

Pay attention to images' path.

Results

Use my idol gakki to do a test.

本当に かわいいねｗｗ 女はみずだ！

content image	style image	transfer image

Difference

Compare to my previous style transfer repo or official implementation,this repo's computation of content loss is different. Using five layers' content loss togeter instead of only one layer's content loss in paper and official implementation.So it may take longer time,but a little bit improvement a think,because if it improves a lot , paper may likely use this kind of loss. You can change loss computation part to do a test.

Reference

Yunjey,Yunjey's code is always very clear,you may learn a lot from him.