fullfanta / real_time_style_transfer

tensorflow implementation of 'Perceptual Losses for Real-Time Style Transfer and Super-Resolution'

Style transfer

This is tensorflow implementation of 'Perceptual Losses for Real-Time Style Transfer and Super-Resolution'.

Download program

$ git clone https://github.com/fullfanta/real_time_style_transfer.git

Download training data

$ cd real_time_style_transfer
$ sh get_coco.sh

Download vgg16 model

$ sh get_vgg16.sh

Train

$ python train.py

If you have multiple GPU cards, use CUDA_VISIBLE_DEVICES to specify GPU card. Trained model is in summary. To adjust weights of content loss and style loss, you can set the parameter alpha. loss equation is alpha * content loss + (1 - alpha) * style loss. Default is 0.1 During training, you can see generated images through tensorboard.

$ tensorboard --logdir=summary

Freeze model

$ sh freeze.sh 40000

It generates pb file which contains weights as contant. First arugment means the iteration number of saved check point file.

Test

$ python stylize.py --model=models/starry_night.pb --input_image=test_images/Aaron_Eckhart_0001.jpg
$ python stylize.py --model=models/starry_night.pb --input_image=test_images/jolie.jpg --resize_ratio=2
$ python stylize.py --model=models/starry_night.pb --input_image=test_images/dinosour.png --resize_ratio=2
$ python stylize.py --model=models/starry_night.pb --input_image=test_images/ryan.png --resize_ratio=3
$ python stylize.py --model=models/starry_night.pb --input_image=test_images/herb.png --resize_ratio=2
$ python stylize.py --model=models/starry_night.pb --input_image=test_images/cheez.png --resize_ratio=4

It generates stylized images and save them to 'test_images'.

Examples

	Input	Output
Aaron Echart
Angelina Jolie
Dinosour
Ryan
Herb
Cheez