Style transfer

This is tensorflow implementation of 'Multimodal Transfer: A Hierarchical Deep Convolutional Neural Network for Fast Artistic Style Transfer' which generates stylized image in high resulution such as 1024 pixels.

Download program

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

Train

To train network, I use MS coco dataset.

$ cd multimodal_style_transfer
$ bash get_coco.sh

• downloaded image is in 'data/train2014'.

For stylization, pretrained VGG16 is necessary.

$ bash get_vgg16.sh

Then training is SIMPLE.

$ python train.py

• If you have multiple GPU cards, use CUDA_VISIBLE_DEVICES to specify GPU card.
• Trained model is in summary.

During training, you can see generated images through tensorboard.

$ tensorboard --logdir=summary

Freeze model

$ sh freeze.sh 10000

• parameter is iteration number among saved check point files.
• It generates pb file which contains weights as contant.

Test

$ python stylize.py --model=models/starry_night.pb --input_image=test_images/jolie.jpg

• It generates hierarchical stylized images and save them to 'test_images/jolie_output_1.jpg', 'test_images/jolie_output_2.jpg', and 'test_images/jolie_output_3.jpg'. Their sizes are 256, 512 and 1024 in short edge.
• Parameters:

--model : freezed model path
--input_image : image file path to stylize
--hierarchical_short_edges : three short edge length to generate images. (default is 256, 512, 1024)

Examples

	Input	Output(256px)	Output(512px)	Output(1024px)
Angelina Jolie
Dinosour
Ryan
Cheez
Herb

Acknowledgement

• For instance normalization, I refer 'https://github.com/ghwatson/faststyle'.
• For pretrained VGG16 network, I refer 'https://www.cs.toronto.edu/~frossard/post/vgg16/'.