Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/JiangQH/Style-Transfer-In-Tensorflow
style transfer with both pycaffe and tensorflow
https://github.com/JiangQH/Style-Transfer-In-Tensorflow
Last synced: 3 months ago
JSON representation
style transfer with both pycaffe and tensorflow
- Host: GitHub
- URL: https://github.com/JiangQH/Style-Transfer-In-Tensorflow
- Owner: JiangQH
- Created: 2017-02-24T05:38:21.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2017-05-08T09:37:19.000Z (over 7 years ago)
- Last Synced: 2024-05-19T23:36:50.111Z (6 months ago)
- Language: Python
- Homepage:
- Size: 29.9 MB
- Stars: 4
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Style Transfer whole in one
This repo contains two copies of code, both do the style transfer job. Using the code you can transfer styles from any image to another.
**The optimized method**, based on [A Neural Algorithm of Artistic Style](https://arxiv.org/abs/1508.06576). It is an optimized way of doing style transfer and it may take several minutes to complete the transfer job. But no pre-trained model is needed, so you can transfer style from any image as you like without the training procedure. If you want to transfer a style without training and can bear the time to deploy, try this method. Refer to the document for more detail.
**Fast Style Transfer**, based on [Perceptual Losses for Real-Time Style Transfer and Super-Resolution](https://arxiv.org/abs/1603.08155)[Instance Normalization: The Missing Ingredient for Fast Stylization](https://arxiv.org/abs/1607.08022). It can do the transfer job in a second, and you can even do the transfer job to videos! pre-trained style model is needed. If you care the deploy time and want to do real time style-transfer, try this method. Refer to the document for more detail.
# Image Stylization
Here we show the results of both two methods. Note here that the image to transfer style from is called style-image and the image to be transfered is called content image. The transfered result can be adjust by the weight and layers to use.
**The optimized way**
**The Fast Style Transfer way**
# Video Stylization
Here only the Fast Style Transfer be deployed to the video style transfer, optimized way consumes too much time which is not feasible. Refere to [Youtube](https://youtu.be/-aJYsQ2Bt60) to watch the whole video.
# Documentation
All two works are done with python2.7, make sure you have it installed.
## The optimized method
### Requirements
[caffe](http://caffe.berkeleyvision.org/) compiled with pycaffe, and make sure the PYTHONPATH is correctly set.
[pre-trained vgg16 model](http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel)
[scipy&numpy](https://www.scipy.org/)
[pillow](https://pillow.readthedocs.io/en/4.1.x/)
### Stylizing images
Use the ``style.py`` to do the optimized job. Usage:
``python style.py -c /pathto/content_img -s /pathto/style_img``
Result is saved in current dir. You can see the code in optimize for more detail. Note since it has to optimize, several minutes can be taken to complete the whole, and time to consume is related to the image size and your device.
## Fast Style Transfer
When doing this fast style transfer, I find caffe's way to add new layer can be really complicated and it is not friendly to python interface. So I change the platform to tensorflow.
### Requirements
[tensorflow](https://www.tensorflow.org/)
[scipy&numpy](https://www.scipy.org/)
[pillow](https://pillow.readthedocs.io/en/4.1.x/)
[opencv2.4](https://opencv-python-tutroals.readthedocs.io/en/latest/) --- needed if you want to style the video.
### Train Style Models
If you want to train your own style models, addition requirements are needed.
[COCO dataset](http://msvocds.blob.core.windows.net/coco2014/train2014.zip), release it in a content folder
[pre-trained vgg16 model](), note this model has been converted to be loaded with numpy interface
[pyyaml](http://pyyaml.org/), needed to parse the config file
Now change to the src/fast folder, change the contents in conf.yml with guidance in it and everything is done.
``python train.py -c conf.yml``
### Stylizing images
Since it can do the style transfer job in a second, we can style many images all in once. Call the ``fast_style.py`` script as follows:
``python fast_style.py -m -i -o ``
For more possible params, run ``python fast_style.py`` to view.
### Styling Videos
Use the ``fast_style_video.py`` to transfer style for videos. Note opencv is needed when using this script. Usage:
``python fast_style_video.py -m -i -o ``
For more detail, refer to the code in it.
### Trained Models
released soon.
### License
Free for noncommercial use, contact me if for commercial use and feel free to email me if you have any question.