prisma

Online fast neural style transfer

Requirements

• Python 3+
• Tensorflow 1.1.0+
• Scipy
• Flask
• Flask-Mail
• Celery
• Redis

Setup

• Dependencies

pip3 install numpy pillow scipy
pip3 install flask flask-mail celery redis
pip3 install tensorflow // for cpu

• Download Models

http://pan.baidu.com/s/1pLPSXdx

mv models/ prisma/

• Mailbox

default_config.py

MAIL_SERVER = 'xxxxx'
MAIL_PORT = xxx
MAIL_USERNAME = 'xxxxxx'
MAIL_PASSWORD = 'xxxxxx'

• Run Redis

default_config.py

CELERY_BROKER_URL = 'redis://localhost:6379/0'
./redis-server

• Run Celery

celery -A server.celery worker

• Run Flask

python server.py

Training

• Download COCO dataset and VGG19 model

VGG19 model

COCO dataset

• Put the model and dataset into "prisma/"

# Recommend using tensorflow of gpu version
python3 train.py --STYLE_IMAGES style-image.jpg --CONTENT_WEIGHT 1.0 --STYLE_WEIGHT 10.0 --MODEL_PATH models/newmodel.ckpt

mv models/newmodel.ckpt-done models/newmodel.ckpt

# Test
python3 eval.py --CONTENT_IMAGE content-image.jpg --MODEL_PATH models/newmodel.ckpt --OUTPUT_FOLDER generate/

• Add to Prisma

default_config.py

MODEL_FILES = set(['newmodel.ckpt', ......])

# Put a example image ("newmodel.jpg") into "static/models_image/"

Screenshot

http://localhost:5000/

Reference

• A Neural Algorithm of Artistic Style
• Perceptual Losses for Real-Time Style Transfer and Super-Resolution
• Instance Normalization: The Missing Ingredient for Fast Stylization
• OlavHN/fast-neural-style