fast-neural-style

what is this?

style transfer for 360 photos and videos.

specs

• ubuntu 14.04
• torch. includes nn, cutorch, image, lua-cjson, cunn and cuDNN bindings - the good stuff.

ml deets

for more information on approach and dataset, refer to “perceptual losses for real time style transfer and super resolution.”

get docker image

building torch/lua from source for osx ruined my weekend. it is not for the faint of heart so you should do one of the following:

option 1: pull from docker hub

docker pull mynameisvinn/unclip_style_transfer

option 2: build locally (recommended for now)

git clone https://github.com/mynameisvinn/unclip_style_transfer
cd unclip_style_transfer
docker build -t mynameisvinn/unclip_style_transfer .

run

option 1 (recommended)

assuming youve placed images in /date/in, from command line, do

docker run -v /Users/vincenttang/dropbox/temp/unclip_style_transfer/data:/root/fast-neural-style/data mynameisvinn/unclip_style_transfer th fast_neural_style.lua \
  -model models/eccv16/the_wave.t7 \
  -image_size 200 \
  -input_dir data/in/ \
  -output_dir data/out/

• -model refers to pretrained models. styles include la muse, composition vii, starry night, and the wave.
• -image_size refers to height/width. use 0 to keep original dimensions.
• -input_dir refers to path to folder containing images.
• -output_dir refers to path to folder for stylized images.

if successful, you should see modified images in /data/out.

option 2 - from container

from command line, do

docker run -it -v /Users/vincenttang/dropbox/temp/unclip_style_transfer/data:/root/fast-neural-style/data mynameisvinn/unclip_style_transfer

then, from container, do

th fast_neural_style.lua \
  -model models/eccv16/the_wave.t7 \
  -image_size 200 \
  -input_dir data/in/ \
  -output_dir data/out/