Fast Image Style Transfer

Gonna try to transfer the style of a woodcut to other images.

Public

Training

Download COCO2014 images.

wget --progress=dot:giga -P /results/ \
  http://images.cocodataset.org/zips/val2014.zip \
  http://images.cocodataset.org/zips/test2014.zip \
  http://images.cocodataset.org/annotations/annotations_trainval2014.zip \
  http://images.cocodataset.org/annotations/image_info_test2014.zip

399.3s

Train (Bash in Python)

annotations_trainval2014.zip

image_info_test2014.zip

test2014.zip

val2014.zip

mkdir -p coco
cd coco
unzip -q test2014.zip
unzip -q test2014.zip

137.2s

Train (Bash in Python)

Train (~2 hours on a P100).

cd /examples/fast_neural_style
python neural_style/neural_style.py train --dataset /coco --cuda 1 \
  --save-model-dir /results --checkpoint-model-dir /shared \
  --style-image FlammarionWoodcut-crop.png
 \
  --epochs 2 --batch-size 14 --log-interval 280 --checkpoint-interval 2000 \
  --content-weight 1e5 --style-weight 1e10

7229.7s

Train (Bash in Python)

epoch_2_Tue_Sep_17_05:29:12_2019_100000.0_10000000000.0.model

Style Transfer

An Antikythera Mechanism recreation.

cd /examples/fast_neural_style
python neural_style/neural_style.py eval --cuda 0 \
  --model Antikythera_model_front_panel_Mogi_Vicentini_2007.JPG
 \
  --content-image Antikythera_model_front_panel_Mogi_Vicentini_2007.JPG
 \
  --output-image /results/out.jpg

33.7s

Eval (Bash)

Woah.

Brutalist architecture.

cd /examples/fast_neural_style
python neural_style/neural_style.py eval --cuda 0 \
  --model 812px-NLNZ_ext_5.jpg
 \
  --content-image 812px-NLNZ_ext_5.jpg
 \
  --output-image /results/out.jpg

6.9s

Eval (Bash)

Gears in a printing press.

cd /examples/fast_neural_style
python neural_style/neural_style.py eval --cuda 0 \
  --model PIXNIO-206511-725x483.jpeg
 \
  --content-image PIXNIO-206511-725x483.jpeg
 \
  --output-image /results/out.jpg

6.7s

Eval (Bash)