chainer-dfi

Implementation of "Deep Feature Interpolation for Image Content Changes"(https://arxiv.org/abs/1611.05507) using Chainer.

Requirements

• Python 2.7

• Chainer 2.0.0

• Cupy 1.0.0

• Pillow 3.1.0

Usage

Download Caffe model and convert

Download Caffe VGG-19 layer model

Download VGG_ILSVRC_19_layers.caffemodel from https://gist.github.com/ksimonyan/3785162f95cd2d5fee77.

Convert Caffe model to Chainer model.

$ python src/create_chainer_model.py

To use Labeled Faces in the Wild (LFW)

Download dataset

• Download "All images aligned with deep funneling"(lfw-deepfunneled.tgz) from LFW web site

• Download "LFW attributes file"(lfw_attributes.txt) from the same site.

• Extract tgz file.

Interpolate feature

Example:

python src/train_lfw.py lfw-deepfunneled lfw_attributes.txt "Silvio Berlusconi" 23 smiling image/lfw_out.jpg -g 0

Output example

person name: "Silvio Berlusconi"
image number: 23

Feature: smiling

|Original|Weight: 0.1|Weight: 0.2|Weight: 0.3|Weight: 0.4|Weight: 0.5| |:---:|:---:|:---:|:---:|:---:|:---:| |||||||

Feature: senior

|Original|Weight: 0.1|Weight: 0.2|Weight: 0.3|Weight: 0.4|Weight: 0.5| |:---:|:---:|:---:|:---:|:---:|:---:| |||||||

To use Large-scale CelebFaces Attributes (CelebA) Dataset

Download dataset

• Download img_align_celeba.zip and list_attr_celeba.txt from CelebA Dataset web site.

• Extract zip file.

Make image list for source and target images

Example:

$ python src/extract_image.py img_align_celeba list_attr_celeba.txt image_normal.txt image_smile.txt smiling young,bags_under_eyes -e eyeglasses,male,pale_skin,narrow_eyes,bushy_eyebrows,chubby,double_chin,bald,bangs,receding_hairline,sideburns,wavy_hair,blond_hair,gray_hair,mouth_slightly_open

Interpolate feature

Example:

$ python src/train.py sample/sample.png image/out/out.png image_normal.txt image_smile.txt -g 0 -c 19,39,159,179

Output example

Feature: smiling

|Original|Weight: 0.1|Weight: 0.2| |:---:|:---:|:---:| ||||

|Weight: 0.3|Weight: 0.4|Weight: 0.5| |:---:|:---:|:---:| ||||

Difference from the original implementation

• feature (x) is not normalized.

• attribute vector w is normalized by w -> ||(x)|| * w / ||w|| (||w|| means L2 norm)

License

MIT License