AdaIN-Caffe2

A Caffe2 implementation of the paper "Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization".

Requirements

• Python 3
• Caffe2
• numpy
• skimage
• scipy

Usage

python adain_caffe2.py <arguments>

• --vgg_init (-vi) : VGG-19 (normalized) Caffe2 init net PB file
• --vgg_predict (-vp) : VGG-19 (normalized) Caffe2 predict net PB file
• --decoder_init (-di) : Decoder Caffe2 init net PB file
• --decoder_predict (-dp) : Decoder Caffe2 predict net PB file
• --content (-c) : input content image (JPG/PNG)
• --content_size (-cs) <num_pixels> : resize resolution (short side) for the content image (default: 512)
• --style (-s) : input style image (JPG/PNG)
• --style_size (-ss) <num_pixels> : resize resolution (short side) for the style image (default: 512)
• --gpu (-g) : GPU device ID for CUDNN (default: -1, CPU)

Examples

An example image stylized by la_muse.jpg with 1024 for both content and style rescaling resolution.

Acknowledgement

Tested environment

• Ubuntu 16.04
• Anaconda Python 3.6