Awesome

SMIT: Stochastic Multi-Label Image-to-image Translation

This repository provides a PyTorch implementation of SMIT. SMIT can stochastically translate an input image to multiple domains using only a single generator and a discriminator. It only needs a target domain (binary vector e.g., [0,1,0,1,1] for 5 different domains) and a random gaussian noise. <br/>

Paper

SMIT: Stochastic Multi-Label Image-to-image Translation <br/> Andrés Romero¹, Pablo Arbelaez¹, Luc Van Gool², Radu Timofte² <br/> ¹Biomedical Computer Vision (BCV) Lab, Universidad de Los Andes. <br/> ²Computer Vision Lab (CVL), ETH Zürich. <br/> <br/>

Citation

@article{romero2019smit,
  title={SMIT: Stochastic Multi-Label Image-to-Image Translation},
  author={Romero, Andr{\'e}s and Arbel{\'a}ez, Pablo and Van Gool, Luc and Timofte, Radu},
  journal={ICCV Workshops},
  year={2019}
}

Dependencies

• Python (2.7, 3.5+)
• PyTorch (0.3, 0.4, 1.0) <br/>

Usage

Cloning the repository

$ git clone https://github.com/BCV-Uniandes/SMIT.git
$ cd SMIT

Downloading the dataset

To download the CelebA dataset:

$ bash generate_data/download.sh

Train command:

./main.py --GPU=$gpu_id --dataset_fake=CelebA

Each dataset must has datasets/<dataset>.py and datasets/<dataset>.yaml files. All models and figures will be stored at snapshot/models/$dataset_fake/<epoch>_<iter>.pth and snapshot/samples/$dataset_fake/<epoch>_<iter>.jpg, respectivelly.

Test command:

./main.py --GPU=$gpu_id --dataset_fake=CelebA --mode=test

SMIT will expect the .pth weights are stored at snapshot/models/$dataset_fake/ (or --pretrained_model=location/model.pth should be provided). If there are several models, it will take the last alphabetical one.

Demo:

./main.py --GPU=$gpu_id --dataset_fake=CelebA --mode=test --DEMO_PATH=location/image_jpg/or/location/dir

DEMO performs transformation per attribute, that is swapping attributes with respect to the original input as in the images below. Therefore, --DEMO_LABEL is provided for the real attribute if DEMO_PATH is an image (If it is not provided, the discriminator acts as classifier for the real attributes).

Pretrained models

Models trained using Pytorch 1.0.

Multi-GPU

For multiple GPUs we use Horovod. Example for training with 4 GPUs:

mpirun -n 4 ./main.py --dataset_fake=CelebA

Qualitative Results. Multi-Domain Continuous Interpolation.

First column (original input) -> Last column (Opposite attributes: smile, age, genre, sunglasses, bangs, color hair). Up: Continuous interpolation for the fake image. Down: Continuous interpolation for the attention mechanism.

Qualitative Results. Random sampling.

CelebA

EmotionNet

RafD

Edges2Shoes

Edges2Handbags

Yosemite

Painters

Qualitative Results. Style Interpolation between first and last row.

CelebA

EmotionNet

RafD

Edges2Shoes

Edges2Handbags

Yosemite

Painters

Qualitative Results. Label continuous inference between first and last row.

CelebA

EmotionNet