CycleCoopNet

Using Tensorflow to learn image-to-image translation without input-output pairs.

CycleCoopNet method is proposed by Chien-Hao Weng in CycleCoopNet: Image-to-Image Translation with Cooperative Learning Networks.

Prepare environment

In our experient environment, we use windows 10 with NVIDIA GTX1060 6GB GPU, Intel i5-7400 CPU.

install CUDA 9.0

For NVIDIA GPU users, we install CUDA 9.0 for GPU device driver. Your can find it on CUDA Toolkit downloads website.

Download link: https://developer.nvidia.com/cuda-90-download-archive

We are NOT sure the newest version of CUDA can work.

install cuDNN 7.0

cuDNN is part of the NVIDIA Deep Learning SDK. We can install it in the cuDNN developer website. Remember to choose the version that matches your Tensorflow version.

Download link: https://developer.nvidia.com/rdp/cudnn-download

We are also NOT sure the newest version of cuDNN can work.

create new Anaconda environment

Here we use Anaconda to help us manage python environment. You can download it on Anaconda website.

Use Anaconda to create a new python environment, and install all python packages below.

conda create -n [your-environment-name] python=3.6.3
conda activate [your-environment-name]

Or you can use Anaconda GUI to create your own python environment:

install python Packages

tensorflow-gpu 1.13.1
numpy 1.15.4
scipy 1.1.0
pillow 5.3.0
pandas 0.23.4

Getting Started

Prepare the model

Clone this repository:

git clone https://github.com/howarder3/CycleCoopNet

Download training dataset

edges2handbags dataset

bash ./download_dataset_bag.sh

vangogh2photo dataset

bash ./download_dataset_vangogh.sh

Start Training a model

Training a model: (e.g. edges2handbags datasets)

python main.py --dataset_name=edges2handbags

Testing

Put the picture your want to test in the "test_datasets" folder, you can find these results in the "sample" folder. (After you start the training, this folder will auto create by our program.)