StyleWood: Training and using GANs to generate microscopic cross-section images of hardwood species

These species don't exist. StyleGAN generated microscopic cross-section images of hardwood species.

This repository was created to provide the original implementation for the paper "Creating high-resolution microscopic cross-section images of hardwood species using generative adversarial networks", written by Dercilio Junior Verly Lopes, Gustavo Fardin Monti, Greg W. Burgreen, Jordão Cabral Moulin, Gabrielly dos Santos Bobadilha, Edward D. Entsminger and Ramon Ferreira Oliveira.

In this repository, you will find the code used for training. Here is a basic summary of the directories in this repository:

• Support/: Supporting files regarding Visual Turing Test analysis.
• collab/: Contains notebook, generator and code for web application. See and use the notebook on Collab.
• dnnlib/: StyleGAN related.
• images/: Images produced for this repository (used in Readme).
• metrics/: Contains FID score calculation
• results/: Contains model's trained snapshot at 7446 kimg selected to show implementation.
• training/: StyleGAN related. Adjusted to use cross-sectional images of hardwoods species.

Due to limited resources using Git Large File Storage, the .pkl files are not available when cloning this repository, and a place-holder file is used instead. To download a copy of the files, click here. This .zip contains two files. The image_generation.pkl file should be pasted to the directory ./stylewood-model-usage/collab/results/. The network-snapshot-007446.pkl file should be pasted to the directory ./stylewood-model-usage/results/00000-sgan-dataset-4gpu.

System requirements

• CentOS Linux 7. Not tested in Windows.
• 64-bit Python 3.7 installation.
• TensorFlow 1.14 with GPU support.
• One or more high-end NVIDIA GPUs with at least 11GB of DRAM. We recommend NVIDIA RTX 2080 Ti GPU.
• NVIDIA driver 411.63, CUDA toolkit 10.0, cuDNN 7.6.4.

How to use?

[step 1.] Prepare dataset

• Make sure to download the TFRecords of the dataset here

[step 2.] Config:

1. Edit train.py with the path of the unzipped file from step 1. Change Line 37.
2. Edit train.py lines 40-43 with the number of GPUS that will be used for training. By default, 4 GPUS will be used.

[step 3.] Training:

1. Run train.py as python train.py.

• Bear in mind that StyleGAN is computationally expensive to train and use.
• Visualization of training progress can be done by TensorBoard.
• Within the results folder, a new directory will be created with training information.
• Training is expected to last days (maybe weeks) depending on computational power.

[step 4.] Image generation:

1a. You can use random_figure.py to generate 100 images using the author's snapshot. If you want to generate a different number of images, change line 33 in the same file.

1b. You can also use your own snapshot generated by your training stage. If that is the case, change line 26 with your own snapshot in random_figure.py.

Don't want to train it, but want to use it?

See our Google Collab notebook here. In the collab you will be able to generate image transitions as the below.

We also prepared a playlist containing video tutorials explaining how to use the Google Collab notebook and the web application. See the playlist on youtube.

You can choose to run the web application locally as well, but you will need to install the pre-requisites.

Here's what we used to run the web application on a Windows Machine:

• Python (We used 3.7).
  • Comes with Anaconda.
• Anaconda (Optional but useful). Open the Anaconda CLI and type
  • conda create --name stylewood python=3.7
  • activate stylewood (Run this every time you need to re-open the CLI)
• cudatoolkit
  • conda install cudatoolkit
    • Requires adding Anaconda Enviroment library bins to path on Windows otherwise missing DLL error may happen
    • The path should be available from the Anaconda Install: Anaconda/envs/envname/Library/bin
    • See how to add a directory to path on Windows.
  • Can also be downloaded from NVidia.
• Tensorflow with GPU support. We used 1.14
  • pip install tensorflow-gpu==1.14
  • 2.0 or greater isn't compatible with StyleGan
• Streamlit. We used v0.82
  • pip install streamlit==0.82
• After installing the packages, run the streamlit application.
  • Go to the directory that contains the Collab files cd C:/user/directory/collab
  • streamlit run webapp.py

Credits and References

For the paper, we trained StyleGAN to generate synthetic images. We have included a copy of the license from the original StyleGAN repository.

• A Style-Based Generator Architecture for Generative Adversarial Networks. Tero Karras (NVIDIA), Samuli Laine (NVIDIA), Timo Aila (NVIDIA). https://arxiv.org/abs/1812.04948. StyleGAN — Official TensorFlow Implementation.

To train StyleGAN, we used a dataset containing identified microscopic cross-section images of hardwoods species by the XDD research team.

• SUGIYAMA, Junji, HWANG, Sung Wook, ZHAI, ShengCheng, KOBAYASHI, Kayoko, KANAI, Izumi, KANAI, Keiko. Xylarium Digital Database for Wood Information Science and Education.

The complete set of references can be found in our paper.