Currently we are focused on autoencoder based solution.
This is a GAN-system for picture to picture translation, where pixel values in images correspond particle momentum before entering material and after exiting the material.
The data is sparse with only few pixels activated in an 256x256 grid both in before and after images and this will be reflected in the default options, but the codebase should be flexible enough for applying this for other purposes. Another design goal is a modular structure which allows extending this work with newer designs for networks, loss-functions, visualization, and logging.
The code has now HDF5 input file support. The data in HDF5 file is supposed to be in root of the file in datasets named: testdata, traindata, rundata, or valdata. At the moment only training is implemented in the code set, so only traindata is being used. The datasets are in a following format: array(index,width,height)=32bit float from 0 to 1, meaning grayscale. Each array has before and after -(image)values, stacked either next to each other or above each other, the default is before on the left and after on the right so the width is double of the single image.
This work is done for Muon-Solutions Oy. The company owns all rights to the code and has agreed to publish it under GPL-3 license. Parts of the code are from BicycleGAN and pytorch-CycleGAN-and-pix2pix and those are owned by their authors.
As the codebase grows it will be easy to recognize familirities with the official pytorch implementation of pix2pix and cycleGAN, which is at minimum an inspiration for this code. After some work I realized that BicycleGAN would have been the best starting point, which has the multimodality implemented. The HDF5 loader is based on this posting in Pytorch forum
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent