/DLA-VPS_3dGAN

Primary LanguagePythonMIT LicenseMIT

DLA-VPS (3D GAN based model)

Introuction

The 3D GAN model which reconstruct Nyx simulation data from the simulation parameter space.

Prerequisites

  • Anaconda 4.10.1
  • python 3.7.11
  • tensorflow-gpu 2.3
  • cuda 10.1
  • cudnn 7.6.5
  • tensorflow_addons-0.15.0

Pipeline

Data

  • We used the data Nyx which is devloped by Lawrence Berkeley National Laboratory.
  • Simulaiotn Info:
    • Input parameters: Omega_M: [0.17, 0.5] , Omega_B: [0.03, 0.08] and hubble: [0.55, 0.85].
    • Output quantities: density, temperature, rho_e, phi_grav, x-momentum, y-momentum and z-momentum.
    • Resolution: 32^3 to 8192^3. (64^3 in our case.)
  • We save the simulation data as TFRecord formate. Click here to download the folder Nyx_tfrecords.

Training

  • Step1: Modify the Nyx_Reconstruction/config/utlis/config.py of the training. The settings that are most needed to be changed are as follows:

    • variable : the quantity choosing. (density, temp, rho_e, phi_grav, xmom. ymom and zmom)
    • dataSetDir: the directory of the tfRecords
    • logDir : the directory to save the trained models (generator and discriminator)
    • epochs : the number of the training epochs
    • batchSize : the batch size during training
    • save_weights_only: save wieghts only or save model
    • save_epochs: how many epochs to save the trained model once

  • Step2: Run main.py to start the training. The trained models will be saved under the logDir/gen and logDir/dis with respective to the generator and discriminator during training. We highly recommend you to traine the models on GPU instead of CPU.

  • Step3: Buy a coffee and have patience!

Architecture

Generator

generator

Discriminator

discriminator

Ressidual block

residual_block

Notes

  • The details of the model architecture and training process can be found in the theses DLA-VPS.
  • Anyone is welcome to let us know how to improve in any way.