PhysGNN: learning generative neural network with physics knowledge

Instruction

Run the following command in current directory

include("compile.jl")

⚠️ Julia≧1.3 is required.

⚠️ If it is the first time you run the command above, it might take you some time for downloading and installing dependencies.

Reproducing Numerical Experiments

⚠️ Please disable GPU for the following examples:

This section provides detailed intructions to reproduce the examples in the paper

MNIST Example

(Julia)

cd("MNIST")
include("DCGAN.jl") # DCGAN
include("PhysGNN.jl") # PhysGNN
include("SHGAN1.jl") # Sinkhorn GAN with the penalty parameter = 1.0
include("SHGAN10.jl") # Sinkhorn GAN with the penalty parameter = 10.0
include("SHGAN100.jl") # Sinkhorn GAN with the penalty parameter = 100.0
include("SHGAN10000.jl") # Sinkhorn GAN with the penalty parameter = 10000.0

Neural Network Architectures

(bash)

cd PhysGNN/JumpDiffusion
sh Gauss.sh 

Poisson's Equation

(Julia)

cd("Poisson/compare_gan_and_physgnn")
include("gan.jl") # adversarial training
include("physgnn.jl") # PhysGNN

LinearElasticity

(Julia)

cd("StochasticElasticity")
include("nn.jl") # PhysGNN
tid = 1 # random
include("gs.jl") # Gaussian 

Jump Diffusion

(bash)

cd("JumpDiffusion")
sh Jump.sh 

MCMC