Tricks for training?

[TankManBeta]

Thank you for sharing your idea and source code. I would like to know whether any other tricks were adopted in the training process for the results I got are not as good as the ones reported in the paper.

[sifanexisted]

Hello, did you try to run our provided scripts or to reproduce the results by yourself?

[TankManBeta]

I run the code provided in your repository. Only for the KS equation(regular), I can't get a good result, the L2 error reported in the paper is 3.49e − 04, but I got an error of 4.155e-02.

[sifanexisted]

Apologies for the delayed response. We recently tested our code for that example and successfully reproduced the results on an NVIDIA A6000. However, we did notice a performance drop when using other GPUs. This may be attributed to the example being sensitive to initial values, causing small errors to accumulate and eventually lead to significant discrepancies.

[braveyjc]

I run the code provided in your repository. Only for the KS equation(regular), I can't get a good result, the L2 error reported in the paper is 3.49e − 04, but I got an error of 4.155e-02.

Can you share your code? I'm using Windows system and have been trying to reproduce the causality PINN without the modified MLP and Taylor-mode automatic differentiation. However, I've noticed that the results are not satisfactory.

[TankManBeta]

I run the code provided in your repository. Only for the KS equation(regular), I can't get a good result, the L2 error reported in the paper is 3.49e − 04, but I got an error of 4.155e-02.

Can you share your code? I'm using Windows system and have been trying to reproduce the causality PINN without the modified MLP and Taylor-mode automatic differentiation. However, I've noticed that the results are not satisfactory.

I reproduced the results using the code provided in the repository and didn't modify anything. I did notice that modified MLP and Taylor-mode automatic differentiation can improve accuracy and efficiency.

[braveyjc]

I run the code provided in your repository. Only for the KS equation(regular), I can't get a good result, the L2 error reported in the paper is 3.49e − 04, but I got an error of 4.155e-02.

Can you share your code? I'm using Windows system and have been trying to reproduce the causality PINN without the modified MLP and Taylor-mode automatic differentiation. However, I've noticed that the results are not satisfactory.

I reproduced the results using the code provided in the repository and didn't modify anything. I did notice that modified MLP and Taylor-mode automatic differentiation can improve accuracy and efficiency.

Can we exchange contact information for further discussion? If you have QQ, my QQ number is: 1194453382.