search

PredictiveIntelligenceLab / CausalPINNs

Other
138 stars 29 forks source link

Tricks for training? #6

Open TankManBeta opened 1 year ago

TankManBeta commented 1 year ago

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 commented 1 year ago

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

TankManBeta commented 1 year ago

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 commented 1 year ago

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 commented 1 year ago

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 commented 1 year ago

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 commented 1 year ago

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.