HI, I have a question. When we trained the same FourCastNet model using both the Modulus framework and the training code available on the official GitHub repository, we observed a significant difference: the version trained with Modulus outperformed the one trained with the GitHub-provided code. This discrepancy raises questions about the underlying reasons for such a pronounced difference in performance.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Modulus vs modulus-sym & Post-processing (using the model after training) | 1 | 896 | May 24, 2023 | |
| Just Released: NVIDIA Modulus 23.11 | 0 | 330 | November 21, 2023 | |
| Just Released: NVIDIA Modulus 23.11 | 0 | 424 | November 30, 2023 | |
| Nvidia Modulus: failed to run cuBLAS routine: CUBLAS_STATUS_EXECUTION_FAILED | 2 | 1815 | May 18, 2022 | |
| Unable to load 'arch/fourier_net' | 8 | 1386 | February 27, 2024 | |
| Custom Pytorch with modulus | 0 | 71 | December 20, 2024 | |
| Develop Physics-Informed Machine Learning Models with Graph Neural Networks | 2 | 493 | June 27, 2023 | |
| Can I use RTX8000? | 2 | 970 | December 21, 2022 | |
| Modulus prediction based on the trained model | 8 | 1785 | October 13, 2022 | |
| Enhancing Digital Twin Models and Simulations with NVIDIA Modulus v22.09 | 0 | 443 | September 21, 2022 |