Reproducing Results in Table 1

[alasdairtran]

Hi Zongyi,

I really enjoyed reading your paper! I was having a go at reproducing your numbers in Table 1. I'm wondering if the numbers on the last row come from the test_l2 variable of the final epoch?

If I clone your repo and run python fourier_1d.py, then I have test_l2 as 0.00254360 on epoch 500. In Table 1 of your paper, you report a value of 0.0160 for the FNO model with s = 1024. This seems to be an order of magnitude bigger. Am I looking at the right metric?

[zongyi-li]

Hi Alasdair, in the paper, we used the burgers_R10.mat dataset with modes=16, width=64, and epoch=500. If you get better results probably you are using burgers_v1000.mat dataset, which has smoother initial conditions and therefore easier to model.

[alasdairtran]

No I was using burgers_data_R10.mat from your google drive. I didn't change any code in fourier_1d.py.

Upon closer inspection, I was able to get a test_l2 closer to the number in Table 1 when I reverted back a few commits. I'm guessing that the one-order-of-magnitude improvement is due to the commit where you remove the batch norm. It's interesting that batch norm makes the performance a lot worse here.

[zongyi-li]

That's interesting. We observed the Fourier neural network usually doesn't require the batchnorm. But it's kind of surprising it gets 1 order of magnitude improvement. Thanks for letting me know.

[zongyi-li]

I confirm I get test_l2 = 0.00286 too.