I encountered a problem and cannot reproduce this result

[StarrMoonn]

As shown in the figure, I am using MSFWI, but the result of running this ipynb is not the same as the original result. I tested it several times, but I did not find the problem.

V9_Marmousi_MSFWI.ipynb

[omarmohamed15]

Hi @StarrMoonn

Thank you for your interest in our work. Some versions of PyTorch may exhibit instability with the optimizer, which can prevent convergence. Please try using the same version we used: pytorch==1.10.1=py3.9_cuda11.3_cudnn8.2.0_0. Alternatively, you can experiment with a different version of PyTorch. Let me know if you encounter the same issue.

[StarrMoonn]

你好@StarrMoonn

感谢您对我们工作的关注。某些版本的 PyTorch 可能与优化器存在不稳定性，从而阻碍收敛。请尝试使用我们之前使用的相同版本：pytorch==1.10.1=py3.9_cuda11.3_cudnn8.2.0_0。或者，您也可以尝试使用其他版本的 PyTorch。如果您遇到同样的问题，请告知我们。

Hello Omar,

First off, your work on Siamese network FWI is fantastic and has been a huge inspiration. The way you tackled MSFWI crosstalk with FFT was a game-changer for my own project.

I'm now digging into the global correlation norm from the paper "Application of multi-source waveform inversion to marine streamer data using the global correlation norm." I'm trying to figure out how you translated that theory into practice within your DeepWave code.

I'd love to learn more about your approach. Is there any chance you could point me to the relevant sections of the code or share some thoughts on the implementation? Thanks again for sharing your work with the community!

[omarmohamed15]

Hi @StarrMoonn

For SiameseFWI, we did not use global correlation as the loss function. Instead, we applied the Euclidean loss, implemented as follows:

loss =
0.1 * F.pairwise_distance(o1[0, 0, :, :], o2[0, 0, :, :], keepdim=True).mean() +
0.1 * F.pairwise_distance(o1[0, 1, :, :], o2[0, 1, :, :], keepdim=True).mean() +
F.pairwise_distance(o1[0, 2, :, :], o2[0, 2, :, :], keepdim=True).mean()

However, if you prefer to use global correlation, you can define the following function and pass in the desired inputs:

def Xcorr1(x, y):
x = x / torch.norm(x)
y = y / torch.norm(y)
loss = -torch.sum(torch.mul(x, y))
return loss