Reproduce Synthetic Experiment (paper sec. 4.1)

[leo-du]

Hi,

Thanks for the amazing paper and code! This is not really an issue but I wonder if the authors can share instructions or code on how to reproduce the synthetic experiments section 4.1 (with Sinkhorn iterations). Any help or pointer are appreciated! Thanks!

[xh3204]

I have the same issue. I can‘t repeat the synthetic experiments when I follow the step and parameters as mentioned in the paper. I don't know whether there are some other hyperparameters to define.

[rusty1s]

Hi and thanks for your interest! Sorry, I missed that issue. I haven't got around to clean up and provide the synthetic experiments yet. However, this is the code we used for applying Sinkhorn on the initial and refined assignment matrix:

def sinkhorn(x, x_mask=None, num_steps=0):
    if x_mask is not None:
        x = x.masked_fill(~x_mask, float('-inf'))

    x = torch.softmax(x, dim=-1)

    if x_mask is not None:
        x = x.masked_fill(~x_mask, 0)

    for _ in range(num_steps):
        x = F.normalize(x, p=1, dim=-2)
        x = F.normalize(x, p=1, dim=-1)

    if x_mask is not None:
        x = x.masked_fill(~x_mask, 0)

    return x

I will let you know once synthetic experiments are included in the repository.

[leo-du]

Hi, I have a follow up question: I saw that your sinkhorn function has a num_steps parameter. Does that mean you run the Sinkhorn iteration for a fixed number of iterations, or do you run it until convergence? Thanks!

[rusty1s]

We run sinkhorn for a very large number of fixed iterations.

[leo-du]

Can you give me an order of magnitude if you still have it? Exact number would be better. Thanks!

[rusty1s]

We run sinkhorn both for 100 and 1000 iterations and they both perform equally well.