longhoangphi225 / multisample-hypernetworks Goto Github PK

Hi @longhoangphi225,

Thank you for your excellent code! It really helps me a lot!

I have a question regarding the sampling method. In your paper (line 2-5 of Algorithm 1, https://arxiv.org/pdf/2212.01130.pdf),
partition sampling is only used when J (the number of objectives) is large than 2; otherwise the Dir sampling is used.

But it seems that in your code, when J=2, the sampling method is uniform, https://github.com/longhoangphi225/MultiSample-Hypernetworks/blob/main/Toy_Examples/MH_1.ipynb

for i in range(n_mo_sol):
      random = np.random.uniform(start, end)
      #random = np.random.uniform(start + i*(end-start)/head, start+ (i+1)*(end-start)/head)
      ray = np.array([np.cos(random),
                        np.sin(random)], dtype='float32')

while J>2, the sampling method seems to be a combination of Dirichlet and partition sampling? https://github.com/longhoangphi225/MultiSample-Hypernetworks/blob/main/Toy_Examples/MH_4.ipynb

  test_rays = get_reference_directions("das-dennis", 3, n_partitions=5).astype(
        np.float32
    )

for j in range(n_mo_sol):
      ray = np.random.dirichlet([1/3]*3, 1).astype(np.float32)
      ray = ray@test_rays[partition[j]]
      ray /= np.sum(ray)
      ray = ray.astype(np.float32)
      rays.append(torch.from_numpy(
                      ray.flatten()
                  ).to(device))  

Are these typos?

Also, I am wondering whether you could give me some explainations on ray = ray@test_rays[partition[j]], why sampling from dirichlet needs to be multiplied by partition sampling?

Thanks so much in advance!

We see in this block that penalty[0] term in direction_loss will always be removed?
is there a reason? or is it an error?