ma921 / sober Goto Github PK

When running BASQ on a SOBER-optimized model for further calculations, sometimes a cryptic error shows up:

Traceback (most recent call last):
    MAP = basq.MAP(map_samples)
  File "sober/BASQ/_basq.py", line 135, in MAP
    samples = self.sampler.sample(n_samples)
  File "sober/_sampler.py", line 427, in sample
    samples = torch.vstack([samples_wkde, samples_prior])
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 0 but got size 5 for tensor number 1 in the list.

I printed the shapes of samples_wkde and samples_prior (in that order) to look at what happens here. When everything runs well, the output looks like this:

Expected log marginal likelihood: 8.68782e+00
Variance log marginal likelihood: -1.30538e+01
torch.Size([995, 5]) torch.Size([0, 5])
torch.Size([39, 5]) torch.Size([0, 5])

This is when I only run one SOBER iteration. But when I run 10 iterations, afterwards the output is:

Expected log marginal likelihood: 7.21550e+00
Variance log marginal likelihood: -1.73687e+01
torch.Size([618, 5]) torch.Size([0, 5])
torch.Size([0]) torch.Size([0, 5])

Which obviously can't be stacked.

I have two questions. Why is samples_prior always empty? And why is samples_wkde sometimes shaped wrongly?

Settings to reproduce: run SOBER with the following sample sizes for 1 or 10 iterations:

• 100 samples from the prior
• sober.next_batch(10000, 100, 100)
  Then, run BASQ with the following sample sizes:
• basq.quadrature(200, 5, 5)
• basq.sampling_posterior(10)
• basq.MAP(100)

Assume I have a 4D categorical variable ranges in [[0,1], [0,1], [0,1,2,3,4], [0,1,2]], how should I define the categorical prior on this variable and pass into SOBER? As I know, each row in a torch.Tensor must has same dimesions.