Restrictions on scatter_add in LDA tutorial

Tutorials
1
  • What tutorial are you running? LDA
  • What version of Pyro are you using? 0.3
  • Please link or paste relevant code, and steps to reproduce.

I believe that there is an error in the LDA code. More precisely, there is a constraint on the input parameters that has not been stated, which is that num-words must be greater than num-words-per-doc.

To see if this is true, run the the code twice:

  1. First run: num-words is 65. The code runs fine.
  2. Second run: num-words is 63. The code crashes.

The problem is with

counts.scatter_add_(0, data[:, ind], torch.tensor(1.).expand(counts.shape))

in the parametrized_guide .There is a restriction on the dimensions of the argument.

My question is: why is there such a restriction? How can I duplicate the functionality of scatter_add without this restriction? Thanks.

2

I found a solution. Simply do the scatter into a larger array and truncate the array. Thanks.

Gordon