Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.
And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.
Make DiagLazyTensor work with any batch size
See original GitHub issueCurrently, DiagLazyTensor assumes either a vector or matrix as input. This has the following weird behavior:
> d1 = DiagLazyTensor(torch.ones(4))
> d1.ndimension(), d1.shape
2, (4, 4)
> d2 = DiagLazyTensor(torch.ones(3, 4))
> d2.ndimension(), d2.shape
3, (3, 4, 4)
> d3 = DiagLazyTensor(torch.ones(2, 3, 4))
> d3.ndimension(), d3.shape
2, (4, 4)
Given @gpleiss’s previous work on supporting general batch shapes, I suggest we modify DiagLazyTensor to support input tensors of general shapes. The last dimension will be interpreted as the diagonal, and the previous dimensions as batches. That is, a DiagLazyTensor(torch.ones(2, 3, 4)) would have 4 dimensions and a batch shape of [2, 3].
I need this for implementing Heteroskedastic Likelihoods for batched Multitask GPs.
Issue Analytics
- State:
- Created 5 years ago
- Comments:8 (3 by maintainers)
Top Results From Across the Web
Top Related Medium Post
Top Related StackOverflow Question
Troubleshoot Live Code
Top Related Reddit Thread
Top Related Hackernoon Post
Top Related Tweet
Top Related Dev.to Post
Top Related Hashnode Post
I’m in agreement with this change for sure, a DiagLazyTensor created with a 2 x 3 x 4 tensor should represent a 2 x 3 x 4 x 4 matrix.
Do our other LazyTensors already support multiple batch dimensions already? I remember @gpleiss working on this, but wasn’t 100% sure whether it was fully done.
this was done in #362