Fix bug in preconditioned KISS-GP / Hadamard Multitask GPs #2090
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Previously in PyTorch, calling on a int/long matrix was a no-op. At some point in the PyTorch releases, this line started throwing an error (since only floating point operations can have gradients). Our implementation of pivoted cholesky previously called `matrix_arg.requires_grad_(True)` on all LazyTensor/LinearOperator arguments, without first checking whether `matrix_arg` was a floating point dtype. KISS-GP makes use of InterpolatedLazyTensor (to become InterpolatedLinearOperator), which has integer matrices (the index matrices for interpolation). This therefore produced an error message when it was used in conjunction with the pivoted cholesky preconditioner. A similar bug exists for preconditioned Hadamard Multitask GPs. (The reason this bug went undetected is because our tests for KISS-GP models and multitask models all use small datasets (N < 100). Preconditioners are not used until N > 2000 or so.) [Fixes #2056]
Balandat
approved these changes
Aug 7, 2022
| @@ -58,8 +58,8 @@ | |||
| "metadata": {}, | |||
| "outputs": [], | |||
| "source": [ | |||
| "train_x1 = torch.rand(50)\n", | |||
| "train_x2 = torch.rand(50)\n", | |||
| "train_x1 = torch.rand(2000)\n", | |||
There was a problem hiding this comment.
By how much does this increase runtime? Should we be concerned here if this is run as part of the unit tests?
There was a problem hiding this comment.
This change (plus the ones in the other two examples) add at most one second to the test suite - I just checked.
| @@ -98,7 +98,11 @@ def backward(ctx, grad_output, _): | |||
|
|
|||
| with torch.enable_grad(): | |||
| # Create a new set of matrix args that we can backpropagate through | |||
| matrix_args = [matrix_arg.detach().requires_grad_(True) for matrix_arg in _matrix_args] | |||
| matrix_args = [] | |||
There was a problem hiding this comment.
I would say we could also catch the error so that if there are any future changes we don't have to adjust the code here. But then raising errors is unfortunately quite slow in ptorch...
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Previously in PyTorch, calling on a int/long matrix was a no-op.
At some point in the PyTorch releases, this line started throwing an error (since only
floating point operations can have gradients).
Our implementation of pivoted cholesky previously called
matrix_arg.requires_grad_(True)on all LazyTensor/LinearOperatorarguments, without first checking whether
matrix_argwas a floatingpoint dtype.
KISS-GP makes use of InterpolatedLazyTensor (to become InterpolatedLinearOperator), which
has integer matrices (the index matrices for interpolation). This
therefore produced an error message when it was used in conjunction with
the pivoted cholesky preconditioner. A similar bug exists for
preconditioned Hadamard Multitask GPs.
(The reason this bug went undetected is because our tests for KISS-GP
models and multitask models all use small datasets (N < 100).
Preconditioners are not used until N > 2000 or so.)
[Fixes #2056]