Adds support for energy based learning with NLL loss (LEO) #30
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
facebook-github-bot
added
the
CLA Signed
mhmukadam
reviewed
Dec 17, 2021
mhmukadam
reviewed
Dec 17, 2021
mhmukadam
reviewed
Dec 17, 2021
mhmukadam
reviewed
Dec 17, 2021
mhmukadam
reviewed
Dec 17, 2021
mhmukadam
reviewed
Dec 18, 2021
mhmukadam
reviewed
Dec 18, 2021
mhmukadam
reviewed
Dec 18, 2021
mhmukadam
reviewed
Dec 18, 2021
luisenp
requested changes
Dec 20, 2021
There was a problem hiding this comment.
Great job on this PR! I see all the functionality is there, but I propose we shuffle some things around a bit. I'm OK leaving that as a separate PR, but we should definitely make issues for the comments so we don't forget.
Regarding the broken test, upper= kwarg for cholesky was only introduced in torch 1.10.0. It used to be in torch.cholesky, but it was removed in torch.linalg.cholesky in v1.9.0 (see screenshot).
For now, can you update the call to match the one used in v1.9.0 so that the test passes?
mhmukadam
reviewed
Dec 23, 2021
mhmukadam
reviewed
Dec 23, 2021
mhmukadam
approved these changes
Dec 23, 2021
mhmukadam
approved these changes
Dec 24, 2021
suddhu
pushed a commit
to suddhu/theseus
that referenced
this pull request
Jan 21, 2023
…esearch#30) * add tests for leo with GN/LM optimizers * add sampler to GN/LM optimizers * run leo on 2d state estimation, add viz, learning_method options
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Outdated suggestions cannot be applied.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.
This PR introduces LEO, learning energy-based models in optimization (https://arxiv.org/abs/2108.02274). LEO is a method to learn models end-to-end within second-order optimizers like Gauss-Newton. The main difference is that instead of unrolling the optimizer and minimizing the MSE tracking loss, this introduces the NLL energy-based loss that does not backpropagate through the optimizer. It requires low-energy samples from the optimizer, pushing up the energy of optimizer samples and pushing down the energy of ground truth samples.
To execute it, run
python examples/state_estimation_2d.pywithlearning_method="leo"This should update the learnable cost weights so that the optimizer trajectory (orange) matches the ground truth trajectory (green).