Summary: 111 instances, 96 unique

Text	Count
# TODO: technically we should make sure that we add a	1
# TODO: Figure out attribute namming weirdness here	1
# TODO: Consider returning an MVN here instead	1
# TODO: improve efficiency for multi-task models	1
# TODO: filter out points that are worse than the reference point first here	1
# TODO (T52818288): Properly use lazy tensors in scalarize_posterior	1
TODO: implement support using independent warping functions	1
# TODO: Allow qMC sampling	2
# TODO: make this be the default KPMatmulLT diagonal method in gpytorch	2
# Use the mean of the previous noise values (TODO: be smarter here).	1
# TODO: Make sure observation noise is transformed correctly	1
# TODO swap out scatter_ so that comparisons could be int instead of long	1
# TODO: add support for outcome transforms.	1
# TODO: Refactor these as PosteriorTransform as well.	1
# TODO: Don't instantiate a generator	1
# TODO: Allow passing in inner samplers directly	1
# TODO: Avoid unnecessary computation by not generating all candidates	1
# TODO: give the Posterior API an add_observation_noise function to avoid	1
# TODO: ensure this works in batch mode, which it does not currently.	1
# TODO: support splitting outcome transforms.	1
# TODO: return a HigherOrderGPPosterior once rescaling constant	1
# TODO: use gpytorch's pivoted cholesky instead once that gets an exposed list	1
# TODO: enforce the diagonalization to return a KPLT for all shapes in	1
# TODO: cleanup the reshaping here	1
# TODO: Consider different aggregation (i.e. max) across q-batch	1
# TODO: Dedup handling of this here and in the constructor (maybe via a	2
# TODO: Investigate differentiability of MVaR.	1
TODO: Support `m > 2` objectives.	1
TODO: write this in C++ for faster looping.	1
# TODO: Add support for outcome transforms.	1
performs the decomposition under minimization. TODO: use maximization	1
TODO: Use the observed values to identify the fantasy sub-tree that is closest to	1
# TODO: expand options for the mean module via batch shaping?	1
# TODO: ensure that this still works for structured noise solves.	1
# @TODO make sure fidelity_dims align in project, expand & cost_aware_utility	1
TODO: Add support for outcome constraints.	1
# TODO: Implement best point computation from training data	1
TODO: Investigate further.	1
# TODO: Find a general way to do this efficiently.	1
TODO: add support for utilizing gradient information	1
# TODO: clean this up after removing AcquisitionObjective.	2
# TODO: remove these variables from `state_dict()` so that when calling	1
# TODO: improve efficiency by not recomputing baseline-baseline	1
# TODO: Add support for HeteroskedasticSingleTaskGP.	3
# TODO: we could use batches to compute (q choose i) and (q choose q-i)	1
# GPyTorchPosterior (TODO: Should we Lazy-evaluate the mean here as well?)	1
# TODO: Allow subsetting of other covar modules	2
# TODO: make this a flag?	1
# TODO: Deduplicate repeated evaluations / deal with numerical degeneracies	1
# Use the mean of the previous noise values (TODO: be smarter here).	1
# TODO: Can we enable backprop through the latent covariances?	1
# TODO (T41270962): Support task-specific noise levels in likelihood	1
# TODO: Clean up once ScalarizedObjective is removed.	1
# TODO: Add prune_baseline functionality as for qNEI	1
TODO: make this method return the sampler object, to avoid doing burn-in	1
TODO: remove this when MultiTask models support outcome transforms.	2
# TODO: Implement more efficient way to compute posterior over both training and	1
# TODO: speed computation of covariance matrix	1
For now uses the same q' as in `full_optimizer`. TODO: allow different `q`.	1
# TODO: Add support for custom likelihoods.	2
# TODO: Make sure this doesn't change base samples in-place	1
# TODO: Delta method, possibly issue warning	4
# TODO: update to follow new gpytorch convention resulting from	1
# TODO: Use sparse representation (not sure if scipy optim supports that)	1
# TODO: when batched kernels are supported in RandomFourierFeatures,	1
TODO: similar to qNEHVI, when we are using sequential greedy candidate	1
# TODO: we could refactor this __init__ logic into a	1
# TODO: support batched inputs (req. dealing with ragged tensors)	2
# TODO: Can we get the model batch shape property from the model?	1
# TODO: Handle approx. zero losses (normalize by min/max loss range)	1
# TODO: remove this once torch.cholesky issue is resolved	1
# TODO: Validate noise shape	1
# TODO: make D a sparse matrix once pytorch has better support for	1
# TODO: get rid of this once cholesky_inverse supports batch mode	1
# TODO: support multiple batch dimensions here	1
# TODO: remove the exception handling, when the pytorch	1
# TODO: enable vectorization/parallelization here	1
# TODO: Allow broadcasting of model batch shapes	1
# TODO: make sure there are not deduplicate constraints	1
# TODO: Do not instantiate full covariance for lazy tensors (ideally we simplify	1
# TODO: use the exposed joint covariances from the prediction strategy	1
# TODO: Find a better way to do this	1
# ensure scalars agree (TODO: Allow different priors for different outputs)	1
TODO: Use DKLV17 and LKF17 for the box decomposition as in [Yang2019]_ for	1
# TODO: Update the exp moving average efficiently	1
# TODO: be smart about how we can update covar matrix here	1
TODO: refactor some/all of this into the MCSampler.	2
# TODO: support batched inputs (req. dealing with ragged tensors)	2
# TODO: Ideally support RFFs for multi-outputs instead of having to	1
# TODO: We can support fixed features, see Max's comment on D33551393. We can	1
"composite_mtbo.ipynb",  # TODO: very slow, figure out if we can make it faster	1
new_weights = getattr(sampler, "base_weights", None)  # TODO: generalize this	1
TODO: Support t-batches of initial conditions.	1
# TODO: use batched `torch.randperm` when available:	1
# TODO: make minimum value dtype-dependent	1
TODO: replace this with a more efficient decomposition. E.g.	1