Extend Posterior API to support torch distributions & overhaul MCSampler API (#1486) #193
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
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D39759489 |
saitcakmak
added a commit
to saitcakmak/aepsych
that referenced
this pull request
Nov 16, 2022
…ler API (#1254) Summary: X-link: facebook/Ax#1254 Pull Request resolved: facebookresearch#193 X-link: pytorch/botorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: 14457c86793aa05489c090fe0f969756fbd22bf6
saitcakmak
force-pushed
the
export-D39759489
branch
from
92a03fc
to
17dc1c3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D39759489 |
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 16, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: a851029dde668c954ce6104c0155baf718dfa860
saitcakmak
added a commit
to saitcakmak/Ax
that referenced
this pull request
Nov 16, 2022
…ler API (facebook#1254) Summary: Pull Request resolved: facebook#1254 X-link: facebookresearch/aepsych#193 X-link: pytorch/botorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: e18a1eb5ecd989aa926e646e71eae27f9f685ae4
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 16, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: f1a7d9b390e4012a35f13cfd04ee1114bc12e53e
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 16, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: 59fa663777555ff6d528dab53d124665ae5e75e7
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 16, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. - Some more listed in T134364907 - Test fixes and new units Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: 64ba2cac975fb347cbf2b61cd0e8fc627edc6a6f
…ler API (#1254) Summary: X-link: facebook/Ax#1254 Pull Request resolved: facebookresearch#193 X-link: pytorch/botorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: d6f2d48c1019370f9727acb3fc2652f048e302a0
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 17, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: 325519e3458dcb5bd1f09cf8e71466f5aecda6ae
saitcakmak
force-pushed
the
export-D39759489
branch
from
17dc1c3
to
5e5a0a7
Compare
|
This pull request was exported from Phabricator. Differential Revision: D39759489 |
saitcakmak
added a commit
to saitcakmak/Ax
that referenced
this pull request
Nov 17, 2022
…ler API (facebook#1254) Summary: Pull Request resolved: facebook#1254 X-link: facebookresearch/aepsych#193 X-link: pytorch/botorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: 7c56a0b185333ba8a90b99d14058dba1add4d2de
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 17, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Differential Revision: https://internalfb.com/D39759489 fbshipit-source-id: 99bb2ee02e9e52a8b390f736c4a0668bce8bb09d
saitcakmak
added a commit
to saitcakmak/botorch
that referenced
this pull request
Nov 17, 2022
…ler API (pytorch#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: pytorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Differential Revision: https://internalfb.com/D39759489 fbshipit-source-id: 2669ea7c5996095e2635e8572184b1c22e969d57
facebook-github-bot
pushed a commit
to pytorch/botorch
that referenced
this pull request
Nov 18, 2022
…ler API (#1254) Summary: X-link: facebook/Ax#1254 X-link: facebookresearch/aepsych#193 Pull Request resolved: #1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: f4db866320bab9a5455dfc0c2f7fe2cc15385453
facebook-github-bot
pushed a commit
to facebook/Ax
that referenced
this pull request
Nov 18, 2022
…ler API (#1254) Summary: Pull Request resolved: #1254 X-link: facebookresearch/aepsych#193 X-link: pytorch/botorch#1486 The main goal here is to broadly support non-Gaussian posteriors. - Adds a generic `TorchPosterior` which wraps a Torch `Distribution`. This defines a few properties that we commonly expect, and calls the `distribution` for the rest. - For a unified plotting API, this shifts away from mean & variance to a quantile function. Most torch distributions implement inverse CDF, which is used as quantile. For others, the user should implement it either at distribution or posterior level. - Hands off the burden of base sample handling from the posterior to the samplers. Using a dispatcher based `get_sampler` method, we can support SAA with mixed posteriors without having to shuffle base samples in a `PosteriorList`, as long as all base distributions have a corresponding sampler and support base samples. - Adds `ListSampler` for sampling from `PosteriorList`. - Adds `ForkedRNGSampler` and `StochasticSampler` for sampling from posteriors without base samples. - Adds `rsample_from_base_samples` for sampling with `base_samples` / with a `sampler`. - Absorbs `FullyBayesianPosteriorList` into `PosteriorList`. - For MC acqfs, introduces a `get_posterior_samples` for sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior using `get_sampler`, eliminating the need to construct a sampler in `__init__`, which we used to do under the assumption of Gaussian posteriors. TODOs: - Relax the Gaussian assumption in acquisition functions & utilities. Some of this might be addressed in a follow-up diff. - Updates to website / docs & tutorials to clear up some of the Gaussian assumption, introduce the new relaxed API. Likely a follow-up diff. Other notables: - See D39760855 for usage of TorchDistribution in SkewGP. - TransformedPosterior could serve as the fallback option for derived posteriors. - MC samplers no longer support resample or collapse_batch_dims(=False). These can be handled by i) not using base samples, ii) just using torch.fork_rng and sampling without base samples from that. Samplers are only meant to support SAA. Introduces `ForkedRNGSampler` and `StochasticSampler` as convenience samplers for these use cases. - Introduced `batch_range_override` for the sampler to support edge cases where we may want to override `posterior.batch_range` (needed in `qMultiStepLookahead`) - Removes unused sampling utilities `construct_base_samples(_from_posterior)`, which assume Gaussian posterior. - Moves the main logic of `_set_sampler` method of CachedCholesky subclasses to a `_update_base_samples` method on samplers, and simplifies these classes a bit more. Reviewed By: Balandat Differential Revision: D39759489 fbshipit-source-id: f4db866320bab9a5455dfc0c2f7fe2cc15385453
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.
Summary:
X-link: pytorch/botorch#1486
The main goal here is to broadly support non-Gaussian posteriors.
TorchPosteriorwhich wraps a TorchDistribution. This defines a few properties that we commonly expect, and calls thedistributionfor the rest.get_samplermethod, we can support SAA with mixed posteriors without having to shuffle base samples in aPosteriorList, as long as all base distributions have a corresponding sampler and support base samples.ListSamplerfor sampling fromPosteriorList.ForkedRNGSamplerandStochasticSamplerfor sampling from posteriors without base samples.rsample_from_base_samplesfor sampling withbase_samples/ with asampler.FullyBayesianPosteriorListintoPosteriorList.get_posterior_samplesfor sampling from the posterior with base samples / a sampler. If a sampler was not specified, this constructs the appropriate sampler for the posterior usingget_sampler, eliminating the need to construct a sampler in__init__, which we used to do under the assumption of Gaussian posteriors.TODOs:
Other notables:
ForkedRNGSamplerandStochasticSampleras convenience samplers for these use cases.batch_range_overridefor the sampler to support edge cases where we may want to overrideposterior.batch_range(needed inqMultiStepLookahead)construct_base_samples(_from_posterior), which assume Gaussian posterior._set_samplermethod of CachedCholesky subclasses to a_update_base_samplesmethod on samplers, and simplifies these classes a bit more.Reviewed By: Balandat
Differential Revision: D39759489