Hyperoptim#26
Conversation
jdrefs
left a comment
jdrefs
left a comment
There was a problem hiding this comment.
Afaics, this MR provides an extensive implementation to run hyperparameter optimization (HPO) for TVAE on bars test using free energies as HPO objective. As significant parts of the code are tailored to this application, i.e., tailored to a concrete model and a concrete HPO objective (e.g., in neural_models.py, test_neural_models.py, workers.py; also see comments), my suggestion wouId be to make this repository part of the examples . Future refactoring may aim at bundling all methods that are generically applicable to different models and experiments and make them part of tvo.utils.
| from .tvae import GaussianTVAE, BernoulliTVAE | ||
| from .gmm import GMM | ||
| from .pmm import PMM | ||
| from .sssc import SSSC |
| # infer hyperparameter status | ||
| self.is_hyperparameter_S = not (S) | ||
| self.is_hyperparameter_H = not (H) | ||
| self.is_hyperparameter_EEM = False |
There was a problem hiding this comment.
EEM->EVO, also in several places below
| import os | ||
|
|
||
|
|
||
| class ValidFreeEnergy: |
There was a problem hiding this comment.
For a general hyperparameter optimization framework, it may be desirable to provide an interface that allows to implement different optimization metrics (including free energies) as generically as possible.
| return { | ||
| "loss": -train_F if -train_F else np.nan, | ||
| "info": { | ||
| "test accuracy": test_F, | ||
| "train accuracy": train_F, | ||
| "validation accuracy": valid_F, | ||
| "number of parameters": model._external_model.number_of_parameters(), | ||
| }, | ||
| } |
There was a problem hiding this comment.
Do I understand correctly that this implies dependency on a particular model and optimization metric? (similar comment above)
There was a problem hiding this comment.
This is for a particular worker. Individual models require individual workers.
2 participants
Add hyperoptimization capabilities