Frequently Asked Questions (F.A.Q.)#

  1. Which search algorithm should I use?

  2. What is num_workers?

  3. How to scale the number of parallel evaluations in a Bayesian optimization search?

  4. Why is the number of evaluations returned larger than max_evals?

  5. How do you consider uncertainty in a neural network training?

  6. How to perform checkpointing with Bayesian optimization?

Which search algorithm should I use?#

  • If you come with an existing machine learning pipeline, then it is better to use algorithms such as CBO search (Centralized Bayesian Optimization) which can fine-tune this pipeline.

  • If you come without an existing machine learning pipeline then it is better to use algorithms which provides an existing supervised-learning pipeline.

What is num_workers?#

The num_workers is related to the Evaluator class from the deephyper.evaluator module. It represents the number of “concurrent” evaluations (“concurrent” because not always “parallel”, for example, with a thread-evaluator and an I/O bound run-function). The num_workers does not restrict or isolate the number of resources that each run-function evaluation will use (e.g., the number of GPUs).

Why is the number of evaluations returned larger than max_evals?#

Algorithms provided in DeepHyper are parallel search algorithms. The search can be performed asynchronously (direclty submitting jobs to idle workers) or synchronously (waiting for all workers to finish their jobs before re-submitting new jobs). If evaluator.num_workers == 1 then the number of results will be equal to max_evals (classic sequential optimization loop). However, if `evaluator.num_workers > 1 then batches of size > 1 can be received anytime which makes the number of results grow with a dynamic increment between 1 and evaluator.num_workers. Therefore, if the current number of collected evaluations is 7, 3 new evaluations are received in one batch and `max_evals=8 the search will return 10 results.

How do you consider uncertainty in a neural network training?#

The surrogate model in Bayesian optimization is estimating both aleatoric (data uncertainty, therefore noise of neural network training from different random initialization) and epistemic uncertainty (areas of the search space unknown to the the model). This is done differently for different surrogate models. For example, the Random-Forest estimator uses the law of total variance to estimate these quantities. The Random-Forest surrogate model follows the “random-best” split rule instead of “best” split to estimate better epistemic uncertainty. By default, the same configuration of parameters is not evaluated multiple times with CBO(..., filter_duplicated=True) because empirically it brings better results. Therefore if the user wants to take into consideration the noise from neural networks training it should be CBO(..., filter_duplicated=False).

How to perform checkpointing with Bayesian optimization?#

The CBO(..., log_dir=".") algorithm will save new results in ``{log_dir}/results.csv``(by default to the current directory) each time they are received. Then in case of failure and the search needs to be re-launched it can be done with the following:

search = CBO(..., log_dir="new-log-dir") # to avoid writting on top of previous set of results
search.fit_surrogate("results.csv") # load checkpoint
results = # continue the search