Hyperparameter Search

In machine learning, a common task is attempting to find good hyperparameters for a learning algorithm. PEDL provides support for hyperparameter search as a first-class workflow. Several hyperparameter search algorithms are supported: single, random, grid, adaptive_simple, adaptive, and pbt.

Here we give a brief introduction to the search methods. The fields mentioned in the descriptions below are specified under the searcher field in the experiment configuration file.

In addition to the parameters that are specific to individual search algorithms, the following parameters may be provided for any algorithm.

  • metric: The name of the validation metric to use when comparing trials.
  • (optional) smaller_is_better: Whether a smaller value of metric is considered better performance (e.g., this would be true for a loss metric and false for an accuracy metric). Defaults to true.
  • (optional) source_trial_id and source_checkpoint_uuid: Only one of these can be specified at a time. Initializes weights of all trials to some prior checkpoint. If not specified, model weights are initialized randomly.

Single

The single search method does not technically "search": it trains a single hyperparameter configuration for max_steps steps, and then performs validation. This method is useful for testing or for training a single model configuration until convergence.

Random

The random search method generates max_trials trials with hyperparameters chosen uniformly at random from the configured hyperparameter space. Each trial is trained for max_steps steps and then the trial's validation metrics are computed.

Grid

The grid search method generates trials on a "grid" of hyperparameter configurations and trains each trial for max_steps steps. The user specifies a set of values for each hyperparameter via the hyperparameters field in the experiment config file. The "grid" of hyperparameter configurations is generated by taking the product of these sets. For example, if the set of values for three separate hyperparameters aparam, bparam, and cparam are specified as {0, 1, 2}, {10, 20}, and {"c"} respectively, then the grid of tuples (aparam, bparam, cparam) generated is:

(0, 10, "c")
(0, 20, "c")
(1, 10, "c")
(1, 20, "c")
(2, 10, "c")
(2, 20, "c")

The way the set of hyperparameter values is specified depends on the type of hyperparameter:

  • const: The set of values contains just the single value. For example, cparam above could be specified as a const hyperparameter with val: c.
  • categorical: The set of values is exactly the set of categorical values. For example, bparam above could be specified as a categorical hyperparameter with vals: [10, 20].
  • int: The set of count values is taken evenly from the range [minval, maxval], inclusive of endpoints. If count is larger than the number of integer values in the range, that is interpreted as the entire range of integers in [minval, maxval]. For example, aparam above could be specified as an int hyperparameter with minval: 0, maxval: 2, and count: 3 or count: 100.
  • double: The set of count values is taken evenly from the range [minval, maxval], inclusive of endpoints. The set {0.1, 0.3, 0.5} could be specified as a double hyperparameter with minval: 0.1, maxval: 0.5, count: 3.
  • log: The set of count values is taken logarithmically evenly from the range [baseminval, basemaxval], inclusive of endpoints. For example, the set {0.00001, 0.0001, 0.001} could be specified as a log hyperparameter with base: 10, minval: -5, maxval: -3, and count: 3.

Under the special case of count: 1 for int, double, or log, the midpoint (with rounding for int and with basemidpoint for log) is returned.

Adaptive (Simple)

The adaptive_simple search method is a theoretically principled and empirically state-of-the-art method that adaptively allocates resources to promising hyperparameter configurations while quickly eliminating poor ones. There are two interfaces to this search algorithm: the adaptive_simple method is easier to configure and provides good defaults for most situations, whereas the adaptive search method (described below) allows advanced users to have more fine-grained control over the behavior of the search.

The adaptive_simple search method takes two configuration settings:

  • max_steps: The maximum number of steps that any trial that survives to the end of the experiment will be trained for (a step is a fixed number of batches). This quantity is domain-specific and should roughly reflect the number of training steps needed for the model to converge on the data set. For users who would like to determine this number experimentally, train a model with reasonable hyperparameters using the single search method.
  • max_trials: The maximum number of hyperparameter configurations that will be explored. Most of these configurations will not be trained to convergence; rather, the search method will use early-stopping to prune hyperparameter configurations that are not performing well.

That is, max_steps is a property of the model itself (how long the model must be trained until convergence), whereas max_trials controls how many resources the user would like the search to consume.

Adaptive (Advanced)

The adaptive search method employs the same underlying algorithm as the adaptive_simple method described above, but it allows users to control the behavior of the search in a more fine-grained way, at the cost of being more difficult to configure. This section explains the configuration settings that influence the behavior of the adaptive searcher and gives recommendations for how to configure those settings.

Quick start

Here are some suggested initial settings for adaptive that typically work well.

Search mode:

  • mode: Set to standard.

Resource budget:

  • target_trial_steps: The maximum number of steps that any trial that survives to the end of the experiment will be trained for (a step is a fixed number of batches). This quantity is domain-specific and should roughly reflect the number of training steps needed for the model to converge on the data set. For users who would like to determined this number experimentally, train a model with reasonable hyperparameters using the single search method.
  • step_budget: Set step_budget to roughly 10 times target_trial_steps. A higher step_budget will result in hyperparameter search that consumes more resources and takes longer to complete.

Details

Conceptually, the adaptive searcher is a carefully tuned strategy for spawning multiple ASHA (asynchronous successive halving algorithm) searchers, themselves hyperparameter search algorithms. ASHA can be configured to make different tradeoffs between exploration and exploitation, i.e., how many trials are explored versus how long a single trial is trained for. Because the right tradeoff between exploration and exploitation is hard to know in advance, the adaptive algorithm tries several ASHA searches with different tradeoffs.

The configuration settings available to PEDL users running in adaptive mode mostly affect the ASHA subroutines directly. The mode configuration is the only one affecting the decisions of the adaptive searcher, by changing the number and types of ASHA subroutines spawned.

The first section here gives a description of ASHA. The second section describes the configuration parameters that influence how this search method behaves. The third section gives a summary of the adaptive configuration settings.

ASHA

At a high level, ASHA prunes ("halves") a set of trials in successive rounds we call rungs. ASHA starts with an initial set of trials. (A trial means one model for training, with a fixed set of hyperparameter values.) ASHA trains all the trials for some number of steps and the trials with the worst validation performance are discarded. In the next rung, the remaining trials are trained for a longer period of time, and then trials with the worst validation performance are pruned once again. This is repeated until the maximum number of training steps is reached.

First, an example of ASHA.

  • Rung 1: ASHA creates N initial trials; the hyperparameter values for each trial are randomly sampled from the hyperparameters defined in the experiment configuration file. Each trial is trained for 3 steps, and then validation metrics are computed.
  • Rung 2: ASHA picks the N/4 top-performing trials according to validation metrics. These are trained for 12 steps.
  • Rung 3: ASHA picks the N/16 top-performing trials according to validation metrics. These are trained for 48 steps.

At the end, the trial with best performance has the hyperparameter setting the ASHA searcher returns.

In the example above, divisor is 4, which determines what fraction of trials are kept in successive rungs, as well as the number of steps in successive rungs. target_trial_steps is 48, which is the maximum number of steps a trial is trained for.

The remaining degree of freedom in this ASHA example is the number N of trials initialized. This is determined by the top-level adaptive algorithm, through step_budget and the number/types of ASHA subroutines called.

In general, ASHA has a fixed divisor d. In the first rung, it generates an initial set of randomly chosen trials and runs until each trial has completed the same number of steps. In the next rung, it keeps 1/d of those trials and closes the rest. Then it runs each remaining trial until it has completed d times as many steps as after the previous rung. ASHA iterates this process until some stopping criterion is reached, such as completing a specified number of rungs or having only one trial remaining. The number of steps, rungs, and trials within rungs are fixed within each ASHA searcher, but vary across different calls to ASHA by the adaptive algorithm. Note that although the name "ASHA" includes the phrase "halving", the fraction of trials pruned after every rung is controlled by divisor.

Adaptive over ASHA

The adaptive algorithm calls ASHA subroutines with varying parameters. The exact calls are configured through the choice of mode, which specifies how aggressively to perform early stopping. One way to think about this behavior is as a spectrum that ranges from "one ASHA run" (aggressive early stopping; eliminate most trials every rung) to "searcher: random" (no early stopping; all initialized trials are allowed to run to completion).

On one end, aggressive applies early stopping in a very eager manner; this mode essentially corresponds to only making a single call to ASHA. With the default divisor of 4, 75% of the remaining trials will be eliminated in each rung after only being trained for 25% as many training steps as will be performed in the next rung. This implies that relatively few of the trials will be allowed to finish even a small fraction of the training steps needed for a full training run (target_trial_steps). This aggressive early stopping behavior allows the searcher to start more trials for a wider exploration of hyperparameter configurations, at the risk of discarding a configuration too soon.

On the other end, conservative mode is more similar to a random search, in that it performs significantly less pruning. Extra ASHA subroutines are spawned with fewer rungs and larger training steps to account for the high percentage of trials eliminated after only a few steps. However, a conservative adaptive search will only explore a small fraction of the configurations explored by an aggressive search, given the same step budget.

Once the number and types of calls to ASHA are determined (via mode), the adaptive algorithm will allocate budgets of steps to the ASHA subroutines, from the overall step_budget for the adaptive algorithm (user-specified through step_budget). This determines the number of trials at each rung (N in the above ASHA example).

Configuration

Users specify configurations for the adaptive searcher through the experiment configuration file. They fall into two categories described below.

Parameters for ASHA:

  • target_trial_steps: The maximum number of steps that any one trial will be trained.
  • (optional, for advanced users only) divisor: The multiplier for eliminating trials and increasing steps trained at each rung. The default is 4.
  • (optional, for advanced users only) max_rungs: The maximum number of rungs. The default is 5.

Parameters for adaptive mode:

  • mode: Options are aggressive, standard, or conservative. Specifies how aggressively to perform early stopping. We suggest using either aggressive or standard mode.
  • step_budget: A budget for total steps taken across all trials and ASHA calls. The budget is split evenly between ASHA calls. The recommendation above was to set step_budget = 10 * target_trial_steps.

Examples

The table below illustrates the difference between aggressive, standard, and conservative for an otherwise fixed configuration. While aggressive tries out 64 hyperparameter configurations, conservative tries only 31 hyperparameter configurations but has the budget to run more of them to the full 16 steps. More ASHA instances are generated by conservative, which are responsible for creating the trials run for the full 16 steps.

The settings are divisor: 4, max_rungs: 3, target_trial_steps: 16, and step_budget: 160.

Total steps trained Number of trials
64 43 31
aggressive standard conservative
ASHA0 ASHA0 ASHA1 ASHA0 ASHA1 ASHA2
1 48 23 14
4 11 7 7 5 5
16 5 2 4 2 2 3

For an experiment generated by a specific .yaml experiment configuration file, this information (ASHA instances and number of trials vs. number of steps) can be found with the command

pedl preview-search <file_name.yaml>

FAQ

Q: How do I control how many batches a trial is trained for before it is potentially discarded?

Two factors affect the number of batches a trial is guaranteed to be trained on. The field batches_per_step affects how many batches is considered one step. The number of steps guaranteed is affected by target_trial_steps, and is at least target_trial_steps / 256 by default, or target_trial_steps / divisormax_rungs-1 in general.

Q: How do I set the initial number of trials? How do I make sure x trials are run the full target_trial_steps steps?

The number of initial trials is determined by a combination of mode, step_budget, divisor, max_rungs, and target_trial_steps. Here is a rule of thumb for the default configuration of max_rungs: 5 and divisor: 4, with mode: aggressive and a large enough step_budget:

  • The initial number of trials is step_budget / (4 * target_trial_steps).
  • To ensure that x trials are run target_trial_steps, set step_budget to be 4 * x * target_trial_steps.

A configuration setting that meets set goals can also be found by trial and error. The command

pedl preview-search <file_name.yaml>
will display information on the number of trials versus number of steps for the configuration specified in file_name.yaml. Increasing step_budget increases both the initial number of trials and the number of trials run the full number of steps. On the other hand, target_trial_steps decreases both. The mode decides on allocation of steps between trials; mode: conservative runs more trials for longer, whereas mode: aggressive eliminates the most trials early in training.

Q: The adaptive algorithm sounds great so far. What are its weaknesses?

One downside of adaptive is that it results in doing more validations, which might be expensive.

Population-based training

Population-based training (PBT) is loosely based on genetic algorithms; see the original paper or blog post for details. The motivation is that it makes sense to explore hyperparameter configurations that are known to perform well, since the performance of a model as a function of the hyperparameters is likely to show some continuity. The algorithm works by repeatedly replacing low-performing hyperparameter configurations with modified versions of high-performing ones.

Quick start

A typical set of configuration values for PBT:

  • population_size: 40
  • num_rounds, steps_per_round: The product of these values is the total number of steps that a trial that survives to the end of the experiment will be trained for; it should be chosen similarly to the value of target_trial_steps for adaptive search. For a given value of the product, decreasing steps_per_round creates more opportunity for evaluation and selection of good configurations at the cost of higher variance and computational overhead.
  • replace_function:
    • truncate_fraction: 0.2
  • explore_function:
    • resample_probability: 0.2
    • perturb_factor: 0.2

Details

At any time, the searcher maintains a fixed number of active trials (the population). Initially, each trial uses a randomly chosen hyperparameter configuration, just as with the random searcher. The difference is that, periodically, every trial stops training and evaluates the validation metric for the trial's current state; some of the worst-performing trials are closed, while an equal number of the best-performing trials are cloned to replace them. Cloning a trial involves checkpointing it and creating a new trial that continues training from that checkpoint. The hyperparameters of the new trial are not generally equal to those of the original trial, but are derived from them in a particular way; see the description of available parameters for details.

There is an important constraint on the hyperparameters that are allowed to vary when PBT is in use: it must always be possible to load a checkpoint from a model that was created with any potential hyperparameter configuration into a model using any other configuration; otherwise, the cloning process could fail. This means that, for instance, the number of hidden units in a neural network layer cannot be such a hyperparameter. If it were, the models for different configurations could have weight matrices of different dimensions, so their checkpoints would not be compatible.

Parameters

One round consists of a period of training followed by a validate/close/clone phase. During each round, each running trial does a fixed amount of training, determined by the experiment configuration.

  • population_size: The number of trials that should run at the same time.
  • num_rounds: The total number of rounds to run.
  • steps_per_round: The number of training steps for each trial to run during each round.

The parameters for the cloning process are also configurable using two nested objects, called replace_function and explore_function, within the searcher fields of the experiment configuration file.

  • replace_function: The configuration for deciding which trials to close.
    • truncate_fraction: The fraction of the population that is closed and replaced by clones at the end of each round.
  • explore_function: The configuration for modifying hyperparameter configurations when cloning. Each hyperparameter is either resampled, meaning that it is replaced by a value drawn independently from the original configuration, or perturbed, meaning that it is multiplied by a configurable factor.
    • resample_probability: The probability that a hyperparameter is replaced with a new value sampled from the original distribution specified in the configuration.
    • perturb_factor: The amount by which hyperparameters that are not resampled are perturbed: each numerical hyperparameter is multiplied by either 1 + perturb_factor or 1 - perturb_factor with equal probability; categorical and const hyperparameters are left unchanged.