deephyper.stopper.LCModelStopper

class deephyper.stopper.LCModelStopper(max_steps: int, min_steps: int = 1, lc_model='mmf4', min_obs_to_fit_lc_model=4, min_done_for_outlier_detection=10, iqr_factor_for_outlier_detection=1.5, prob_promotion=0.9, early_stopping_patience=0.25, reduction_factor=3, objective_returned='last', random_state=None)

Bases: Stopper

Stopper based on learning curve extrapolation (LCE) to evaluate if the iterations of the learning algorithm should be stopped.

Single-Objective	Multi-Objectives	Failures
✅	❌	❌

The LCE is based on a parametric learning curve model (LCM) which is modeling the score as a function of the number of training steps. Training steps can correspond to the number of training epochs, the number of training batches, the number of observed samples or any other quantity that is iterated through during the training process. The LCE is based on the following steps:

1. An early stopping condition is always checked first. If the early stopping condition is met, the LCE is not applied.

2. Then, some safeguard conditions are checked to ensure that the LCE can be applied (number of observed steps must be greater or equal to the number of parameters of the LCM).

3. If the LCM cannot be fitted (number of observed steps is less than number of parameters of the model), then the last observed step is compared to hitorical performance of others at the same step to check if it is a low-performing outlier (outlier in the direction of performing worse!) using the IQR criterion.

4. If the LCM can be fitted, a least square fit is performed to estimate the parameters of the LCM.

5. The probability of the current LC to perform worse than the best observed score at the maximum iteration is computed using Monte-Carlo Markov Chain (MCMC).

To use this stopper, you need to install the following dependencies:

$ jax>=0.3.25
$ numpyro

Parameters:

• max_steps (int) – The maximum number of training steps which can be performed.

• min_steps (int, optional) – The minimum number of training steps which can be performed. Defaults to 4. It is better to have at least as many steps as the number of parameters of the fitted learning curve model. For example, if lc_model="mmf4" then min_steps should be at least 4.

• lc_model (str, optional) – The parameteric learning model to use. It should be a string in the following list: ["lin2", "loglin2", "loglin3", "loglin4", "pow3","mmf4", "vapor3", "logloglin2", "hill3", "logpow3", "pow4", "exp4", "janoschek4", "weibull4", "ilog2"]. The number in the name corresponds to the number of parameters of the parametric model. Defaults to "mmf4".

• min_done_for_outlier_detection (int, optional) – The minimum number of observed scores at the same step to check for if it is a lower-bound outlier. Defaults to 10.

• iqr_factor_for_outlier_detection (float, optional) – The IQR factor for outlier detection. The higher it is the more inclusive the condition will be (i.e. if set very large it is likely not going to detect any outliers). Defaults to 1.5.

• prob_promotion (float, optional) – The threshold probabily to stop the iterations. If the current learning curve has a probability greater than prob_promotion to be worse that the best observed score accross all evaluations then the current iterations are stopped. Defaults to 0.9 (i.e. probability of 0.9 of being worse).

• early_stopping_patience (float, optional) – The patience of the early stopping condition. If it is an int it is directly corresponding to a number of iterations. If it is a float then it is corresponding to a proportion between [0,1] w.r.t. max_steps. Defaults to 0.25 (i.e. 25% of max_steps).

• objective_returned (str, optional) – The returned objective. It can be a value in ["last", "max", "alc"] where "last" corresponds to the last observed score, "max" corresponds to the maximum observed score and "alc" corresponds to the area under the learning curve. Defaults to “last”.

• random_state (int or np.RandomState, optional) – The random state of estimation process. Defaults to None.

Raises:

ValueError – parameters are not valid.

Methods

observe	Observe a new objective value.
stop	Returns True if the evaluation should be stopped and False otherwise.
to_json	Returns a dict version of the stopper which can be saved as JSON.
transform_objective	Replaces the currently observed objective by the maximum objective observed from the start.

Attributes

objective	Last observed objective.
observations	Returns a copy of the list of observations with 0-index the budgets and 1-index the objectives.
step	Last observed step.

property objective

Last observed objective.

property observations: list

Returns a copy of the list of observations with 0-index the budgets and 1-index the objectives.

observe(budget: float, objective: float)

Observe a new objective value.

Parameters:

• budget (float) – the budget used to obtain the objective (e.g., the number of epochs).

• objective (float) – the objective value to observe (e.g, the accuracy).

property step

Last observed step.

stop() → bool

Returns True if the evaluation should be stopped and False otherwise.

Returns:

(step >= max_steps).

Return type:

bool

to_json()

Returns a dict version of the stopper which can be saved as JSON.

transform_objective(objective: float)

Replaces the currently observed objective by the maximum objective observed from the start. Identity transformation by default.