Hyperparameters

Functions to handle the definition and manipulation of hyperparameters

metatrain.utils.hypers.get_hypers_list(hypers_cls: Type[HypersType]) → list[str][source]

Get the list of hyperparameter names defined in a TypedDict hypers class.

Inheritance of parameters is allowed from parent classes, but make sure that the parent classes only contain hyperparameters as attributes! (i.e., no methods allowed). Private attributes (starting with “_”) are not considered as hyperparameters, so one can have arbitrary private methods or attributes in the class and its parents, although this is not recommended.

Parameters:

hypers_cls (Type[HypersType]) – The class defining the hyperparameters.

Returns:

A list with the names of the hyperparameters.

Return type:

list[str]

metatrain.utils.hypers.init_with_defaults(hypers_cls: Type[HypersType]) → dict[source]

Initialize a TypedDict hypers class with its default values.

Inheritance of parameters is allowed from parent classes, but make sure that the parent classes only contain hyperparameters as attributes! (i.e., no methods allowed). Private attributes (starting with “_”) are not considered as hyperparameters, so one can have arbitrary private methods or attributes in the class and its parents, although this is not recommended.

Parameters:

hypers_cls (Type[HypersType]) – The class defining the hyperparameters.

Returns:

A dict with the default hyperparameters.

Return type:

dict

metatrain.utils.hypers.overwrite_defaults(hypers_cls: Type, new_defaults: dict) → None[source]

Overwrite the default hyperparameters.

This function does not check that the new defaults correspond to valid hyperparameters of the given hypers class. If the new defaults contain keys that are not hyperparameters of the class, they will simply be ignored.

Parameters:

  • hypers_cls (Type) – The hypers class whose defaults to overwrite.

  • new_defaults (dict) – A dict with the new default hyperparameters.

Return type:

None

metatrain.utils.hypers.get_hypers_diff(old_hypers: Mapping[str, Any], new_hypers: Mapping[str, Any]) → dict[str, tuple[Any, Any]][source]

Get the difference between two hypers dictionaries.

Only keys present in new_hypers are compared. Missing keys are treated as “keep the checkpoint value”. Nested mappings are compared the same way, so a partial YAML block such as soap: {max_radial: 4} does not mismatch a checkpoint that also stores cutoff.

Parameters:
Returns:

A dict of dotted paths to (old, new) pairs that differ. A missing old key is reported as "<not present>".

Return type:

dict[str, tuple[Any, Any]]

metatrain.utils.hypers.raise_hypers_mismatch(hypers_diff: Mapping[str, tuple[Any, Any]]) → None[source]

Raise an error if the hypers diff is not empty.

The error shows a report of the mismatched hyperparameters.

Parameters:

hypers_diff (Mapping[str, tuple[Any, Any]]) – A dict with the hyperparameters that are different. It can be computed using get_hypers_diff().

Return type:

None

metatrain.utils.hypers.raise_if_hypers_mismatch(old_hypers: Mapping[str, Any], new_hypers: Mapping[str, Any]) → None[source]

Raise an error if the new hypers do not match the old hypers.

Parameters:
Return type:

None