lczerolens.xai#

XAI module.

class lczerolens.xai.ActivationLens(module_exp: str | None = None)#

Class for activation-based XAI methods.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Any#: Cache the activations for a given model and input.

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → Dict[str, Any] | None#: Cache the activations for a given model and dataset.

is_compatible(wrapper: ModelWrapper) → bool#: Caching is compatible with all torch models.

class lczerolens.xai.AndBinaryConcept(*concepts: BinaryConcept)#

Class for binary concept-based XAI methods.

compute_label(board: Board) → Any#: Compute the label for a given model and input.

class lczerolens.xai.BestLegalMoveConcept(wrapper: ModelWrapper)#

Class for move concept-based XAI methods.

compute_label(board: Board) → int#: Compute the label for a given model and input.

class lczerolens.xai.BinaryConcept#

Class for binary concept-based XAI methods.

static compute_metrics(predictions, labels)#: Compute the metrics for a given model and input.

Class for concept

save(file_name: str, n_history: int = 0, indices=None)#: Save the dataset to a file.

Note

As the board needs to be unpiled use the preprocess_game method.

class lczerolens.xai.CrpLens#

Class for wrapping the LCZero models.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Tensor#: Computes the heatmap for a given board.

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → dict#: Computes the statistics for a given board.

is_compatible(wrapper: ModelWrapper) → bool#: Returns whether the lens is compatible with the model.

class lczerolens.xai.HasMateThreatConcept#

Class for material concept-based XAI methods.

compute_label(board: Board) → int#: Compute the label for a given model and input.

class lczerolens.xai.HasMaterialAdvantageConcept(relative: bool = True)#

Class for material concept-based XAI methods.

piece_values#

The piece values.

Type:: Dict[int, int]

compute_label(board: Board, piece_values: Dict[int, int] | None = None) → int#: Compute the label for a given model and input.

class lczerolens.xai.HasPieceConcept(piece: str, relative: bool = True)#

Class for material concept-based XAI methods.

compute_label(board: Board) → int#: Compute the label for a given model and input.

class lczerolens.xai.HasThreatConcept(piece: str, relative: bool = True)#

Class for material concept-based XAI methods.

compute_label(board: Board) → int#: Compute the label for a given model and input.

class lczerolens.xai.Lens#

Generic lens class.

abstract analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Any#: Computes the heatmap for a given board.

abstract analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → Dict[Any, Any] | None#: Computes the statistics for a given board.

classmethod from_name(name: str, **kwargs) → Lens#: Returns the lens from its name.

classmethod get_subclass(name: str) → Type[Lens]#: Returns the subclass.

abstract is_compatible(wrapper: ModelWrapper) → bool#: Returns whether the lens is compatible with the model.

classmethod register(name: str)#: Registers the lens.

class lczerolens.xai.LrpLens#

Class for wrapping the LCZero models.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Tensor#

Runs basic LRP on the model.

Parameters:

board (chess.Board) – The board to compute the heatmap for.
wrapper (ModelWrapper) – The model wrapper.

Returns:

The heatmap for the given board.

Return type:

torch.Tensor

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → Dict[int, Tensor] | None#: Cache the activations for a given model and dataset.

static context(wrapper: ModelWrapper, composite: Composite | None = None, replace_onnx2torch: bool = True, linearise_softmax: bool = False)#: Context manager for the lens.

is_compatible(wrapper: ModelWrapper) → bool#

Returns whether the lens is compatible with the model.

Parameters:: wrapper (ModelWrapper) – The model wrapper.
Returns:: Whether the lens is compatible with the model.
Return type:: bool

class lczerolens.xai.OrBinaryConcept(*concepts: BinaryConcept)#

Class for binary concept-based XAI methods.

compute_label(board: Board) → Any#: Compute the label for a given model and input.

class lczerolens.xai.PatchingLens(patching_dict: Dict[str, Callable])#

Class for patching-based XAI methods.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Tensor#: Analyse a single board with the probing lens.

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → dict#: Analyse a dataset with the probing lens.

is_compatible(wrapper: ModelWrapper) → bool#: Returns whether the lens is compatible with the model.

class lczerolens.xai.PieceBestLegalMoveConcept(wrapper: ModelWrapper, piece: str)#

Class for move concept-based XAI methods.

compute_label(board: Board) → int#: Compute the label for a given model and input.

class lczerolens.xai.PolicyLens#

Class for wrapping the LCZero models.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Tensor#: Compute the policy for a given board.

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → Dict[int, Any] | None#: Computes the statistics for a given board.

is_compatible(wrapper: ModelWrapper) → bool#: Returns whether the lens is compatible with the model.

class lczerolens.xai.ProbingLens(probe_dict: Dict[str, Probe])#

Class for probing-based XAI methods.

analyse_board(board: Board, wrapper: ModelWrapper, **kwargs) → Tensor#: Analyse a single board with the probing lens.

analyse_dataset(dataset: Dataset, wrapper: ModelWrapper, batch_size: int, collate_fn: Callable | None = None, save_to: str | None = None, **kwargs) → dict#: Analyse a dataset with the probing lens.

is_compatible(wrapper: ModelWrapper) → bool#: Returns whether the lens is compatible with the model.