autointent.modules.scoring.KNNScorer#

class autointent.modules.scoring.KNNScorer(embedder_name, k, weights='distance', db_dir=None, embedder_device='cpu', batch_size=32, max_length=None, embedder_use_cache=False)#

Bases: autointent.modules.abc.ScoringModule

K-nearest neighbors (KNN) scorer for intent classification.

This module uses a vector index to retrieve nearest neighbors for query utterances and applies a weighting strategy to compute class probabilities.

Variables:

weights – Weighting strategy used for scoring.
_vector_index – VectorIndex instance for neighbor retrieval.
name – Name of the scorer, defaults to “knn”.
prebuilt_index – Flag indicating if the vector index is prebuilt.

Parameters:

embedder_name (str)
k (int)
weights (autointent.custom_types.WEIGHT_TYPES)
db_dir (str | None)
embedder_device (str)
batch_size (int)
max_length (int | None)
embedder_use_cache (bool)

Examples#

from autointent.modules.scoring import KNNScorer
utterances = ["hello", "how are you?"]
labels = [0, 1]
scorer = KNNScorer(
    embedder_name="sergeyzh/rubert-tiny-turbo",
    k=5,
    db_dir=db_dir,
)
scorer.fit(utterances, labels)
test_utterances = ["hi", "what's up?"]
probabilities = scorer.predict(test_utterances)
print(probabilities)  # Outputs predicted class probabilities for the utterances

[[0.67297815 0.32702185]
 [0.44031678 0.55968322]]

weights: autointent.custom_types.WEIGHT_TYPES#

name = 'knn'#

prebuilt_index: bool = False#

max_length: int | None#

embedder_name#

k#

embedder_device = 'cpu'#

batch_size = 32#

embedder_use_cache = False#

property db_dir: str#

Get the database directory for the vector index.

Returns:: Path to the database directory.
Return type:: str

classmethod from_context(context, k, weights, embedder_name=None)#

Create a KNNScorer instance using a Context object.

Parameters:

context (autointent.context.Context) – Context containing configurations and utilities.
k (int) – Number of closest neighbors to consider during inference.
weights (autointent.custom_types.WEIGHT_TYPES) – Weighting strategy for scoring.
embedder_name (str | None) – Name of the embedder, or None to use the best embedder.

Returns:

Initialized KNNScorer instance.

Return type:

KNNScorer

get_embedder_name()#

Get the name of the embedder.

Returns:: Embedder name.
Return type:: str

fit(utterances, labels)#

Fit the scorer by training or loading the vector index.

Parameters:

utterances (list[str]) – List of training utterances.
labels (list[autointent.custom_types.LabelType]) – List of labels corresponding to the utterances.

Raises:

ValueError – If the vector index mismatches the provided utterances.

Return type:

None

predict(utterances)#

Predict class probabilities for the given utterances.

Parameters:: utterances (list[str]) – List of query utterances.
Returns:: Array of predicted probabilities for each class.
Return type:: numpy.typing.NDArray[Any]

predict_with_metadata(utterances)#

Predict class probabilities along with metadata for the given utterances.

Parameters:: utterances (list[str]) – List of query utterances.
Returns:: Tuple of predicted probabilities and metadata with neighbor information.
Return type:: tuple[numpy.typing.NDArray[Any], list[dict[str, Any]] | None]

clear_cache()#

Clear cached data in memory used by the vector index.

Return type:: None

dump(path)#

Save the KNNScorer’s metadata and vector index to disk.

Parameters:: path (str) – Path to the directory where assets will be dumped.
Return type:: None

load(path)#

Load the KNNScorer’s metadata and vector index from disk.

Parameters:: path (str) – Path to the directory containing the dumped assets.
Return type:: None