tfrs.layers.factorized_top_k.ScaNN

ScaNN approximate retrieval index for a factorized retrieval model.

Inherits From: TopK

Used in the notebooks

Used in the tutorials

This layer uses the state-of-the-art ScaNN library to retrieve the best candidates for a given query.

To understand how to use this layer effectively, have a look at the efficient retrieval tutorial.

To deploy this layer in TensorFlow Serving you can use our customized TensorFlow Serving Docker container, available on Docker Hub. You can also build the image yourself from the Dockerfile.

query_model Optional Keras model for representing queries. If provided, will be used to transform raw features into query embeddings when querying the layer. If not provided, the layer will expect to be given query embeddings as inputs.
k Default number of results to retrieve. Can be overridden in call.
distance_measure Distance metric to use.
num_leaves Number of leaves.
num_leaves_to_search Number of leaves to search.
training_iterations Number of training iterations when performing tree building.
dimensions_per_block Controls the dataset compression ratio. A higher number results in greater compression, leading to faster scoring but less accuracy and more memory usage.
num_reordering_candidates If set, the index will perform a final refinement pass on num_reordering_candidates candidates after retrieving an initial set of neighbours. This helps improve accuracy, but requires the original representations to be kept, and so will increase the final model size."
parallelize_batch_searches Whether batch querying should be done in parallel.
name Name of the layer.

ImportError if the scann library is not installed.

Methods

call

View source

Query the index.

Args
queries Query features. If query_model was provided in the constructor, these can be raw query features that will be processed by the query model before performing retrieval. If query_model was not provided, these should be pre-computed query embeddings.
k The number of candidates to retrieve. If not supplied, defaults to the k value supplied in the constructor.

Returns
Tuple of (top candidate scores, top candidate identifiers).

Raises
ValueError if index has not been called.

index

View source

Builds the retrieval index.

When called multiple times the existing index will be dropped and a new one created.

Args
candidates Matrix of candidate embeddings.
identifiers Optional tensor of candidate identifiers. If given, these will be used as identifiers of top candidates returned when performing searches. If not given, indices into the candidates tensor will be returned instead.

Returns
Self.

index_from_dataset

View source

Builds the retrieval index.

When called multiple times the existing index will be dropped and a new one created.

Args
candidates Dataset of candidate embeddings or (candidate identifier, candidate embedding) pairs. If the dataset returns tuples, the identifiers will be used as identifiers of top candidates returned when performing searches. If not given, indices into the candidates dataset will be given instead.

Returns
Self.

Raises
ValueError if the dataset does not have the correct structure.

is_exact

View source

Indicates whether the results returned by the layer are exact.

Some layers may return approximate scores: for example, the ScaNN layer may return approximate results.

Returns
True if the layer returns exact results, and False otherwise.

query_with_exclusions

View source

Query the index.

Args
queries Query features. If query_model was provided in the constructor, these can be raw query features that will be processed by the query model before performing retrieval. If query_model was not provided, these should be pre-computed query embeddings.
exclusions [query_batch_size, num_to_exclude] tensor of identifiers to be excluded from the top-k calculation. This is most commonly used to exclude previously seen candidates from retrieval. For example, if a user has already seen items with ids "42" and "43", you could set exclude to [["42", "43"]].
k The number of candidates to retrieve. Defaults to constructor k parameter if not supplied.

Returns
Tuple of (top candidate scores, top candidate identifiers).

Raises
ValueError if index has not been called. ValueError if queries is not a tensor (after being passed through the query model).