merlin.models.tf.CachedUniformSampler

class merlin.models.tf.CachedUniformSampler(*args, **kwargs)

Bases: merlin.models.tf.blocks.sampling.cross_batch.CachedCrossBatchSampler

Provides a cached uniform negative sampling for two-tower item retrieval model. It is similar to the CachedCrossBatchSampler, the main difference is that the CachedUniformSampler is a popularity-based sampler and CachedUniformSampler only keeps unique item embeddings in the queue for uniform sampling. The caches consists of two fixed capacity internal FIFO queues that hold both item ids and item embeddings. It ensures that each item id (and corresponding embedding) is added only once into the queue. If the item id was already included in the queue by a previous batch, the embedding is updated. As the queues reach their capacity of unique items, new items will replace the first items added to the queue.

This is a cached implementation of 1, where those authors proposed combining in-batch sampling (our InBatchSampler()) with uniform sampling. Differently from [1] which requires a separate dataset with the all unique items (and corresponding features) to generate the item embeddings, our streaming approach in CachedUniformSampler keeps caching new items as they appear in the batches. That means that the very first processed batches will have less negative samples.

P.s. Ignoring the false negatives (negative items equal to the positive ones) is managed by ItemRetrievalScorer(…, sampling_downscore_false_negatives=True)

References

1

Yang, Ji, et al. “Mixed negative sampling for learning two-tower neural networks in recommendations.” Companion Proceedings of the Web Conference 2020. 2020.

Parameters

• capacity (int) – The queue capacity to store samples

• ignore_last_batch_on_sample (bool) – Whether should include the last batch in the sampling. By default False, as for sampling from the current batch we recommend InBatchSampler(), which allows computing gradients for in-batch negative items

• item_id_feature_name (str) – Name of the column containing the item ids Defaults to item_id

__init__(capacity: int, ignore_last_batch_on_sample: bool = True, item_id_feature_name: str = 'item_id', **kwargs)

Methods

__init__(capacity[, …])
add(inputs[, training])	Updates the FIFO queue with batch item embeddings (for items whose ids were already added to the queue) and adds to the queue the items seen for the first time
add_loss(losses, **kwargs)	Add loss tensor(s), potentially dependent on layer inputs.
add_metric(value[, name])	Adds metric tensor to the layer.
add_update(updates)	Add update op(s), potentially dependent on layer inputs.
add_variable(*args, **kwargs)	Deprecated, do NOT use! Alias for add_weight.
add_weight([name, shape, dtype, …])	Adds a new variable to the layer.
build(input_shapes)
call(inputs[, training])	Adds the current batch to the FIFO queue cache and samples all items embeddings from the last N cached batches.
compute_mask(inputs[, mask])	Computes an output mask tensor.
compute_output_shape(input_shape)	Computes the output shape of the layer.
compute_output_signature(input_signature)	Compute the output tensor signature of the layer based on the inputs.
count_params()	Count the total number of scalars composing the weights.
finalize_state()	Finalizes the layers state after updating layer weights.
from_config(config)	Creates a layer from its config.
get_config()	Returns the config of the layer.
get_input_at(node_index)	Retrieves the input tensor(s) of a layer at a given node.
get_input_mask_at(node_index)	Retrieves the input mask tensor(s) of a layer at a given node.
get_input_shape_at(node_index)	Retrieves the input shape(s) of a layer at a given node.
get_output_at(node_index)	Retrieves the output tensor(s) of a layer at a given node.
get_output_mask_at(node_index)	Retrieves the output mask tensor(s) of a layer at a given node.
get_output_shape_at(node_index)	Retrieves the output shape(s) of a layer at a given node.
get_weights()	Returns the current weights of the layer, as NumPy arrays.
sample()
set_max_num_samples(value)
set_weights(weights)	Sets the weights of the layer, from NumPy arrays.
with_name_scope(method)	Decorator to automatically enter the module name scope.

Attributes

activity_regularizer	Optional regularizer function for the output of this layer.
compute_dtype	The dtype of the layer’s computations.
dtype	The dtype of the layer weights.
dtype_policy	The dtype policy associated with this layer.
dynamic	Whether the layer is dynamic (eager-only); set in the constructor.
inbound_nodes	Return Functional API nodes upstream of this layer.
input	Retrieves the input tensor(s) of a layer.
input_mask	Retrieves the input mask tensor(s) of a layer.
input_shape	Retrieves the input shape(s) of a layer.
input_spec	InputSpec instance(s) describing the input format for this layer.
item_embeddings_queue
losses	List of losses added using the add_loss() API.
max_num_samples
metrics	List of metrics added using the add_metric() API.
name	Name of the layer (string), set in the constructor.
name_scope	Returns a tf.name_scope instance for this class.
non_trainable_variables
non_trainable_weights	List of all non-trainable weights tracked by this layer.
outbound_nodes	Return Functional API nodes downstream of this layer.
output	Retrieves the output tensor(s) of a layer.
output_mask	Retrieves the output mask tensor(s) of a layer.
output_shape	Retrieves the output shape(s) of a layer.
required_features
stateful
submodules	Sequence of all sub-modules.
supports_masking	Whether this layer supports computing a mask using compute_mask.
trainable
trainable_variables
trainable_weights	List of all trainable weights tracked by this layer.
updates
variable_dtype	Alias of Layer.dtype, the dtype of the weights.
variables	Returns the list of all layer variables/weights.
weights	Returns the list of all layer variables/weights.

add(inputs: Dict[str, tensorflow.python.framework.ops.Tensor], training: bool = True) → None

Updates the FIFO queue with batch item embeddings (for items whose ids were already added to the queue) and adds to the queue the items seen for the first time

Parameters

• inputs (TabularData) –

  Dict with two keys:

  ”items_embeddings”: Items embeddings tensor “items_metadata”: Dict like {“<feature name>”: “<feature tensor>”} which contains features that might be relevant for the sampler (e.g. item id, item popularity, item recency). The CachedUniformSampler requires the “item_id” feature to identify items already added to the queue.

• training (bool, optional) – Flag indicating if on training mode, by default True