Skip to content

Documentation for MinAggregator

Functionality

MinAggregator is designed to return the minimum value among a list of rank scores. It calculates the minimum value from a list or tensor of rank scores, and when the list is empty, it returns a default value provided at initialization. This helps in scenarios where a minimal measure is needed, including support for gradient computation using differentiable operations.

Parameters

  • if_empty_value (float): The value returned when the rank list is empty.
  • ranks: A list of floats or a tensor containing rank scores for the aggregation.

Usage

  • Purpose: Aggregates rank values by selecting the lowest score. This functionality is implemented in two methods:

    • _aggregate: For standard minimum calculation from lists or tensors.
    • _aggregate_differentiable: For minimum calculation that allows gradient propagation during training.

  • Motivation: To capture the minimal performance among several candidates in ranking tasks.

  • Inheritance: Inherits from RanksAggregator, aligning with a set of aggregation strategies.

Example

Using _aggregate method: 

aggregator = MinAggregator(if_empty_value=0.0)
result = aggregator._aggregate([3.5, 2.1, 4.7])
print(result)  # Outputs: 2.1

Using _aggregate_differentiable method: 

import torch

ranks = torch.tensor([2.0, 5.0, 1.0])
min_value = aggregator._aggregate_differentiable(ranks)

Key Points

  • The _aggregate method calculates the minimum from a list or tensor of rank scores.
  • The _aggregate_differentiable method uses the "differentiable_extreme" function to compute the differentiable minimum, allowing for gradient flow during training.