Documentation for differentiable_extreme

Functionality

Differentiable approximation of max or min using softmax.

Parameters

• x: Input tensor.
• beta: Scaling parameter. A larger value sharpens the softmax, approximating the true max/min more closely.
• mode: Either "max" or "min" to specify the approximation target.

Usage

• Purpose: Provides a differentiable method to compute an extreme value over a tensor.

Example

Simple usage:

import torch
from embedding_studio.embeddings.models.utils.differentiable_extreme \
     import differentiable_extreme

x = torch.tensor([1.0, 2.5, 0.5])
result = differentiable_extreme(x, beta=1e5, mode="max")
print(result)