Activation Output

A Python utility library for computing activation outputs in machine learning classification tasks. This package provides a unified interface for handling binary, multiclass, and multilabel classification scenarios.

Features

• Multiple Classification Types: Support for binary, multiclass, and multilabel classification
• Numerically Stable: Implements stable computation methods to prevent overflow/underflow
• Type Safe: Full type hints and validation for robust usage
• NumPy Integration: Built on NumPy for efficient array operations

Installation

This project uses uv for dependency management. To install:

# Clone the repository
git clone <repository-url>
cd krish-sandbox

# Install dependencies using uv
uv sync

Requirements

• Python >= 3.13
• NumPy >= 2.3.1

Usage

Basic Usage

import numpy as np
from krish_sandbox import activation_output

# Binary classification
logits = np.array([2.5, -1.0, 0.8])
probabilities = activation_output(logits, task_type='binary')
print(probabilities)  # [0.92414182 0.26894142 0.68997448]

# Multiclass classification
logits = np.array([[1.0, 2.0, 0.5], [0.1, 0.2, 0.3]])
probabilities = activation_output(logits, task_type='multiclass')
print(probabilities)
# [[0.24472847 0.66524096 0.09003057]
#  [0.30060961 0.33222499 0.3671654 ]]

# Multilabel classification
logits = np.array([[1.0, -0.5, 2.0], [0.1, 0.8, -0.3]])
probabilities = activation_output(logits, task_type='multilabel')
print(probabilities)
# [[0.73105858 0.37754067 0.88079708]
#  [0.52497919 0.68997448 0.42555748]]

Task Types

Binary Classification

• Input: Scalar or 1D array of logits
• Output: Probabilities between 0 and 1
• Activation: Sigmoid function

Multiclass Classification

• Input: 2D array of logits (batch_size, num_classes)
• Output: Probability distribution over classes (sums to 1)
• Activation: Softmax function with numerical stability

Multilabel Classification

• Input: 2D array of logits (batch_size, num_labels)
• Output: Independent probabilities for each label
• Activation: Sigmoid function applied to each logit

API Reference

activation_output(logits, task_type='binary')

Computes activation output for classification tasks.

Parameters:

• logits (Union[npt.NDArray, Sequence]): Raw model outputs
• task_type (str): One of ['binary', 'multiclass', 'multilabel']. Defaults to 'binary'

Returns:

• npt.NDArray: Probabilities after applying appropriate activation function

Raises:

• ValueError: If logits array is empty or contains NaN/infinite values
• ValueError: If task_type is not one of the valid options

Development

Setup Development Environment

# Install development dependencies
uv sync --dev

# Activate virtual environment
source .venv/bin/activate  # On Unix/macOS
# or
.venv\Scripts\activate     # On Windows