SNN FrameWork Documentation

Welcome to the SNN (Siamese Neural Network) FrameWork - a comprehensive, highly customizable framework for experimenting with Siamese Neural Networks and Reinforcement Learning approaches to network data analysis.

Overview

The SNN FrameWork is designed for researchers and practitioners working with network traffic analysis, video quality prediction, and machine learning on time-series data. It provides a modular architecture that supports both traditional supervised learning and reinforcement learning approaches.

Key Features:

Siamese Neural Networks: Specialized architecture for similarity learning and metric learning tasks
Reinforcement Learning Integration: Actor-Critic models for adaptive decision making
Video Quality Assessment: VMAF-based quality prediction and optimization
Modular Design: Easily extensible with custom models, callbacks, and data processing pipelines
Configuration-Driven: INI-based configuration system for reproducible experiments
Comprehensive Logging: Detailed experiment tracking and result analysis

Quick Start

Installation

# Clone the repository
git clone <repository-url>
cd SNN2

# Install dependencies
pip install -r requirements.txt

# Install the package
pip install -e .

Basic Usage

# Run with default configuration
snn -c config/basic

# Enable reinforcement learning
snn -c config/basic --reinforcement

# Run inference only
snn -c config/basic --inference "model_label"

# Enable verbose logging
snn -c config/basic -vvv

Architecture Overview

The framework is organized into several key components:

Core Components

Data Processing (src/core/data/): Handles data preprocessing, windowing, and separation
Models (src/model/): Siamese networks, custom architectures, and reinforcement learning models
Experiments (src/core/experiment.py): Orchestrates training, evaluation, and inference
Actions (src/actions/): Data transformation and separation strategies
I/O System (src/io/): Configuration management, logging, and file handling

Reinforcement Learning

Actor-Critic Models: Policy gradient methods for adaptive learning
Reward Functions: Customizable reward systems for different objectives
Action Policies: Various exploration and exploitation strategies
Environment Management: Configurable RL environments for different scenarios

Configuration System

The framework uses a hierarchical INI-based configuration system with the following structure:

Default Configuration: Base settings in default_ini/
Experiment-Specific: Override configurations in Conf/ directories
Parameter Interpolation: Dynamic parameter substitution and inheritance

Example configuration sections:

model.ini: Neural network architecture settings
reinforcement.ini: RL-specific parameters (γ, β, action spaces)
flow.ini: Data flow and preprocessing parameters
callbacks.ini: Training callbacks and monitoring

Use Cases

Video Quality Optimization

The framework excels at video quality prediction and optimization tasks:

VMAF Prediction: Learn to predict video quality metrics
Adaptive Streaming: RL-based bitrate adaptation
Quality-Aware Encoding: Optimize encoding parameters for perceptual quality

Network Traffic Analysis

Apply Siamese networks to network data:

Traffic Classification: Identify similar network flows
Anomaly Detection: Detect unusual patterns in network behavior
Performance Prediction: Forecast network performance metrics

Research Applications

Metric Learning: Learn meaningful distance functions for complex data
Few-Shot Learning: Leverage Siamese architectures for limited data scenarios
Multi-Objective Optimization: Balance multiple competing objectives with RL

Advanced Features

Custom Model Development

Extend the framework with custom models:

from SNN2.src.model.custom.customWrapper import CustomWrapper

class MyCustomModel(CustomWrapper):
    def build_model(self):
        # Implement your custom architecture
        pass

Reinforcement Learning Extensions

Create custom reward functions and action policies:

from SNN2.src.model.reward.rewardWrapper import RewardWrapper

class CustomReward(RewardWrapper):
    def compute_reward(self, state, action, next_state):
        # Implement custom reward logic
        pass

Indices and Tables

Support and Contributing

Documentation: Complete API reference and examples
Issues: Report bugs and request features on the issue tracker
Contributing: See CONTRIBUTING.md for development guidelines
License: BSD-3-Clause license

The SNN FrameWork is actively developed and maintained. We welcome contributions from the community!