nets

Instruction file imported from janelia-cellmap/LeibNetz (.github/instructions/nets.instructions.md). Copyright stays with the author.

Network Architecture Instructions

Overview

This directory contains pre-built network architectures that demonstrate how to compose nodes from the nodes/ directory into complete neural networks.

Architecture Development Guidelines

Network Composition Patterns

Each network architecture should:

1. Use the LeibNet class as the base for composition
2. Compose nodes in a logical, reusable manner
3. Provide clear configuration options
4. Handle different input/output requirements

Existing Architectures

U-Net (unet.py)

• Classic encoder-decoder architecture with skip connections
• Demonstrates hierarchical feature extraction
• Shows proper handling of spatial dimension matching

ScaleNet (scalenet.py)

• Multi-scale processing architecture
• Demonstrates parallel processing paths
• Shows feature fusion techniques

AttentiveScaleNet (attentive_scalenet.py)

• Combines multi-scale processing with attention mechanisms
• Demonstrates integration of attention nodes
• Shows advanced feature selection patterns

Network Development Best Practices

Configuration Management

def build_network(input_channels, output_channels, **config):
    # Use configuration dictionaries for flexibility
    conv_config = config.get('conv_config', {})
    attention_config = config.get('attention_config', {})

    # Build network with configurable components
    return LeibNet([...])

Skip Connections and Feature Fusion

• Use consistent naming for skip connection handling
• Implement proper spatial alignment for feature fusion
• Consider feature channel alignment

Multi-Scale Processing

• Design for different input resolutions
• Handle scale-specific feature extraction
• Implement proper feature aggregation

Testing Network Architectures

Each network should have tests that verify:

• End-to-end forward pass functionality
• Correct output shapes for various input sizes
• Proper gradient flow (if training components are included)
• Memory efficiency for large inputs

Documentation Requirements

• Document the architectural design rationale
• Provide usage examples with typical configurations
• Document expected input/output formats
• Include performance characteristics and memory requirements

Common Network Patterns

Encoder-Decoder

# Encoder path
encoder_nodes = [
    ConvPassNode(...),
    ResampleNode(scale_factor=0.5),  # Downsample
    # ... more layers
]

# Decoder path with skip connections
decoder_nodes = [
    ResampleNode(scale_factor=2.0),  # Upsample
    # Feature fusion node for skip connections
    ConvPassNode(...),
]

Multi-Path Processing

# Parallel processing paths
path1 = LeibNet([...])
path2 = LeibNet([...])

# Feature fusion
fusion_node = FusionNode()

Integration Guidelines

• New architectures should integrate well with existing ModelWrapper functionality
• Consider compatibility with training scripts in examples/
• Design for extensibility and modification
• Provide sensible default configurations