feneck 0.0.1

Feature pyramid network implementations for computer vision tasks

🦊 FE-Neck

Feature extraction neck modules for computer vision models.

• Github repository: https://github.com/vncntcmnn/feneck/
• Documentation: https://vncntcmnn.github.io/feneck/

Installation

# Clone the repository
git clone https://github.com/vncntcmnn/feneck.git
cd feneck

# Install dependencies (CPU version)
uv sync --extra cpu

# Or with CUDA support
uv sync --extra cu118  # CUDA 11.8
uv sync --extra cu124  # CUDA 12.4

Quick Start

import torch
from feneck import FPN, PAFPN, BiFPN

# Classic FPN
fpn = FPN(
    in_channels=[256, 512, 1024, 2048],
    in_strides=[4, 8, 16, 32],
    out_channels=256
)

# Backbone features
features = [
    torch.randn(1, 256, 64, 64),   # stride 4
    torch.randn(1, 512, 32, 32),   # stride 8
    torch.randn(1, 1024, 16, 16),  # stride 16
    torch.randn(1, 2048, 8, 8),    # stride 32
]

# Forward pass
pyramid_features = fpn(features)

Available Necks

Module	Description	Best For
FPN	Feature Pyramid Network	Standard multi-scale detection
PAFPN	Path Aggregation FPN	Enhanced feature fusion
BiFPN	Bidirectional FPN	Efficient multi-scale fusion
NASFPN	NAS-discovered FPN	Learned fusion patterns
SimpleFPN	FPN for transformers	Single-scale to multi-scale
CustomCSPPAN	CSP-PAN + transformer	Advanced aggregation
HRFPN	High-Resolution FPN	Multi-scale aggregation
LRFPN	Location-Refined FPN	Remote sensing detection
CARAFE	Content-Aware upsampling	Adaptive feature reassembly
DyHead	Dynamic Head	Post-FPN refinement
FeaturePyramidExtender	Preprocessing utility	Backbone adaptation

See the documentation for detailed usage examples and API reference.

Development

Setup Development Environment

# Install with dev dependencies
make install

# Run pre-commit hooks
uv run pre-commit run -a

Running Tests

# Run tests
make test

# Run tests with coverage
make test-cov

Requirements

• Python >= 3.10
• PyTorch

License

Apache License 2.0

---

Repository initiated with fpgmaas/cookiecutter-uv.