sdf-loss 0.1.0

Signed Distance Function based loss functions for deep learning semantic segmentation to miss fewer instances.

SDF Loss

Signed Distance Function (SDF) based loss functions for deep learning semantic segmentation.

Overview

This library provides PyTorch loss functions that use Signed Distance Functions to weight pixels based on their distance from object boundaries. This approach puts heavier penalties on false positives and false negatives that are farther from the correct boundary, leading to more accurate segmentation results.

Installation

Install directly from PyPI:

pip install sdf-loss

Or using uv:

uv add sdf-loss

For development:

git clone https://github.com/Halyjo/sdf_loss.git
cd sdf_loss
uv sync

Quick Start

import torch
from sdf_loss import DiSCoLoss

# Initialize the loss function
criterion = DiSCoLoss()

# Your model predictions (logits) and ground truth
pred_logits = model(images)  # Shape: (B, 1, H, W)
target = ground_truth         # Shape: (B, 1, H, W), binary mask

# Compute loss
loss = criterion(pred_logits, target)
loss.backward()

Available Loss Functions

DiSCoLoss (Recommended)

Distance-scaled combination loss that combines BCE, Dice, and their SDF-weighted variants.

from sdf_loss import DiSCoLoss

# Default: only SDF-weighted losses
criterion = DiSCoLoss(
    normalize=True,           # Normalize SDF to [-1, 1]
    baseloss_weight=0,       # Weight for BCE + Dice
    sdfweighted_weight=1,    # Weight for SDF-weighted losses
    clip_negatives=False     # Clip negative distances
)

loss = criterion(pred_logits, target)

SDFWeightedBCELoss

Binary cross-entropy loss weighted by SDF differences.

from sdf_loss import SDFWeightedBCELoss

criterion = SDFWeightedBCELoss(
    reduction="mean",        # 'mean', 'sum', or 'none'
    normalize=True,
    clip_negatives=False
)

loss = criterion(pred_logits, target)

SDFWeightedDiceLoss

Dice loss weighted by SDF differences.

from sdf_loss import SDFWeightedDiceLoss

criterion = SDFWeightedDiceLoss(
    from_logits=True,
    normalize=True,
    clip_negatives=False
)

loss = criterion(pred_logits, target)

DiceLoss

Standard Dice loss with optional custom weighting function.

from sdf_loss import DiceLoss

criterion = DiceLoss(
    from_logits=True,
    smooth=0.0
)

loss = criterion(pred_logits, target)

Drop-in Replacement for BCE Loss

You can easily replace torch.nn.BCEWithLogitsLoss with DiSCoLoss:

# Before
criterion = torch.nn.BCEWithLogitsLoss()

# After - simple drop-in replacement
criterion = DiSCoLoss()

# Usage remains the same
loss = criterion(pred_logits, target)

Key Features

• Boundary-aware: Focuses on pixels near object boundaries
• Distance-weighted: Penalizes errors proportional to distance from correct boundary
• PyTorch native: Fully compatible with PyTorch training loops
• GPU compatible: Works with CUDA tensors
• Differentiable: Full gradient flow for backpropagation
• Flexible: Multiple loss functions and customizable parameters

Parameters

Common Parameters

• normalize (bool): Normalize SDF values to [-1, 1] range. Default: True
• clip_negatives (bool): Clip negative distance values to 0. Default: False
• from_logits (bool): Whether input is logits (before sigmoid). Default: True

DiSCoLoss Specific

• baseloss_weight (float): Weight for standard BCE + Dice losses. Default: 0
• sdfweighted_weight (float): Weight for SDF-weighted losses. Default: 1

Examples

Basic Training Loop

import torch
from sdf_loss import DiSCoLoss

model = YourSegmentationModel()
criterion = DiSCoLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)

for epoch in range(num_epochs):
    for images, masks in dataloader:
        optimizer.zero_grad()

        pred_logits = model(images)
        loss = criterion(pred_logits, masks)

        loss.backward()
        optimizer.step()

Combining Base and SDF-weighted Losses

# Use both base losses and SDF-weighted losses
criterion = DiSCoLoss(
    baseloss_weight=0.3,      # 30% base losses
    sdfweighted_weight=0.7    # 70% SDF-weighted losses
)

Custom Reduction

from sdf_loss import SDFWeightedBCELoss

# Get per-pixel losses for custom weighting
criterion = SDFWeightedBCELoss(reduction="none")
loss = criterion(pred_logits, target)  # Shape: (B, 1, H, W)

# Apply custom weighting
custom_weights = compute_your_weights(target)
weighted_loss = (loss * custom_weights).mean()

Requirements

• Python >= 3.10
• PyTorch >= 1.10.0
• NumPy >= 1.20.0
• SciPy >= 1.7.0
• scikit-image >= 0.19.0

Testing

Run the test suite:

pytest tests/ -v

Citation

If you use this library in your research, please cite:

@article{your_paper,
  title={Your Paper Title},
  author={Your Name},
  journal={Your Journal},
  year={2025}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgments

The DiceLoss implementation is inspired by segmentation-models-pytorch (SMP).