cv-tversky-loss
Tversky loss with independent alpha/beta constants to separately penalize false positives and false negatives in imbalanced segmentation.
SKILL.mdskills/cv/tversky-loss/SKILL.md
GitHubTversky Loss
Overview
Dice loss weights false positives and false negatives equally, but in many segmentation tasks they have different costs — missing a tumor (FN) is worse than a false alarm (FP). Tversky loss generalizes Dice by introducing alpha and beta to independently control FP and FN penalties: Tversky = TP / (TP + alpha*FP + beta*FN). Setting alpha=beta=0.5 recovers Dice; alpha=0.3, beta=0.7 penalizes false negatives more, improving recall on small or rare objects.
Quick Start
import torch
import torch.nn as nn
import torch.nn.functional as F
class TverskyLoss(nn.Module):
def __init__(self, alpha=0.3, beta=0.7, smooth=1.0):
super().__init__()
self.alpha = alpha
self.beta = beta
self.smooth = smooth
def forward(self, inputs, targets):
inputs = torch.sigmoid(inputs).view(-1)
targets = targets.view(-1)
tp = (inputs * targets).sum()
fp = ((1 - targets) * inputs).sum()
fn = (targets * (1 - inputs)).sum()
tversky = (tp + self.smooth) / (tp + self.alpha * fp + self.beta * fn + self.smooth)
return 1 - tversky
# Usage: penalize FN more than FP
criterion = TverskyLoss(alpha=0.3, beta=0.7)
loss = criterion(logits, masks)
Workflow
- Apply sigmoid to raw logits
- Compute true positives, false positives, false negatives
- Compute Tversky index with alpha/beta weighting
- Return
1 - Tverskyas loss
Key Decisions
- alpha/beta:
alpha < betaboosts recall (fewer FN);alpha > betaboosts precision (fewer FP) - alpha=beta=0.5: Equivalent to Dice loss — use as a sanity check
- Focal Tversky: Add
(1-Tversky)^gammato further focus on hard examples - Smooth: 1.0 prevents division by zero; lower values (0.01) for sharper gradients