---

name: losses-metrics description: "Choose, configure, and validate Segmentation Models losses and metrics for Keras compile and evaluation workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

Segmentation Models losses and metrics

Use this sub-skill when a task needs to select, configure, debug, or sanity-check segmentation_models losses and metrics for Keras / TensorFlow Keras compile, evaluate, or deterministic tensor checks.

Route here for

• Binary segmentation metrics such as IOUScore(threshold=0.5) and FScore(threshold=0.5).
• Multiclass or multilabel choices among Jaccard, Dice, cross-entropy, and focal losses.
• class_indexes, class_weights, threshold, per_image, smooth, beta, alpha, and gamma behavior.
• Loss aliases and combinations such as bce_jaccard_loss, categorical_focal_dice_loss, or custom sums like DiceLoss(...) + 0.5 * BinaryFocalLoss(...).
• Reproducible hand checks for IoU/F-score/precision/recall/Dice/Jaccard math.

Route elsewhere

• Model constructors, encoder/backbone selection, Unet, Linknet, FPN, or PSPNet architecture details: use the model-construction sub-skill.
• Full data loaders, generators, augmentation, preprocessing pipelines, fitting loops, checkpointing, or plotting: use the training-utilities sub-skill.

Operating prerequisites

In modern environments, prefer TensorFlow Keras and initialize Segmentation Models before constructing loss or metric objects:

import os
os.environ.setdefault("SM_FRAMEWORK", "tf.keras")  # set before importing segmentation_models
import segmentation_models as sm

The package constructs KerasObject-based aliases at import time and injects Keras backend/layers/models/utils submodules during framework initialization. If these submodules are missing, loss and metric construction fails before model training begins.

What to read

• references/api-reference.md — class signatures, aliases, formulas, and exact parameter semantics.
• references/recipes.md — binary, multiclass, multilabel, class-imbalance, and ignored-background compile recipes.
• references/troubleshooting.md — common shape, threshold, class-weight, activation/loss, smooth, and framework errors.

Bundled check

Run the bundled smoke check in the target environment when a task needs deterministic validation without external data:

python scripts/check_losses_metrics.py --threshold-demo

The script builds tiny masks, evaluates the packaged losses/metrics, and prints expected IoU, F-score, precision, recall, loss, class-index, and optional threshold signals.