---

name: dual_contrastive_loss description: Compute and inspect a DCL-style contrastive proxy loss with same-latent positives and real-target anti-collapse negatives.

Dual Contrastive Loss

Use this skill when a recovery needs a deterministic feature-level version of the paper's DCL objective. It accepts a latent-paired batch and produces an InfoNCE-like scalar loss plus mechanism checks. Do not use it as a replacement for full StyleGAN training when full models are available.

Inputs

• Batch from latent_pair_protocol.
• Temperature scalar.
• Optional real-target negative weight.

Outputs

• Mean contrastive proxy loss.
• Positive and negative similarity diagnostics.
• Checks showing same-latent positives and real-target negatives were used.

Workflow

1. Compute dot-product similarity divided by temperature.
2. For each source feature, treat the same-latent target feature as the positive.
3. Treat other generated target features and real-target exemplars as negatives.
4. Average the negative log softmax probability assigned to the positive.

Validation

Run python tests/test_dual_contrastive_loss.py from this skill directory.

Limitations

The script uses small numeric vectors and standard-library math; it preserves objective structure but not neural feature extraction.

Refinement Note

A shuffled-positive stress check should produce higher loss than aligned same-latent positives; see cycle1_shuffled_stress.json in recovery logs.