---

name: clip_contrastive_objective description: Compute CLIP-style normalized image/text similarity and symmetric InfoNCE loss for paired multimodal embeddings.

CLIP Contrastive Objective

Use this skill when implementing or checking the core CLIP training mechanism: paired image/text embeddings, L2 normalization, temperature-scaled cosine logits, and symmetric InfoNCE loss.

Inputs

• Image embedding matrix and text embedding matrix with equal batch size.
• Positive pairs aligned by row index.
• Positive logit_scale.

Outputs

• Normalized embeddings, logits, symmetric loss, and diagnostics.

Workflow

1. Reject mismatched batches, empty matrices, zero vectors, or non-positive logit scale.
2. Normalize each row to unit length.
3. Compute logits = logit_scale * image_norm @ text_norm.T.
4. Compute cross-entropy for image-to-text and text-to-image directions.
5. Average the two losses and report diagonal separation diagnostics.

Validation

Run python tests/test_contrastive_objective.py.

Limitations

A tiny batch confirms the mechanism but is not full-scale distributed CLIP pretraining.

Cycle 3 normalization ablation

Stress checks include unequal embedding magnitudes to confirm L2 normalization makes the contrastive mechanism scale-invariant before logits are computed.