brainstorming
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).
Explore skills
Development · SKILL.md
clip_contrastive_loss
clip_contrastive_lossCompute CLIP-style normalized similarity logits and symmetric contrastive loss for image-text batches.
Independent GitHub snapshot · Automated classification · Original documentation governs
Skill documentation
GitHubname: clip_contrastive_loss description: Compute CLIP-style normalized similarity logits and symmetric contrastive loss for image-text batches.
CLIP Contrastive Loss
Use this skill when implementing or checking the CLIP training mechanism on paired image/text embeddings. It is appropriate for full, reduced, or synthetic recovery if the result is clearly labeled.
Inputs
- Equal-length image and text embedding batches.
- A positive logit scale, corresponding to inverse temperature.
Outputs
- Scaled cosine-similarity logits.
- Image-to-text and text-to-image cross entropy.
- Average symmetric loss and top-1 retrieval accuracy.
Workflow
- Normalize each image and text feature vector.
- Compute all pairwise dot products and multiply by the logit scale.
- Use diagonal indices as positives for both image-to-text and text-to-image directions.
- Average the two cross-entropy losses and report retrieval accuracy.
Validation
Run python tests/test_contrastive_loss.py or the generated skill tree validator.
Limitations
This skill operates on supplied embeddings. It does not implement the visual or text transformer encoders.