CLIP

Use this skill when a task involves OpenAI CLIP: zero-shot image classification, image-text similarity, prompt template design, tokenization, CLIP image/text embeddings, linear-probe evaluation, or troubleshooting CLIP model loading.

Fast Route

• For model names, clip.load, preprocessing, image-text logits, local checkpoints, JIT/device choices, and cache/download behavior, read model-loading-inference.
• For class-name prompt templates, clip.tokenize, context-length failures, zero-shot classifier weights, and prompt ensembling, read prompt-engineering.
• For batch image/text feature extraction, .npz embedding outputs, similarity/search prototypes, linear probes, and responsible evaluation limits, read feature-evaluation.
• For package-wide install/import, dependency, checkpoint download, dataset, and safety issues, read references/troubleshooting.md.
• For a no-download runtime check, run python scripts/validate_clip_runtime.py --json.

Use This Skill When

• The user names clip, OpenAI CLIP, ViT-B/32, RN50, clip.load, clip.tokenize, encode_image, or encode_text.
• The task asks for zero-shot image classification, image-to-text ranking, image/text embedding extraction, CLIP feature caching, or CLIP linear-probe evaluation.
• The task mentions prompt templates for vision-language classification, prompt ensembling, class taxonomy design, or CLIP context length 77.
• The task has errors involving CLIP checkpoint downloads, checksum/cache problems, CPU/CUDA tensor placement, TorchScript/JIT loading, or overlong tokenized text.

Avoid This Skill When

• The user is using OpenCLIP, Hugging Face transformers.CLIPModel, SigLIP, BLIP, or another CLIP-like implementation rather than this package.
• The user wants to train CLIP from scratch; this repository primarily exposes released model loading, inference, feature extraction, and evaluation examples.
• The request is about deploying CLIP in production, surveillance, facial recognition, or sensitive demographic classification without explicit research-only framing and safety review.
• The task is a general PyTorch, torchvision, or image-classification issue with no CLIP-specific API, model, tokenizer, or prompt signal.

Minimal Setup Pattern

Install PyTorch/torchvision first using the wheel or conda variant appropriate for the user's CPU/CUDA environment, then install CLIP and its lightweight text/tokenizer dependencies:

pip install ftfy regex tqdm
pip install git+https://github.com/openai/CLIP.git

Run a no-download import/tokenizer check before any checkpoint download:

python scripts/validate_clip_runtime.py --json

A healthy runtime imports clip, lists the nine released model names, tokenizes sample text to shape [n, 77], and reports that no checkpoint downloads were attempted.

Checkpoint and Dataset Policy

• clip.load("ViT-B/32") and other named models may download checkpoints if they are not cached; ask before triggering downloads in restricted, offline, or expensive environments.
• Use download_root= to point CLIP at a cache, or pass a local checkpoint path as name for strict offline workflows.
• README examples using CIFAR100 and dataset docs for Country211, Rendered SST2, and YFCC100M require network and potentially large downloads; treat them as opt-in verification or evaluation evidence, not default skill actions.
• The native JIT/eager consistency test can download every available model, so run it only when checkpoint downloads and time budget are explicitly allowed.

Core API Facts

• clip.available_models() returns the released model names: RN50, RN101, RN50x4, RN50x16, RN50x64, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336px.
• clip.load(name, device=..., jit=False, download_root=None) returns (model, preprocess) for a named model or local checkpoint path.
• clip.tokenize(texts, context_length=77, truncate=False) returns a [number_of_texts, context_length] token tensor and raises on overlong prompts unless truncate=True.
• Loaded models expose encode_image(image), encode_text(text), and model(image, text); the forward call returns image/text logit matrices derived from normalized cosine similarities.

Responsible Use

CLIP was released as a research model for studying robustness and zero-shot transfer. Model-card evidence warns that performance depends strongly on class taxonomy and prompts, that the model is English-oriented, and that deployment, surveillance, facial recognition, and sensitive demographic classification are out of scope without careful task-specific testing and safety review. Route evaluation and safety questions through feature-evaluation responsible guidance.