Zero-Shot ImageNet

Key Insight

Zero-shot classification is CLIP's most surprising trick, and this project runs it on the full 1,000-class ImageNet benchmark: with no training on its labels, you write each class name as a prompt-template sentence like "a photo of a golden retriever", encode it with the text encoder, and label each picture with the class whose sentence sits closest (by cosine similarity) to its image embedding. What the project makes concrete is how much that wording moves the score: swapping a bare label for a natural caption, and averaging several templates (prompt ensembling), each lift accuracy by a measurable margin. The lesson is that with a model trained on enough image–text pairs, classification becomes retrieval — you are simply searching for the label whose description best fits the picture.