CLIP

Learning Transferable Visual Models From Natural Language Supervision

Alec Radford et al. · ICML 2021

Read paper · arXiv:2103.00020

Align image and text representations using a symmetric in-batch contrastive objective.

What you will implement

Implement feature normalization, temperature scaling, and bidirectional contrastive loss with a tiny paired-retrieval experiment.

• 25. Symmetric Contrastive Learning

  Implement feature normalization, learnable temperature, and the symmetric image-text contrastive objective.

  Implement contrastive_logits and clip_loss; paired-vector training is provided, without training full image/text encoders.