CLIP-KD: An Empirical Study of CLIP Model Distillation

TL;DR

This paper investigates various strategies for distilling smaller CLIP models from larger ones, demonstrating that simple feature mimicry and contrastive learning significantly enhance performance on image classification and retrieval tasks.

Contribution

It introduces multiple distillation strategies for CLIP models, highlighting the effectiveness of feature mimicry and contrastive learning in model compression.

Findings

Feature mimicry with MSE loss is highly effective.

Interactive contrastive learning improves student model performance.

Distilled models outperform original CLIP in zero-shot tasks.

Abstract

Contrastive Language-Image Pre-training (CLIP) has become a promising language-supervised visual pre-training framework. This paper aims to distill small CLIP models supervised by a large teacher CLIP model. We propose several distillation strategies, including relation, feature, gradient and contrastive paradigms, to examine the effectiveness of CLIP-Knowledge Distillation (KD). We show that a simple feature mimicry with Mean Squared Error loss works surprisingly well. Moreover, interactive contrastive learning across teacher and student encoders is also effective in performance improvement. We explain that the success of CLIP-KD can be attributed to maximizing the feature similarity between teacher and student. The unified method is applied to distill several student models trained on CC3M+12M. CLIP-KD improves student CLIP models consistently over zero-shot ImageNet classification and cross-modal retrieval benchmarks. When using ViT-L/14 pretrained on Laion-400M as the teacher, CLIP-KD achieves 57.5\% and 55.4\% zero-shot top-1 ImageNet accuracy over ViT-B/16 and ResNet-50, surpassing the original CLIP without KD by 20.5\% and 20.1\% margins, respectively. Our code is released on https://github.com/winycg/CLIP-KD.