Model card for clipkd-released-ViT-T-16-baseline-cc12m

ViT-T/16 baseline student trained on CC3M+12M (no KD). ImageNet acc: 30.55%

Model Details

• Architecture: ViT-T-16
• Method: CLIP-KD
• Training Data: CC3M+CC12M
• ImageNet Top-1: 30.6%
• Original Repo: https://github.com/winycg/CLIP-KD
• Paper: CLIP-KD: An Empirical Study of CLIP Model Distillation (CVPR 2024)

Model Usage

import open_clip

model, _, preprocess = open_clip.create_model_and_transforms(
    'hf-hub:fassabilf/clipkd-released-ViT-T-16-baseline-cc12m'
)
tokenizer = open_clip.get_tokenizer('hf-hub:fassabilf/clipkd-released-ViT-T-16-baseline-cc12m')