Pretrained Weights

This release tag is being used to host weights for various models trained with this codebase.

NOTE: The one included metric, zero-shot top-1 on ImageNet-1k does capture the full characteristics of the given pretrained weights. Evaluation on a broader set of zero-shot and validation tasks is required for a full comparison.

model	dataset	weights	In1k zero-shot top-1
RN50	CC12M	rn50-quickgelu-cc12m	36.45
RN50	YFCC15M	rn50-quickgelu-yfcc15m	32.73
RN101	YFCC15M	rn101-quickgelu-yfcc15m	34.86
ViT-B-32	LAION-400M	vit_b_32-quickgelu-laion400m_e31	62.96
ViT-B-32	LAION-400M	vit_b_32-quickgelu-laion400m_e32	62.94
ViT-B-32	LAION-2B	vit_b_32-laion2b_e16	65.62
ViT-B-16	LAION-400M	vit_b_16-laion400m_e31	66.98
ViT-B-16	LAION-400M	vit_b_16-laion400m_e32	67.07
ViT-B-16-plus-240	LAION-400M	vit_b_16-laion400m_e31	69.06
ViT-B-16-plus-240	LAION-400M	vit_b_16-laion400m_e32	69.21
ViT-L-14	LAION-400M	vit_b_14-laion400m_e31	72.70
ViT-L-14	LAION-400M	vit_b_14-laion400m_e32	72.77

Initial release.

Welcome to the initial release of open_clip, an open source implementation of OpenAI's CLIP (Contrastive Language-Image Pre-training).

The goal of this repository is to enable training models with contrastive image-text supervision, and to investigate their properties such as robustness to distribution shift. Our starting point is an implementation of CLIP that matches the accuracy of the original CLIP models when trained on the same dataset.