EVA a new state-of-the-art in zero-shot on ImageNet 83.26.

By modifying the inference process and utilizing the EVA02-CLIP-bigE-14-plus model, we achieved a zero-shot accuracy of 83.26%. Consistent improvements were also observed across other model scales.I hope we can continue our conversation further.

Jun 10 '23 14:06 pandaupc

Hi @pandaupc,

We are very happy to hear that. It's a really big improvement. We are wondering if you have a plan to release your methods.

Jun 11 '23 14:06 Quan-Sun

How to achieve 83.26%? I got only 82.1% with CuPL prompts.

Jun 14 '23 08:06 wztxywww123

We have submitted our method to NeurIPS 2023, and we are currently conducting more experiments based on EvaClip, hoping to achieve better results.

Jun 16 '23 08:06 pandaupc

We referenced CUPL, but we also employed a variety of other techniques.

Jun 16 '23 08:06 pandaupc

We have submitted our method to NeurIPS 2023, and we are currently conducting more experiments based on EvaClip, hoping to achieve better results.

Hi, @pandaupc Did you release the paper?

Dec 26 '23 05:12 laserwave