ZERO-SHOT Classification On SUN397

[xiaohoua]

I load ViT-H-14-378-quickgelu model huggingface,and try to do zero-shot classification on SUN397, but i get very low accurancy.
So I want to know how openclip does classification on SUN397。
I think the reasons for my mistake may be as follows：
1.I downloaded the model weights and categorized them on the data set without fine-tuning training. Maybe I need to do some fine-tuning? (I tested 50 images from each category.)
2.The text prompt I used was probably wrong, I used the image category name as the text prompt, as shown in the image below:

How do I do the classification code as predict_accurancy.py
Did openclip open source the code and details of the experiments on the SUN dataset? Where can I get the information?
Looking forward to reply!

[xiaohoua]

I get right result by CLIP_benchmark

[rwightman]

this is the main source I know with open source / config for reproducing various benchmarks for these models, though some of the original openai evals are still hard to reproduce exactly due to some missing split specifications