Lower results in local reproduction and a question about pretraining parameters

[sodaGH]

Thank you for your great work. I'm currently facing two issues while trying to reproduce the experiments locally:

1. I'm getting lower results locally compared to those reported in the paper, and my training script is as follows:

python train.py
--dataset mvtec
--train_data_path ./data/mvtec
--save_path ./exps/mvtec_EXP0_defaultCheck/
--exp_name EXP0_multiADs_default
--batch_size 8
--epoch 10
--pretrained "openai"
--model "ViT-L-14-336"

All parameters used the default settings. Is there any problem with my training script configuration?

2. I noticed a discrepancy between the paper and the actual code regarding the choice of pretraining parameters. Specifically, the paper's description doesn't match the implementation.
   Could you please clarify the exact experimental configuration used for training? Regarding the "model" and "pretrained" parameter combinations, did you use "VIT-L/14-336" + "openai" or "VIT-L/14" + "laion400m_e32"?

I find that in the pertained.py file, "laion400m_e32" only exists in the "VITL14" model, not "VITL14_336" model tag:

[hk-zh]

Hi, thank you for your question. We actually used VIT-L/14-336. I would suggest you use less training epoch instead of 10. 1 or 2 should be enough.

[sodaGH]

Hi, thank you for your question. We actually used VIT-L/14-336. I would suggest you use less training epoch instead of 10. 1 or 2 should be enough.

Thanks for your reply! So you suggest to use "VIT-L/14-336" + "openai" for the configuration of "model"+"pretrained", right?

[hk-zh]

Yes. I think we used this setting. And I think we get the best results on the first or second epoch. You could have a try.