search

Artanic30 / HOICLIP

CVPR 2023 Accepted Paper HOICLIP: Efficient Knowledge Transfer for HOI Detection with Vision-Language Models
55 stars 6 forks source link

Is the training_free_enhancement part required during zero-shot training? #16

Closed sandipan211 closed 10 months ago

sandipan211 commented 10 months ago

Hi @Artanic30 ,

Kudos for this impressive work. I was trying to train your network using the command:

sh ./scripts/train_hico_rf_uc.sh

However, in all these scripts you have used the flag: "--training_free_enhancement_path". This turns your hyperparameter tuning step on at this line. However, the results from this tuning step do not have any effect on training. Instead, the resulting mAPs from the evaluation step done at the end (what you call as "Baseline performance") is returned and decides whether to not to update your training checkpoint.

So my question is: Is having the flag "--training_free_enhancement_path" really necessary while training? Or can it be omitted, since it is very time-consuming?

Artanic30 commented 10 months ago

Thanks for interested in our work. The --training_free_enhancement_path is only used in evaluation, so it can be omitted in training.