search

orrzohar / FOMO

Official Pytorch code for Open World Object Detection in the Era of Foundation Models
Apache License 2.0
57 stars 4 forks source link

Is there a training process of FOMO? #5

Closed synsin0 closed 9 months ago

synsin0 commented 9 months ago

Thanks for your great work. I'd like to ask whether there is a training process of FOMO, or we only need to load from pretrained parameters from OWL-ViT? It seems all start scripts are evaluation of different cases? I'm really confused with training.

synsin0 commented 9 months ago

I may understand that most configs may belong to BASE-ZS/BASE-FS in the paper so they do not need training, which config is for FOMO use?

orrzohar commented 9 months ago

Hi @synsin0, The ones for FOMO are run_rwd.sh and run_rwd_t2.sh.

There is some training process, where we finetune the attribute embeddings, however it is very light -- I could release the trained models if you would like. The training is light because it is done in the few-shot regime.

Best, Orr

synsin0 commented 9 months ago

I have successfully started the process of run_rwd.sh and run_rwd_t2.sh and finds the ultra-fast training process. Thanks for your timely response. I will close the issue.

orrzohar commented 9 months ago

Hi @synsin0, Great! Please note that we will be updating the main table soon as we had some variation between the published values and the reproduced ones -- mostly came from discovering a bug where we had duplicates of the same object in the few_shot_data.json file, leading to a lower few_shot value then expected. Best, Orr