search

janghyuncho / DECOLA

Code release for "Language-conditioned Detection Transformer"
https://arxiv.org/abs/2311.17902
82 stars 4 forks source link

Can not reproduce Direct zero-shot transfer to LVIS v1.0 results #9

Closed fushh closed 5 months ago

fushh commented 8 months ago

command: python train_net.py --num-gpus 4 --config-file configs/DECOLA_PHASE2_O365IN21k_CLIP_SwinT.yaml --eval-only MODEL.WEIGHTS DECOLA_PHASE2_O365IN21k_CLIP_SwinT.pth

there are several errors:

  1. there is no config named DECOLA_PHASE1_O365_CLIP_SwinT_4x.yaml
  2. there are some redundant config items: NO_FED_LOSS_LIST, ONLINE_LABELING

I try to fix the problems by renaming the file to DECOLA_PHASE1_O365_CLIP_SwinT.yaml and deleting the config items. It works but with poor inference results. image

Looking forward to your help.

janghyuncho commented 5 months ago

Hello, I am sorry for the late response! The evaluation configs have been updated. In short, AP_fixed metric needs some special setting here to be careful with. My apologies!