jkli1998 / DRM

Code for paper 'Leveraging Predicate and Triplet Learning for Scene Graph Generation'. (CVPR 2024)
MIT License
22 stars 0 forks source link

About stage 2 evaluation #5

Open kekekek111 opened 1 week ago

kekekek111 commented 1 week ago

Thank you for your great work! I was exploring the code and encountered some errors when I tried to evaluate the provided stage2 model (vg_stage2_predcls). I seems that I would need some files under folder ./experiments/vg_stage2_predcls/infer_train_feat. Could you please provide these files or give some instructions on how to generate them?

Thanks again!

jkli1998 commented 1 week ago

Thanks for your interest. Acturally, the infer_train_feat is generated by the Extract features for calculating distribution: of the training command. One way to directly evaluate the stage2 is to comment out the relevant loading code (line 189-191 in DRM/maskrcnn_benchmark/modeling/roi_heads/relation_head/roi_relation_predictors.py). Another way is to load the weights of stage1 and generate the corresponding files by runing the Extract features for calculating distribution: step. Note that the infer_train_feat is only used in training.