About the experimental reproduction results

hbdat / iccv21_relational_direction

Interaction Compass: Multi-Label Zero-Shot Learning of Human-Object Interactions via Spatial Relations @ ICCV21

MIT License

13 stars 2 forks source link

About the experimental reproduction results #2

Open wlsh1up opened 2 years ago

wlsh1up commented 2 years ago

Hi hdbat: thanks for your nice work and I am very interested in this paper. However, after the code is finished trained, I get the following results. In 1A setting.
"detail_human.csv & model_final.pt & partition_F1_3.csv & ranking_k_10.csv & stats.csv detail_hoi.csv & detail_object.csv & partition_AP.csv & partition_F1_5.csv & ranking_k_7.csv"

This is inconsistent with the results reported in the paper. Can you tell me how to reproduce it Best

hbdat commented 2 years ago

Hi @wlsh1up

Thanks for your interests in the project. Could you provide more information about the training setting you use?

Best, Dat

wlsh1up commented 2 years ago

Sorry for your late reply. I've read your code carefully these days. and i am new to this field, so I hope my questions are not irrelevant.
The complete reproduction results are shown below

111

the training setting I use is HICO which iniclud 38116 + 9658 = 47774 pkl.gz in /data/hico_20160224_det/features_pad

I only changed epoch to 15, Other hype-parameters are not changed.

hbdat commented 2 years ago

Hi @wlsh1up

I am not sure what the issue is. Have you tried to reproduce the results using the pretrained models? I have included the instruction for evaluating on pretrained models in the README.

Please let me know what the results on your side are for these pretrained models.

Best, Dat