search

Li-Wanhua / GR2N

PyTorch implementation of Graph-Based Social Relation Reasoning (ECCV 2020)
16 stars 3 forks source link

Problems on PISC #3

Closed vampirely closed 2 years ago

vampirely commented 2 years ago

Hello, the accuracy of the results I obtained by running on the PISC data set is only about 60%. I set the training parameters by installing the parameters in the article. May I ask which corresponding parameters need to be set in the PISC test to reproduce the results.

Li-Wanhua commented 2 years ago

Hi, thanks for the interest. PISC dataset has two settings: Coarse relationships and Fine relationships. So which one are you referring to? But something to notice includes 1) the metric is mAP NOT accuracy for the PISC dataset. 2) we adopted the reweighting strategy for fine-grained relationship recognition on the PISC dataset. 3) The config file for fine-grained relationship recognition on the PISC dataset is listed as below:

CUDA_VISIBLE_DEVICES=1,2,3 python SR_train.py --batch-size=32 \ --test-batch-size=16 --max-person=8 --image-size=448 \ --max-epochs=10 --lr=0.00001 --fc-lr=0.00001 \ --save-model='./Domain_Model/' \ --loss-weight-path='./relation_split/loss_weight.npy' \ --images-root='/home/disk0/Social_relation/PISC/image/' \ --train-file-pre='./relation_split/relation_train' \ --test-file-pre='./relation_split/relation_test' \ --num-workers=2 --num-classes=6 --time-steps=1 \ --manualSeed=768

Hope this can help you.

Li-Wanhua commented 2 years ago

I think it's best to use English. If you find it inconvenient to communicate in English, you can contact me in Chinese by email. My email is li-wh17@mails.tsinghua.edu.cn