search

Gaoyiminggithub / Graphonomy

Graphonomy: Universal Human Parsing via Graph Transfer Learning
MIT License
295 stars 68 forks source link

Question about trained model #14

Closed chmxu closed 5 years ago

chmxu commented 5 years ago

Thanks for sharing this repo. I have downloaded the checkpoint 'inference.pth' in the Model section in README. And when i load the model weight i find that the 'target_featuremap_2_graph.pre_fea' and 'target_featuremap_2_graph.weight', which are contained in the feature to graph projection module, are all 0, which i think is not so reasonable. Am i having any misunderstanding or maybe I have downloaded a wrong model?

Gaoyiminggithub commented 5 years ago

Thanks for your attention. I have checked the model in README.md in which some modules are 0. The reason is that when I re-trained our models based on the code in Github, I commented the reset_parameters() function. Without the double-check for the model, we are very sorry for the inconvenience. And I start training with the correct version of the code. Thanks for your kindly help.

chmxu commented 5 years ago

Thanks. Hope to see the update version soon.

Gaoyiminggithub commented 5 years ago

Hi @loadder Thanks for your attention. The correct pretrained model of CIHP is in here, and the model of PASCAL is in here. We are so sorry for the inconvenience.

yarin05 commented 5 years ago

@Gaoyiminggithub Thank you for keeping support and maintenance the project. Should I expect changes also in the 'Universal' model?

chmxu commented 5 years ago

Thanks!