search

zhangxi1997 / ECMR-VCR

The coder for the paper "Explicit Cross-Modal Representation Learning for Visual Commonsense Reasoning"
4 stars 0 forks source link

About the pre-trained attribute capturing visual representations #1

Open Pefect96 opened 2 years ago

Pefect96 commented 2 years ago

Great job!! I want to train your model from begining, but I have some questions about the attribute_feature, Can you provide visual features of the VCR dataset? Please provide me with the link or tell me how to generate it in detail. I cannot get it from bottom-up-attention repo at present. I'm very grateful to you.

zhangxi1997 commented 2 years ago

Thanks, please refer to tab-vcr to download the visual features. Note that we use the attribute features instead of the new_tag features.

Pefect96 commented 2 years ago

Get it! But when I run _python my_train.py -params CMR_sGCN_attr/default.json -folder saves/flagship_answer -plot saves/flagshipanswer according to your Readme in models, I meet the following question, i do not know the solution. I'm very grateful to you. image

zhangxi1997 commented 2 years ago

The bug has been solved and the code has been updated. Specifically, I moved the "init.py" into the "models" folder.

vicky-abr commented 2 years ago

Great job!! I want to train your model from begining, but I have some questions about the attribute_feature, Can you provide visual features of the VCR dataset? Please provide me with the link or tell me how to generate it in detail. I cannot get it from bottom-up-attention repo at present. I'm very grateful to you.

Get it! But when I run _python my_train.py -params CMR_sGCN_attr/default.json -folder saves/flagship_answer -plot saves/flagshipanswer according to your Readme in models, I meet the following question, i do not know the solution. I'm very grateful to you. image

Could you please tell me if the code runs successfully? Where did you get the .h5 file like span{self.mode}{self.split}_q.h5?