KunpengLi1994 / VSRN

PyTorch code for ICCV'19 paper "Visual Semantic Reasoning for Image-Text Matching"
293 stars 47 forks source link

Question about the similarity function used while training #12

Closed YAOYI626 closed 4 years ago

YAOYI626 commented 4 years ago

Hi KunPeng,

Appreciate for your excellent paper and code. I noticed that you state your similarity function as inner product in eq.(8)'s caption in your paper. But I find there is two similarity function setting in your code: measure ”order“ for subtraction function and measure "cos" for measure cosine similarity function, this line. So did your only use the inner product for testing? Or because the code you upload is not the final version?Cause my results with your code could not achieve that is stated in your paper.

Thanks a lot for your coming reply.

KunpengLi1994 commented 4 years ago

Hi,

For this part, we simply follow VSE++ and keep the default setting as this line, which is implemented by inner product as shown here.

To reproduce the results, please check the environment settings such as using Python 2.7 and Pytorch 0.4. You can also refer to issue #11 and issue #5 .

YAOYI626 commented 4 years ago

Hi KunPeng,

Thanks a lot. The "cosine" here misleads me.

And it seems that the version of PyTorch do have a big impact on the results, strange as it is, it is. I will close this issue.