rank1 rate 79.3% not same with the paper.

dapengchen123 / video_reid

We plan to release our code in CVPR 2018 "Video Person Re-identification with Competitive Snippet-similarity Aggregation and Co-attentive Snippet Embedding"

41 stars 13 forks source link

rank1 rate 79.3% not same with the paper. #10

Open ccq195 opened 5 years ago

ccq195 commented 5 years ago

Hi,Why i get the rank1 rate is 79.3% on ililds dataset?But the mAP is 87.2%. The rank1 rate can not get the result of the paper.

ccq195 commented 5 years ago

Hi,when i change the 'cnn_model= torch.nn.DataParallel(cnn_model).to(device)' to 'cnn_model= cnn_model.to(device)', the rank1 rate is 85.3%. But i don't know the reason. Do you konw the reason?

dapengchen123 commented 5 years ago

Actually, I don't know why. It may depend on your devices such as how many cards you use and the version of the pytorch. Anyway, it is good that it works.

ccq195 commented 5 years ago

Thank you for your reply。