search

Wanggcong / Spatial-Temporal-Re-identification

[AAAI 2019] Spatial Temporal Re-identification
MIT License
384 stars 77 forks source link

Why is st-ReID + RE lower than in the paper? #8

Closed Adorablepet closed 5 years ago

Adorablepet commented 5 years ago

environment: pytorch 1.0/python 3.6/one-gpu I train st-Reid+RE model according to the parameters you code provide.(original parameter) st-Reid+RE map:87%, top1:96.7% st-Reid+RE + re-rank map: 94.5%, top1:96.8%

Wanggcong commented 5 years ago

Our baseline (appearance feature) is ~91%. It seems that your baseline is not so good?

Adorablepet commented 5 years ago

@Wanggcong My baseline is provided in your code. I have not modified it.I clone your code to train with original parameter.I'm confused.

Wanggcong commented 5 years ago

The reason may be attributed to different versions of pytorch. We will provide the trained model using this code.

Adorablepet commented 5 years ago

@Wanggcong Another problem is that when I use the mixed-precision to train the model, there will be a "gradient overflow". Is there a way to solve this?according to the way with fp16 layumi provides. val_loss: acc: 1.0