about Accuracy

[ynu-yangpeng]

Sorry, the following is the result of my reproduction of your code on NVIDIA GeForce 3090. I ran 50 epochs on Market-1501.
Mean AP: 58.5%
CMC Scores:
top-1 78.4%
top-5 85.6%
top-10 87.9%
Is my configuration errors?

[yoonkicho]

Hi, thank your for your interest.
Did you train a model with a single GPU?
Unsupervised re-ID training is sensitive to the number of GPUs and batch sizes, and we trained a model with 4 GPUs.
I think you can get similar results in the paper with training with 4 GPUs.

[KleinYuan]

@yoonkicho how come the number of GPUs will impact the performances ?

[helloworld-wu]

@YpSokm I have the same problem with you, and I set 4 GPUs to train the model, which can get the much better performances. But I do not know why

[RominaZakerian]

Hi, thank your for your interest. Did you train a model with a single GPU? Unsupervised re-ID training is sensitive to the number of GPUs and batch sizes, and we trained a model with 4 GPUs. I think you can get similar results in the paper with training with 4 GPUs.

Hi. sorry to bother you. I have a question. how can the number of GPUs impact the performance?

[newbie521]

确实是的，使用两个3090运行代码结果是63.3%，而四个3090是81.2%，但是不知道这是为什么

[qingsonghu08]

The source code configuration file sets the batchsize to 64. Do I need to change this to 256 when using 4 4090 sheets?