search

TencentYoutuResearch / CrowdCounting-P2PNet

The official codes for the ICCV2021 Oral presentation "Rethinking Counting and Localization in Crowds: A Purely Point-Based Framework"
Other
429 stars 112 forks source link

精度差 #19

Open FengJin-cv opened 2 years ago

FengJin-cv commented 2 years ago

你好，请问我使用ShanghaiTech_Crowd_Counting_Dataset/part_A_final数据进行训练，参数使用的是readme里面的参数，但是精度很差，是什么原因？ 2779][lr 0.0001000][11.72s]mae:329.14285714285717, mse:466.4266075383996, time:10.338678121566772, best mae:295.75274725274727metric/mae@2778: 329.14285714285717metric/mse@2778: 466.4266075383996loss/loss@2780: 0.04982551398712236loss/loss_ce@2780: 0.04982551398712236[ep 2780][lr 0.0001000][12.00s]

FengJin-cv commented 2 years ago

已解决。

Ssssssyz commented 2 years ago

已解决。

您好，我也遇到相似的问题了，请问您是怎么解决的呢？

Tomhardy13 commented 1 year ago

你好，为训练的过程会出现显存爆炸，哪怕设置bachsize=1都如此，请问你怎么解决的？