RangiLyu / nanodet

NanoDet-Plus⚡Super fast and lightweight anchor-free object detection model. 🔥Only 980 KB(int8) / 1.8MB (fp16) and run 97FPS on cellphone🔥
Apache License 2.0
5.63k stars 1.03k forks source link

为什么显存占用这么高Why is the memory usage so high #538

Open BucolicLife opened 8 months ago

BucolicLife commented 8 months ago

INFO:NanoDet:Train|Epoch1/100|Iter200(201/3000)| mem:1.4G| lr:4.00e-04| loss_qfl:0.3616| loss_bbox:1.0301| loss_dfl:0.4526| aux_loss_qfl:0.4193| aux_loss_bbox:0.5471| aux_loss_dfl:0.2657| [NanoDet][11-01 10:05:29]INFO:Train|Epoch1/100|Iter250(251/3000)| mem:15.3G| lr:5.00e-04| loss_qfl:0.8047| loss_bbox:0.8445| loss_dfl:0.3974| aux_loss_qfl:0.6471| aux_loss_bbox:0.7851| aux_loss_dfl:0.3845| INFO:NanoDet:Train|Epoch1/100|Iter250(251/3000)| mem:15.3G| lr:5.00e-04| loss_qfl:0.8047| loss_bbox:0.8445| loss_dfl:0.3974| aux_loss_qfl:0.6471| aux_loss_bbox:0.7851| aux_loss_dfl:0.3845| 同一个epoch显存从1.4G猛涨到15.3G 训练集3000条样本,验证集300条,图像大小[480,640] In the same epoch, the video memory jumped from 1.4G to 15.3G. There are 3000 samples in the training set, 300 samples in the verification set, and the image size is [480, 640]