Open zhangdandan-git opened 8 months ago
HRNet有四个分支，每个分支都是由basicblock组成，我将basicblock中的结构调整为了bn1->relu->conv1->bn2->relu->conv2的顺序，在稀疏化训练时，会出现bn1的grad为None,请问这是什么原因？ HRNet网络;
None module.stage4.2.branches.0.1.bn1
tensorC[ 1.5367e-09-3.9298e-099.5592e-092.7138e-095.1612e-091.1661e-08,5.5898e-09-1.5646e-09-9.1451e-09-8.7433e-092.4812e-08-1.1339e-086.6680e-10-1.7799e-082.5298e-091.1260e-083.3191e-09-1.4545e-097.2233e-105.3598e-092.8224e-099.2047e-09,5.3420e-10-6.6877e-11-6.0079e-097.2883e-09-9.4481e-098.5220e-101.9245e-086.5513e-09-1.1795e-08,-1.4566e-09]device-'cuda:0 module.stage4.2.branches.0.1.bn2
HRNet有四个分支，每个分支都是由basicblock组成，我将basicblock中的结构调整为了bn1->relu->conv1->bn2->relu->conv2的顺序，在稀疏化训练时，会出现bn1的grad为None,请问这是什么原因？ HRNet网络;
None module.stage4.2.branches.0.1.bn1
tensorC[ 1.5367e-09-3.9298e-099.5592e-092.7138e-095.1612e-091.1661e-08,5.5898e-09-1.5646e-09-9.1451e-09-8.7433e-092.4812e-08-1.1339e-086.6680e-10-1.7799e-082.5298e-091.1260e-083.3191e-09-1.4545e-097.2233e-105.3598e-092.8224e-099.2047e-09,5.3420e-10-6.6877e-11-6.0079e-097.2883e-09-9.4481e-098.5220e-101.9245e-086.5513e-09-1.1795e-08,-1.4566e-09]device-'cuda:0 module.stage4.2.branches.0.1.bn2