About the problem that the loss value is zero in trainning

[chenyu19880302]

hello! when I use your code to train the model,the initial loss is: .loss == 0.0064889, root joint loss == 0.0034980, body joint loss == 0.0029910 for epoch 0 step 100 ....loss == 0.0000841, root joint loss == 0.0000399, body joint loss == 0.0000442 for epoch 0 step 200 ....loss == 0.0001365, root joint loss == 0.0000708, body joint loss == 0.0000657 for epoch 0 step 300

and later the loss is : for epoch 0 step 13500 ....loss == 0.0000063, root joint loss == 0.0000063, body joint loss == 0.0000000 for epoch 0 step 13600 ....loss == 0.0000083, root joint loss == 0.0000083, body joint loss == 0.0000000 for epoch 0 step 13700 ....loss == 0.0000066, root joint loss == 0.0000066, body joint loss == 0.0000000 for epoch 0 step 13800 ....loss == 0.0000051, root joint loss == 0.0000051, body joint loss == 0.0000000 for epoch 0 step 13900 ....loss == 0.0000064, root joint loss == 0.0000062, body joint loss == 0.0000002 for epoch 0 step 14000 ....loss == 0.0000046, root joint loss == 0.0000046, body joint loss == 0.0000000

I don't the reason,do you encounter the problem？how did you solve it ?

[murdockhou]

Hi, thanks for your concern. Actually based on paper, the author used L2 loss for center point and SmoothL1 Loss for SPR regression. I tried to use these in code but get very smaller value as you said. I just change the way calculating loss and test for some steps but no very well result got. I'll keep trying this and hope that you can also do help for this.

[thancaocuong]

@chenyu19880302 I've found that when use smooth L1 loss for dense displacements regression, Losses are quite small due to multiplying with dense_weight. And then, the loss values are averaged. It make loss value quite small