jindongwang / tlbook-code Goto Github PK

在您的这本书提供的models.py的code中，在forward函数中，source和target是先过了classifier_layer,再通过bottleneck_layer
但是在DeepDA提供的code中，是先通过bottleneck_layer再通过classifier_layer
想问下以上两种方式哪种是正确的，抑或是都可以？

东哥好，我使用DANN训练自己的问题时，其域分类精度一开始就是50%左右（transfer loss大约在0.7），并且一直维持在这个值，未曾有变化，感觉域分类器并没有很好的work，已调整过学习率、transfer_loss权重等参数，但均无改变。不知到应该怎么调整代码？

部分输出如下：
Epoch [1/10] - T_Loss:2.7179, R_loss:2.016,D_loss:0.702,D_acc:50.117%
Val_Loss(MSE):2.262,Val_MAE:1.242,Val_R2:0.633
lr [0.0001]
Epoch [2/10] - T_Loss:2.5103, R_loss:1.803,D_loss:0.708,D_acc:45.898%
Val_Loss(MSE):1.774,Val_MAE:0.929,Val_R2:0.713
lr [0.0001]
Epoch [3/10] - T_Loss:1.9520, R_loss:1.252,D_loss:0.700,D_acc:49.336%
Val_Loss(MSE):1.301,Val_MAE:0.828,Val_R2:0.789
lr [0.0001]
Epoch [4/10] - T_Loss:1.9570, R_loss:1.255,D_loss:0.702,D_acc:48.945%
Val_Loss(MSE):1.319,Val_MAE:0.777,Val_R2:0.786
lr [0.0001]
Epoch [5/10] - T_Loss:1.6281, R_loss:0.927,D_loss:0.701,D_acc:48.398%
Val_Loss(MSE):1.368,Val_MAE:0.901,Val_R2:0.778
lr [0.0001]
Epoch [6/10] - T_Loss:1.4805, R_loss:0.779,D_loss:0.701,D_acc:47.474%
Val_Loss(MSE):1.161,Val_MAE:0.759,Val_R2:0.812
lr [0.0001]
Epoch [7/10] - T_Loss:1.3071, R_loss:0.607,D_loss:0.701,D_acc:48.737%
Val_Loss(MSE):1.031,Val_MAE:0.697,Val_R2:0.833
lr [0.0001]
Epoch [8/10] - T_Loss:1.3438, R_loss:0.643,D_loss:0.701,D_acc:46.276%
Val_Loss(MSE):1.019,Val_MAE:0.691,Val_R2:0.835
lr [0.0001]
Epoch [9/10] - T_Loss:1.2132, R_loss:0.513,D_loss:0.700,D_acc:47.826%
Val_Loss(MSE):1.019,Val_MAE:0.650,Val_R2:0.835
lr [0.0001]
Epoch [10/10] - T_Loss:1.1207, R_loss:0.421,D_loss:0.700,D_acc:48.333%
Val_Loss(MSE):0.911,Val_MAE:0.627,Val_R2:0.852

作者您好！近期读到第二版85页“带有概率密度比的SVM”公式第二项β前为什么有负号，在“Density ratio estimation in support vector machine for better generalization: study on direct marketing prediction”论文中该项没有负号，不是很理解
期待您的回复！