SFace
SFace copied to clipboard
zhongyy
intraloss和interloss是否总是分别在减小和增大？
我最近在声音的一些识别工作上尝试使用SFace，发现在第一个epoch中的最初intraloss是在不停减少的，但是到后面的epoch甚至会逐渐增大，稳定在一个范围，但是错误率在下降，这是比较反优化目标，或者说这样的稳定就是把loss控制在一个合适的范围？ 这是我的训练loss，我使用的是adam+Cyclical Learning Rates学习率调整策略（学习率会周期从最低到最高），这和您在论文中使用的SGD+固定缩减lr不太相同。 epoch: 1, lr: 7.27e-03 - train loss: -6.43e+00, train intra_loss: -2.75e+01, train inter_loss: 21.07 - valid loss: -1.13e+01, valid intra_loss: -3.15e+01, valid inter_loss: 20.16, valid ErrorRate: 5.17e-02 epoch: 2, lr: 5.47e-03 - train loss: -1.18e+01, train intra_loss: -2.86e+01, train inter_loss: 16.84 - valid loss: -1.21e+01, valid intra_loss: -2.97e+01, valid inter_loss: 17.54, valid ErrorRate: 3.19e-02 epoch: 3, lr: 1.80e-03 - train loss: -1.23e+01, train intra_loss: -2.86e+01, train inter_loss: 16.27 - valid loss: -1.09e+01, valid intra_loss: -2.58e+01, valid inter_loss: 14.86, valid ErrorRate: 2.51e-02 epoch: 4, lr: 9.07e-03 - train loss: -1.16e+01, train intra_loss: -2.53e+01, train inter_loss: 13.62 - valid loss: -1.59e+01, valid intra_loss: -3.04e+01, valid inter_loss: 14.43, valid ErrorRate: 3.47e-02 epoch: 5, lr: 3.67e-03 - train loss: -1.07e+01, train intra_loss: -2.33e+01, train inter_loss: 12.57 - valid loss: -1.01e+01, valid intra_loss: -2.24e+01, valid inter_loss: 12.34, valid ErrorRate: 2.12e-02 epoch: 6, lr: 3.60e-03 - train loss: -1.18e+01, train intra_loss: -2.48e+01, train inter_loss: 13.01 - valid loss: -1.16e+01, valid intra_loss: -2.46e+01, valid inter_loss: 13.06, valid ErrorRate: 2.38e-02 epoch: 7, lr: 9.14e-03 - train loss: -1.00e+01, train intra_loss: -2.07e+01, train inter_loss: 10.69 - valid loss: -1.65e+01, valid intra_loss: -2.86e+01, valid inter_loss: 12.16, valid ErrorRate: 2.80e-02 epoch: 8, lr: 1.88e-03 - train loss: -1.03e+01, train intra_loss: -2.16e+01, train inter_loss: 11.32 - valid loss: -7.19e+00, valid intra_loss: -1.79e+01, valid inter_loss: 10.75, valid ErrorRate: 1.86e-02 epoch: 9, lr: 5.39e-03 - train loss: -1.10e+01, train intra_loss: -2.21e+01, train inter_loss: 11.10 - valid loss: -1.31e+01, valid intra_loss: -2.48e+01, valid inter_loss: 11.78, valid ErrorRate: 2.52e-02 epoch: 10, lr: 7.35e-03 - train loss: -8.92e+00, train intra_loss: -1.82e+01, train inter_loss: 9.27 - valid loss: -1.56e+01, valid intra_loss: -2.55e+01, valid inter_loss: 9.91, valid ErrorRate: 2.32e-02
