为什么卷积自编码器中的训练损失和验证损失没有显著下降？训练数据的维度为10496x1024，而CAE使用keras训练32x32大小的图像块。我已经尝试了l2regularization，但效果不明显。我训练了20个周期。还有哪些其他可行的方法？

输出如下：

Epoch 1/20 10496/10496 [========] – 52s – loss: 0.4029 – val_loss: 0.3821

Epoch 2/20 10496/10496 [========] – 52s – loss: 0.3825 – val_loss: 0.3784

Epoch 3/20 10496/10496 [=======] – 52s – loss: 0.3802 – val_loss: 0.3772

Epoch 4/20 10496/10496 [=======] – 51s – loss: 0.3789 – val_loss: 0.3757

Epoch 5/20 10496/10496 [=======] – 52s – loss: 0.3778 – val_loss: 0.3752

Epoch 6/20 10496/10496 [=======] – 51s – loss: 0.3770 – val_loss: 0.3743

Epoch 7/20 10496/10496 [=======] – 54s – loss: 0.3763 – val_loss: 0.3744

Epoch 8/20 10496/10496 [=======] – 51s – loss: 0.3758 – val_loss: 0.3735

Epoch 9/20 10496/10496 [=======] – 51s – loss: 0.3754 – val_loss: 0.3731

Epoch 10/20 10496/10496 [=======] – 51s – loss: 0.3748 – val_loss: 0.3739

Epoch 11/20 10496/10496 [=======] – 51s – loss: 0.3745 – val_loss: 0.3729

Epoch 12/20 10496/10496 [=======] – 54s – loss: 0.3741 – val_loss: 0.3723

Epoch 13/20 10496/10496 [=======] – 51s – loss: 0.3736 – val_loss: 0.3718

Epoch 14/20 10496/10496 [=======] – 52s – loss: 0.3733 – val_loss: 0.3716

Epoch 15/20 10496/10496 [=======] – 52s – loss: 0.3731 – val_loss: 0.3717

Epoch 16/20 10496/10496 [=======] – 51s – loss: 0.3728 – val_loss: 0.3712

Epoch 17/20 10496/10496 [=======] – 49s – loss: 0.3725 – val_loss: 0.3709

Epoch 18/20 10496/10496 [=======] – 36s – loss: 0.3723 – val_loss: 0.3710

Epoch 19/20 10496/10496 [=======] – 37s – loss: 0.3721 – val_loss: 0.3708

Epoch 20/20 10496/10496 ========] – 37s – loss: 0.3720 – val_loss: 0.3704

回答：

你的网络仍在学习中，到第20个周期时并没有明显减速。你可以尝试使用更高的学习率，并在有足够数据的情况下使用更多的周期和提前停止方法。这种方法也可以结合正则化方法和k折交叉验证来使用。