量子图像的奇怪自动编码器训练损失

Question

我正在训练自动编码器来分析量子系统的测量结果。测量结果具有形状（100,100,1），并且正在压缩到形状为（13、13、8）的潜在空间。我正在使用以下参数：

count = 40000
validation_split = 0.20
optimizer = 'adadelta'
learning_rate = 1.0
epochs = 50
batch_size = 128
loss = 'mean_absolute_error' # this is because the pixel values can be negative
shuffle = True

使用以下架构：

Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 100, 100, 1)       0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 100, 100, 16)      160       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 50, 50, 16)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 50, 50, 8)         1160      
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 25, 25, 8)         0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 25, 25, 8)         584       
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 13, 13, 8)         0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 13, 13, 8)         584       
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 26, 26, 8)         0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 26, 26, 8)         584       
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 52, 52, 8)         0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 50, 50, 16)        1168      
_________________________________________________________________
up_sampling2d_3 (UpSampling2 (None, 100, 100, 16)      0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 100, 100, 1)       145       
=================================================================
Total params: 4,385
Trainable params: 4,385
Non-trainable params: 0
_________________________________________________________________

这给出了以下训练曲线：

我以前从未见过这种行为的培训。 问题出在优化器还是网络本身的结构？