python - 为什么我的模型同时具有较低的 MAE 和较低的 R2 分数？

Question

我正在尝试构建 CNN 回归模型。输入数据为10年的卫星图像。

输入形状为[10, 256,256, 10]，表示[年份、图像形状、图像形状、 channel /频段]

模型的输出是0-1之间的数字，即图像中区域的百分比值。

这些是使用的参数

CHANNELS=5
BATCH_SIZE=16
INPUT_SHAPE=(10,IMG_SIZE,IMG_SIZE,CHANNELS)
SAMPLES=100
LR=1e-7
EPOCHES=10

我使用 Conv3D 层作为输入层，因为它能够向模型提供体积数据，并使用 Dense 层作为输出。

Model: sequential_FLATTEN_100_5_16_SGD_1e-07_30_v1
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv3d_3 (Conv3D)            (None, 10, 254, 254, 32)  1472      
_________________________________________________________________
max_pooling3d_3 (MaxPooling3 (None, 10, 127, 127, 32)  0         
_________________________________________________________________
conv3d_4 (Conv3D)            (None, 10, 125, 125, 64)  18496     
_________________________________________________________________
max_pooling3d_4 (MaxPooling3 (None, 10, 62, 62, 64)    0         
_________________________________________________________________
flatten_3 (Flatten)          (None, 2460160)           0         
_________________________________________________________________
dense_24 (Dense)             (None, 256)               629801216 
_________________________________________________________________
dense_25 (Dense)             (None, 1)                 257       
=================================================================
Total params: 629,821,441
Trainable params: 629,821,441
Non-trainable params: 0
_________________________________________________________________

该模型在训练集上给出以下分数:

mean_absolute_error: 0.09013315520024737
mean_squared_error: 0.11449361186977994
explained_variance_score: -0.2407465861253424
r2_score: -0.9382254392540899

在验证集上:

mean_absolute_error: 0.1923245317002776
mean_squared_error: 0.2579017795812263
explained_variance_score: -5.067052299015521
r2_score: -5.4177061135705475

我还尝试了以下不同的模型: 其中只有第一层是 Conv3D，其余都是 Dense 层

Model: "sequential_FLATTEN_100_5_16_Adam_1e-07_30_v1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv3d_1 (Conv3D)            (None, 4, 250, 250, 32)   54912     
_________________________________________________________________
max_pooling3d_1 (MaxPooling3 (None, 1, 83, 83, 32)     0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 220448)            0         
_________________________________________________________________
dense_16 (Dense)             (None, 512)               112869888 
_________________________________________________________________
dense_17 (Dense)             (None, 256)               131328    
_________________________________________________________________
dense_18 (Dense)             (None, 128)               32896     
_________________________________________________________________
dense_19 (Dense)             (None, 64)                8256      
_________________________________________________________________
dense_20 (Dense)             (None, 32)                2080      
_________________________________________________________________
dense_21 (Dense)             (None, 16)                528       
_________________________________________________________________
dense_22 (Dense)             (None, 8)                 136       
_________________________________________________________________
dense_23 (Dense)             (None, 1)                 9         
=================================================================
Total params: 113,100,033
Trainable params: 113,100,033
Non-trainable params: 0
_________________________________________________________________

这给了我在训练集上的以下分数:

mean_absolute_error: 0.08475626941395917
mean_squared_error: 0.1637630610914996
explained_variance_score: 0.19943303382780664
r2_score: 0.19214565669613703

在验证集上:

mean_absolute_error: 0.15135902269457854
mean_squared_error: 0.2650686092962602
explained_variance_score: -1.7471740284409094
r2_score: -1.7776585146674124

正如您所看到的，该模型的 MAE 和 MSE 非常低，但 R2 得分和解释方差得分同时也很低。

如何改进这些结果？ 此外，当样本量增加时，模型开始预测所有输入的相似值。

Answer 1

我只是注意到此类任务的参数数量如此之大。可能患有Vanishing or Exploding gradient 。尝试将特征提取器的维度减小得尽可能小。您也可以在两者之间应用 dropout 和正则化。