python - val损失没有减少

Question

我在 Windows 10 上使用 Keras 2.07、Python 3.5、Tensorflow 1.3.0

我正在使用全连接网络进行视频编码来测试论文帧内预测中使用的架构。

我希望将其用于我自己的数据。

我使用了测试数据，我认为它会很快收敛。

我在模型中使用的学习率为 0.1。那么我的val损失并没有减少。

有人可以查看或尝试此代码吗？我是否假设错误、编码错误或不耐烦？谢谢

def multi_input_model():
   input1_ = Input(shape=(4,20,1), name='input1')
   input2_ = Input(shape=(16,4,1), name='input2')

   x1 = Flatten()(input1_)
   x2 = Flatten()(input2_)
   x = Concatenate()([x1, x2])

   for i in range(3):
      #x = Dropout(0.3)(x)
      x = Dense(1024,
                kernel_initializer = RandomNormal(stddev=std),
                use_bias=True,
                bias_initializer = 'zeros', 
                activation='relu'
                )(x)

      #x = Dropout(0.3)(x)
      x = Dense(64,
                kernel_initializer = RandomNormal(stddev=std),
                use_bias=True,
                bias_initializer = 'zeros',
                )(x)

   output_ = Reshape((8,8,1), name='output')(x)

   model = Model(inputs=[input1_, input2_], outputs=[output_])
   model.summary()

return model


sgd = SGD(lr = 0.1, momentum=0.9)

model = multi_input_model()
model.compile(optimizer=sgd,
              loss='mean_squared_error',
              metrics=[mse]
              )

模型摘要:

Using TensorFlow backend.
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input1 (InputLayer)             (None, 4, 20, 1)     0                                            
__________________________________________________________________________________________________
input2 (InputLayer)             (None, 16, 4, 1)     0                                            
__________________________________________________________________________________________________
flatten_1 (Flatten)             (None, 80)           0           input1[0][0]                     
__________________________________________________________________________________________________
flatten_2 (Flatten)             (None, 64)           0           input2[0][0]                     
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 144)          0           flatten_1[0][0]                  
                                                                 flatten_2[0][0]                  
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 1024)         148480      concatenate_1[0][0]              
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 1024)         1049600     dense_1[0][0]                    
__________________________________________________________________________________________________
dense_3 (Dense)                 (None, 1024)         1049600     dense_2[0][0]                    
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 64)           65600       dense_3[0][0]                    
__________________________________________________________________________________________________
output (Reshape)                (None, 8, 8, 1)      0           dense_4[0][0]                    
==================================================================================================
Total params: 2,313,280
Trainable params: 2,313,280
Non-trainable params: 0
__________________________________________________________________________________________________

训练回溯

Train on 985373 samples, validate on 246344 samples
Epoch 1/100
985373/985373 [==============================] - 7s 7us/step - loss: 0.0054 - mse: 353.9386 - val_loss: 0.0087 - val_mse: 566.5364
Lr :  0.08
Epoch 2/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0044 - mse: 288.5897 - val_loss: 0.0082 - val_mse: 534.4153
Lr :  0.08
Epoch 3/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0042 - mse: 270.7345 - val_loss: 0.0080 - val_mse: 517.2601
Lr :  0.08
Epoch 4/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0040 - mse: 259.6213 - val_loss: 0.0078 - val_mse: 504.8340
Lr :  0.08
Epoch 5/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0039 - mse: 251.3669 - val_loss: 0.0076 - val_mse: 495.0704
Lr :  0.08
Epoch 6/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0038 - mse: 244.7449 - val_loss: 0.0075 - val_mse: 486.6413
Lr :  0.08
Epoch 7/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0037 - mse: 239.2320 - val_loss: 0.0074 - val_mse: 480.2631
Lr :  0.08
Epoch 8/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0036 - mse: 234.5702 - val_loss: 0.0073 - val_mse: 473.3974
Lr :  0.08
Epoch 9/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0035 - mse: 230.5504 - val_loss: 0.0072 - val_mse: 468.4981
Lr :  0.08
Epoch 10/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0035 - mse: 227.0740 - val_loss: 0.0071 - val_mse: 463.1125
Lr :  0.08
Epoch 11/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0034 - mse: 224.0050 - val_loss: 0.0071 - val_mse: 459.0103
Lr :  0.08
Epoch 12/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0034 - mse: 221.3146 - val_loss: 0.0070 - val_mse: 455.0988
Lr :  0.08
Epoch 13/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0034 - mse: 218.9012 - val_loss: 0.0069 - val_mse: 451.7273
Lr :  0.08
Epoch 14/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0033 - mse: 216.6984 - val_loss: 0.0069 - val_mse: 448.4461
Lr :  0.08
Epoch 15/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0033 - mse: 214.7047 - val_loss: 0.0069 - val_mse: 446.2414
Lr :  0.08
Epoch 16/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0033 - mse: 212.8751 - val_loss: 0.0068 - val_mse: 443.6305
Lr :  0.08
Epoch 17/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0032 - mse: 211.1775 - val_loss: 0.0068 - val_mse: 441.6251
Lr :  0.08
Epoch 18/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0032 - mse: 209.5941 - val_loss: 0.0068 - val_mse: 439.3194
Lr :  0.08
Epoch 19/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0032 - mse: 208.1229 - val_loss: 0.0067 - val_mse: 436.9524
Lr :  0.08
Epoch 20/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0032 - mse: 206.7133 - val_loss: 0.0067 - val_mse: 435.1670
Lr :  0.08
Epoch 21/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0032 - mse: 205.4114 - val_loss: 0.0067 - val_mse: 432.8971
Lr :  0.08
Epoch 22/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 204.1449 - val_loss: 0.0066 - val_mse: 431.2357
Lr :  0.08
Epoch 23/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 202.9737 - val_loss: 0.0066 - val_mse: 430.8359
Lr :  0.08
Epoch 24/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 201.8537 - val_loss: 0.0066 - val_mse: 429.2760
Lr :  0.08
Epoch 25/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 200.7699 - val_loss: 0.0066 - val_mse: 428.2598
Lr :  0.08
Epoch 26/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 199.7556 - val_loss: 0.0066 - val_mse: 426.2759
Lr :  0.08
Epoch 27/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0031 - mse: 198.7569 - val_loss: 0.0065 - val_mse: 425.0138
Lr :  0.08
Epoch 28/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 197.8049 - val_loss: 0.0065 - val_mse: 423.5451
Lr :  0.08
Epoch 29/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 196.8937 - val_loss: 0.0065 - val_mse: 422.2164
Lr :  0.08
Epoch 30/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 196.0350 - val_loss: 0.0065 - val_mse: 421.1507
Lr :  0.08
Epoch 31/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 195.1807 - val_loss: 0.0065 - val_mse: 420.2791
Lr :  0.08
Epoch 32/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 194.3681 - val_loss: 0.0065 - val_mse: 420.0271
Lr :  0.08
Epoch 33/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 193.5861 - val_loss: 0.0064 - val_mse: 418.1573
Lr :  0.08
Epoch 34/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 192.8048 - val_loss: 0.0064 - val_mse: 417.7967
Lr :  0.08
Epoch 35/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0030 - mse: 192.0708 - val_loss: 0.0064 - val_mse: 416.3381
Lr :  0.08
Epoch 36/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 191.3404 - val_loss: 0.0064 - val_mse: 416.3695
Lr :  0.08
Epoch 37/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 190.6411 - val_loss: 0.0064 - val_mse: 415.9791
Lr :  0.08
Epoch 38/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 189.9464 - val_loss: 0.0064 - val_mse: 414.0931
Lr :  0.08
Epoch 39/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 189.2725 - val_loss: 0.0064 - val_mse: 413.8717
Lr :  0.08
Epoch 40/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 188.6250 - val_loss: 0.0064 - val_mse: 413.0042
Lr :  0.08
Epoch 41/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 187.9844 - val_loss: 0.0064 - val_mse: 413.0950
Lr :  0.08
Epoch 42/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 187.3532 - val_loss: 0.0063 - val_mse: 412.4408
Lr :  0.08
Epoch 43/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 186.7512 - val_loss: 0.0063 - val_mse: 411.1885
Lr :  0.08
Epoch 44/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 186.1417 - val_loss: 0.0063 - val_mse: 410.7527
Lr :  0.08
Epoch 45/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0029 - mse: 185.5806 - val_loss: 0.0063 - val_mse: 409.3184
Lr :  0.08
Epoch 46/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 185.0026 - val_loss: 0.0063 - val_mse: 410.3592
Lr :  0.08
Epoch 47/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 184.4485 - val_loss: 0.0063 - val_mse: 409.0613
Lr :  0.08
Epoch 48/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 183.8902 - val_loss: 0.0063 - val_mse: 409.2569
Lr :  0.08
Epoch 49/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 183.3491 - val_loss: 0.0063 - val_mse: 408.1287
Lr :  0.08
Epoch 50/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 182.8373 - val_loss: 0.0063 - val_mse: 407.0794
Lr :  0.08
Epoch 51/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 182.3068 - val_loss: 0.0063 - val_mse: 407.4011
Lr :  0.08
Epoch 52/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 181.7958 - val_loss: 0.0062 - val_mse: 405.9963
Lr :  0.08
Epoch 53/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 181.2954 - val_loss: 0.0062 - val_mse: 406.3628
Lr :  0.08
Epoch 54/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 180.8150 - val_loss: 0.0062 - val_mse: 405.7962
Lr :  0.08
Epoch 55/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 180.3107 - val_loss: 0.0062 - val_mse: 405.0873
Lr :  0.08
Epoch 56/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 179.8412 - val_loss: 0.0062 - val_mse: 405.1387
Lr :  0.08
Epoch 57/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 179.3717 - val_loss: 0.0062 - val_mse: 404.4157
Lr :  0.08
Epoch 58/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0028 - mse: 178.9016 - val_loss: 0.0062 - val_mse: 403.5622
Lr :  0.08
Epoch 59/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 178.4647 - val_loss: 0.0062 - val_mse: 403.3207
Lr :  0.08
Epoch 60/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 177.9997 - val_loss: 0.0062 - val_mse: 403.5156
Lr :  0.08
Epoch 61/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 177.5587 - val_loss: 0.0062 - val_mse: 403.2921
Lr :  0.08
Epoch 62/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 177.1332 - val_loss: 0.0062 - val_mse: 402.8525
Lr :  0.08
Epoch 63/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 176.6831 - val_loss: 0.0062 - val_mse: 402.3887
Lr :  0.08
Epoch 64/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 176.2728 - val_loss: 0.0062 - val_mse: 401.6309
Lr :  0.08
Epoch 65/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 175.8403 - val_loss: 0.0062 - val_mse: 401.4650
Lr :  0.08
Epoch 66/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 175.4337 - val_loss: 0.0062 - val_mse: 401.6886
Lr :  0.08
Epoch 67/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 175.0153 - val_loss: 0.0062 - val_mse: 401.1379
Lr :  0.08
Epoch 68/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 174.6212 - val_loss: 0.0062 - val_mse: 401.4452
Lr :  0.08
Epoch 69/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 174.2141 - val_loss: 0.0062 - val_mse: 401.5032
Lr :  0.08
Epoch 70/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 173.8265 - val_loss: 0.0062 - val_mse: 400.2583
Lr :  0.08
Epoch 71/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 173.4480 - val_loss: 0.0062 - val_mse: 399.9907
Lr :  0.08
Epoch 72/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 173.0551 - val_loss: 0.0062 - val_mse: 400.5084
Lr :  0.08
Epoch 73/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0027 - mse: 172.6990 - val_loss: 0.0062 - val_mse: 400.7729
Lr :  0.08
Epoch 74/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 172.3093 - val_loss: 0.0061 - val_mse: 398.8733
Lr :  0.08
Epoch 75/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 171.9405 - val_loss: 0.0061 - val_mse: 399.6324
Lr :  0.08
Epoch 76/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 171.5539 - val_loss: 0.0062 - val_mse: 400.0404
Lr :  0.08
Epoch 77/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 171.2163 - val_loss: 0.0061 - val_mse: 399.0176
Lr :  0.08
Epoch 78/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 170.8428 - val_loss: 0.0061 - val_mse: 398.2502
Lr :  0.08
Epoch 79/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 170.5107 - val_loss: 0.0061 - val_mse: 399.4756
Lr :  0.08
Epoch 80/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 170.1418 - val_loss: 0.0061 - val_mse: 398.6772
Lr :  0.08
Epoch 81/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 169.7998 - val_loss: 0.0061 - val_mse: 398.6140
Lr :  0.08
Epoch 82/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 169.4597 - val_loss: 0.0061 - val_mse: 398.5729
Lr :  0.08
Epoch 83/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 169.1229 - val_loss: 0.0061 - val_mse: 397.8155
Lr :  0.08
Epoch 84/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 168.7856 - val_loss: 0.0061 - val_mse: 397.5096
Lr :  0.08
Epoch 85/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 168.4588 - val_loss: 0.0061 - val_mse: 397.9902
Lr :  0.08
Epoch 86/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 168.1268 - val_loss: 0.0061 - val_mse: 397.3359
Lr :  0.08
Epoch 87/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 167.8185 - val_loss: 0.0061 - val_mse: 397.9455
Lr :  0.08
Epoch 88/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 167.4879 - val_loss: 0.0061 - val_mse: 397.7877
Lr :  0.08
Epoch 89/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 167.1769 - val_loss: 0.0061 - val_mse: 397.3274
Lr :  0.08
Epoch 90/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 166.8585 - val_loss: 0.0061 - val_mse: 397.2146
Lr :  0.08
Epoch 91/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 166.5457 - val_loss: 0.0061 - val_mse: 397.8896
Lr :  0.08
Epoch 92/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 166.2354 - val_loss: 0.0061 - val_mse: 396.7620
Lr :  0.08
Epoch 93/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0026 - mse: 165.9282 - val_loss: 0.0061 - val_mse: 397.0715
Lr :  0.08
Epoch 94/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 165.6314 - val_loss: 0.0061 - val_mse: 397.2557
Lr :  0.08
Epoch 95/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 165.3255 - val_loss: 0.0061 - val_mse: 396.8092
Lr :  0.08
Epoch 96/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 165.0327 - val_loss: 0.0061 - val_mse: 396.3043
Lr :  0.08
Epoch 97/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 164.7327 - val_loss: 0.0061 - val_mse: 397.1202
Lr :  0.08
Epoch 98/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 164.4545 - val_loss: 0.0061 - val_mse: 397.8727
Lr :  0.08
Epoch 99/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 164.1428 - val_loss: 0.0061 - val_mse: 397.3806
Lr :  0.08
Epoch 100/100
985373/985373 [==============================] - 6s 6us/step - loss: 0.0025 - mse: 163.8922 - val_loss: 0.0061 - val_mse: 396.7912
Lr :  0.08

Answer 1

您的代码工作正常。从损失图中，您可以看到模型收敛得很好。您的问题是您的训练损失和验证损失之间存在很大差距。这有时被称为“低偏差，高方差”。

解决方案包括:

• 正则化
• 提早停止
• 获取更多训练数据
• 确保您的验证数据和训练数据来自同一分布

有很多资源可以为您提供想法，例如:

• https://towardsdatascience.com/understanding-the-bias-variance-tradeoff-165e6942b229
• https://www.tensorflow.org/tutorials/keras/overfit_and_underfit