不管学习率如何，损失都不会改变

Question

我建立了一个深度学习模型，与VGG网络有点类似。我正在将Keras与Tensorflow后端一起使用。模型摘要如下：

model = Sequential()
model.add(Conv2D(64, 3, border_mode='same', activation='relu', input_shape=(180,320,3)))
model.add(Conv2D(64, 3, border_mode='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=2))
model.add(Conv2D(64, 3, border_mode='same', activation='relu'))
model.add(Conv2D(64, 3, border_mode='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=2))
model.add(Conv2D(128, 3, border_mode='same', activation='relu'))
model.add(Conv2D(128, 3, border_mode='same', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=2))
model.add(Conv2D(128, 3, border_mode='same', activation='relu'))
model.add(Conv2D(128, 3, border_mode='same', activation='relu'))
model.add(Flatten())
model.add(Dropout(0.5))
model.add(Dense(1024, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(9,  activation='relu'))

我尝试了优化器（SGD，Adam等），损失（MSE，MAE等），批处理大小（32和64）的不同组合。我什至试验了学习率的范围从0.001到10000。但是，即使经过20个纪元，无论我使用哪种损失函数，验证损失都保持不变。训练损失的变化很小。我究竟做错了什么？

我的网络应该训练什么：给定输入图像，网络需要预测可以从该图像得出的9个真实值的集合。

培训期间的终端输出：

    Epoch 1/100
    4800/4800 [==============================] - 96s 20ms/step - loss: 133.6534 - mean_absolute_error: 133.6534 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 2/100
    4800/4800 [==============================] - 49s 10ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 3/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 4/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 5/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 6/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 7/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 8/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 9/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 10/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 11/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 12/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 13/100
    4800/4800 [==============================] - 50s 10ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 14/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 15/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 16/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 17/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 18/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 19/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 20/100
    4800/4800 [==============================] - 51s 11ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744
    Epoch 21/100
    4800/4800 [==============================] - 50s 10ms/step - loss: 132.8033 - mean_absolute_error: 132.8033 - val_loss: 132.3744 - val_mean_absolute_error: 132.3744

Answer 1

relu

不要随意使用relu！它具有恒定的零区域，没有梯度。卡住是完全正常的。

• 最严重的错误是在最后一层有relu。
  • 如果您希望输出从0到无穷大，请使用'softplus'。
  • 如果要在0和1之间使用 'sigmoid'
  • 如果要在-1和+1之间使用 'tanh'
• 你的学习速度是巨大的。随着RELU，需要小的学习率：
  • 去0.0001和下面。
• 尝试其他不会卡住的激活
• 尝试在激活之前添加批处理规范化（这样，无论什么情况，您都可以确保某些值大于零）：
  • 这也使您获得更高的学习率

。

model.add(Conv2D(..... , activation='linear'))
model.add(BatchNormalization())
model.add(Activation('relu'))