如何在 CNN 中添加 Dropout

Question

我正在使用 CNN 训练一个时尚 MNIST 数据。由于过度拟合，我尝试添加 Dropout 层。但它不起作用

在我添加 Dropout 之前，模型运行良好。

def fashion_model()
    batch_size = 64
    epochs = 20
    num_classes = 10
    fashion_drop_model = Sequential()
    fashion_drop_model.add(Conv2D(32, kernel_size=(3, 3),activation='linear',padding='same',input_shape=(28,28,1)))
    fashion_drop_model.add(LeakyReLU(alpha=0.1))
    fashion_drop_model.add(MaxPooling2D((2, 2),padding='same'))
    fashion_drop_model.add(Dropout(0.25))

    fashion_drop_model.add(Conv2D(64, (3, 3), activation='linear',padding='same'))
    fashion_drop_model.add(LeakyReLU(alpha=0.1))
    fashion_drop_model.add(MaxPooling2D(pool_size=(2, 2),padding='same'))
    fashion_drop_model.add(Dropout(0.25))

    fashion_drop_model.add(Conv2D(128, (3, 3), activation='linear',padding='same'))
    fashion_drop_model.add(LeakyReLU(alpha=0.1))                  
    fashion_drop_model.add(MaxPooling2D(pool_size=(2, 2),padding='same'))
    fashion_drop_model.add(Dropout(0.4))

    fashion_drop_model.add(Flatten())
    fashion_drop_model.add(Dense(128, activation='linear'))
    fashion_drop_model.add(LeakyReLU(alpha=0.1))           
    fashion_drop_model.add(Dropout(0.3))
    fashion_drop_model.add(Dense(num_classes, activation='softmax'))

    return fashion_drop_model.summary()

fashion_model()

我得到的错误是：UnboundLocalError: local variable 'a' referenced before assignment

PS：在逐行浏览代码后，我发现错误正在蔓延到第 8 行 (fashion_drop_model.add(Dropout(0.25)))

Answer 1

您的 Python 函数定义中缺少一个冒号：

def fashion_model(): #<--

完成此操作后，代码应该运行。在 Google Colaboratory 中运行此程序，您会看到生成了您的模型的摘要：

注意

强烈建议不要在卷积层之后使用 Dropout 层。卷积层的重点是利用空间邻域内的像素来提取正确的特征以馈送到密集层。 Dropout 会破坏这种关系，从而阻止您的模型成功学习这些特征。

有关详细信息，请参阅 Reddit 上的此讨论：https://www.reddit.com/r/MachineLearning/comments/42nnpe/why_do_i_never_see_dropout_applied_in/