如何在TensorFlow的CNN与BiLSTM组合架构中添加Dropout解决过拟合问题

你现有代码里仅在BiLSTM层配置了循环dropout，卷积层部分未做dropout正则化，这是过拟合的主要诱因之一，具体添加方式如下：

一、Dropout添加位置选择 #

• 卷积层部分：优先使用SpatialDropout2D（更适配卷积输出的空间相关性，会随机丢弃整个特征通道，比普通Dropout效果更好），添加在每个卷积层的激活函数之后、池化层之前；网络浅层dropout率设为0.1_{0.2，深层可提升到0.2}0.3
• 全连接/输出层部分：在BiLSTM输出之后、最终的Dense层之前添加普通Dropout，率设为0.3~0.5
• 原有BiLSTM的dropout参数可以保留0.5，也可根据拟合情况微调

二、修改后完整代码 #

from keras.layers import SpatialDropout2D, Dropout # 需提前导入对应层

# input with shape of height=42 and width=600
inputs = Input(shape=(42,600,1))

# convolution layer with kernel size (3,3)
conv_1 = Conv2D(64, (3,3), activation = 'relu', padding='same')(inputs)
conv_1 = SpatialDropout2D(0.2)(conv_1) # 新增SpatialDropout
# poolig layer with kernel size (2,2)
pool_1 = MaxPool2D(pool_size=(2, 2), strides=2)(conv_1)

conv_2 = Conv2D(128, (3,3), activation = 'relu', padding='same')(pool_1)
conv_2 = SpatialDropout2D(0.2)(conv_2) # 新增SpatialDropout
pool_2 = MaxPool2D(pool_size=(2, 2), strides=2)(conv_2)

conv_3 = Conv2D(256, (3,3), activation = 'relu', padding='same')(pool_2)
conv_3 = SpatialDropout2D(0.25)(conv_3) # 新增SpatialDropout

conv_4 = Conv2D(256, (3,3), activation = 'relu', padding='same')(conv_3)
conv_4 = SpatialDropout2D(0.25)(conv_4) # 新增SpatialDropout
# poolig layer with kernel size (2,1)
pool_4 = MaxPool2D(pool_size=(2, 1))(conv_4)

conv_5 = Conv2D(512, (3,3), activation = 'relu', padding='same')(pool_4)
batch_norm_5 = BatchNormalization()(conv_5)
batch_norm_5 = SpatialDropout2D(0.3)(batch_norm_5) # 新增SpatialDropout

conv_6 = Conv2D(512, (3,3), activation = 'relu', padding='same')(batch_norm_5)
batch_norm_6 = BatchNormalization()(conv_6)
batch_norm_6 = SpatialDropout2D(0.3)(batch_norm_6) # 新增SpatialDropout
pool_6 = MaxPool2D(pool_size=(2, 1))(batch_norm_6)

conv_7 = Conv2D(512, (2,2), activation = 'relu')(pool_6)
conv_7 = Dropout(0.3)(conv_7) # 新增普通Dropout

squeezed = Lambda(lambda x: K.squeeze(x, 1))(conv_7)

# bidirectional LSTM layers with units=128，原有dropout保留
blstm_1 = Bidirectional(LSTM(128, return_sequences=True, dropout = 0.5))(squeezed)
blstm_2 = Bidirectional(LSTM(128, return_sequences=True, dropout = 0.5))(blstm_1)

blstm_2 = Dropout(0.3)(blstm_2) # 新增普通Dropout
outputs = Dense(len(char_list)+1, activation = 'softmax')(blstm_2)

# model to be used at test time
act_model = Model(inputs, outputs)

三、注意事项 #

• Dropout率不要设置过高，否则会导致模型欠拟合，建议从低到高逐步调试
• 可以配合早停（EarlyStopping）、训练数据增强、L2正则化等方式共同使用，进一步缓解过拟合

训练准确率损失曲线参考 #

内容的提问来源于stack exchange，提问作者maryam mehboob