keras搭建模型架构的方式及自定义层（继承Layer）

1. Sequential()+add

详见博客：https://blog.csdn.net/weixin_43178406/article/details/97000893

# 网络
model = Sequential()
model.add(Dense(units = 10, input_dim = 1))
model.add(Activation('tanh'))
model.add(Dense(units =1, activation='tanh'))

或者

model = Sequential([
                    Dense(units = 500, input_shape = (X_train.shape[1],), kernel_regularizer=l2(0.003)),# kernel_regularizer:权值正则化项,bias_regularizer:偏置项的正则化， activity_regularizer：激活项的正则化
                    Activation('tanh'),
                    Dropout(0.1),
                    Dense(units = 200, activation='tanh', kernel_regularizer=l2(0.003)),
                    Dropout(0.2),
                    Dense(units=100, activation='tanh', kernel_regularizer=l2(0.003)),
                    Dropout(0.3),
                    Dense(units=10, activation='softmax', kernel_regularizer=l2(0.003))
                      
])

2. Input&Model

def bi_lstm(max_len, max_cnt, embed_size, embedding_matrix):
    _input = Input(shape=(max_len,), dtype='int32')
    _embed = Embedding(max_cnt, embed_size, input_length=max_len, weights=[embedding_matrix], trainable=False)(_input)
    _embed = SpatialDropout1D(0.2)(_embed) # 词向量的dropout
    lstm_result = Bidirectional(LSTM(100, return_sequences=False), merge_mode='sum')(_embed)
    fc = Dropout(0.5)(lstm_result)
    fc = Dense(10)(fc)
    fc = Dropout(0.1)(fc)
    preds = Activation('softmax')(fc)

    model = Model(inputs=_input, outputs=preds)
    return model

可以看到这种方式下，Input指定了输入的shape，下面调用keras定义好的层时，可将上一步的结果通过括号+结果名称的方式输入，如fc = Dropout(0.5)(lstm_result)lstm_result就是lstm的输出结果

3. 自定义层（继承Layer）

在keras中也可以自定义层，类似于pytorch继承Module，keras通过继承Layer来自定义自己的层（类）等