如何在Keras中合并3个模型：按性别分类结果分支预测对应人群年龄

现有代码问题修正 #

首先你当前的子网络实现有两个明显问题需要先调整：

• 性别分类头（A10）使用1个神经元搭配softmax激活是错误的，二分类场景下1神经元的softmax输出永远为1，无法实现0/1分类，需要把激活函数换成sigmoid
• 你当前的男性/女性年龄分支直接对接性别分类的输出不符合设计逻辑，年龄预测的输入应该是原始输入特征，性别输出仅作为分支选择的判断依据，不应该作为年龄模型的输入特征

分支网络实现方案 #

基于TensorFlow Keras Functional API结合条件判断算子即可实现你要的分支路由逻辑，完整可运行代码如下：

import tensorflow as tf
from tensorflow.keras import layers

# 定义性别分类子模型
def build_gender_model(input_shape):
    A1 = layers.Input(shape=input_shape, name='gender_input')
    A2 = layers.Dense(512, activation='relu')(A1)
    A3 = layers.Dropout(0.5)(A2)
    A4 = layers.Dense(1024, activation='relu')(A3)
    A5 = layers.Dropout(0.5)(A4)
    A6 = layers.Dense(1024, activation='relu')(A5)
    A7 = layers.Dropout(0.4)(A6)
    A8 = layers.Dense(256, activation='relu')(A7)
    A9 = layers.Dropout(0.1)(A8)
    A10 = layers.Dense(1, activation='sigmoid', name='gender_output')(A9)
    return tf.keras.Model(inputs=A1, outputs=[A10, A1], name='gender_model')

# 定义男性年龄预测子模型
def build_male_age_model(input_shape):
    B1 = layers.Input(shape=input_shape, name='male_age_input')
    B2 = layers.Dense(512, activation='relu')(B1)
    B3 = layers.Dropout(0.5)(B2)
    B4 = layers.Dense(1024, activation='relu')(B3)
    B5 = layers.Dropout(0.5)(B4)
    B6 = layers.Dense(1024, activation='relu')(B5)
    B7 = layers.Dropout(0.4)(B6)
    B8 = layers.Dense(256, activation='relu')(B7)
    B9 = layers.Dropout(0.1)(B8)
    B10 = layers.Dense(8, activation='softmax', name='male_age_output')(B9)
    return tf.keras.Model(inputs=B1, outputs=B10, name='male_age_model')

# 定义女性年龄预测子模型
def build_female_age_model(input_shape):
    C1 = layers.Input(shape=input_shape, name='female_age_input')
    C2 = layers.Dense(512, activation='relu')(C1)
    C3 = layers.Dropout(0.5)(C2)
    C4 = layers.Dense(1024, activation='relu')(C3)
    C5 = layers.Dropout(0.5)(C4)
    C6 = layers.Dense(1024, activation='relu')(C5)
    C7 = layers.Dropout(0.4)(C6)
    C8 = layers.Dense(256, activation='relu')(C7)
    C9 = layers.Dropout(0.1)(C8)
    C10 = layers.Dense(8, activation='softmax', name='female_age_output')(C9)
    return tf.keras.Model(inputs=C1, outputs=C10, name='female_age_model')

# 组装端到端模型
input_shape = pca.shape[1]
inputs = layers.Input(shape=input_shape, name='main_input')

# 先执行性别分类
gender_pred, raw_features = build_gender_model(input_shape)(inputs)
# 二值化性别预测结果
gender_mask = tf.cast(tf.greater(gender_pred, 0.5), tf.bool)

# 两个年龄分支并行计算
male_age_pred = build_male_age_model(input_shape)(raw_features)
female_age_pred = build_female_age_model(input_shape)(raw_features)

# 按性别选择对应分支的输出
final_age_pred = layers.Lambda(lambda x: tf.where(x[0], x[1], x[2]))([gender_mask, male_age_pred, female_age_pred])

# 生成最终端到端模型
end2end_model = tf.keras.Model(inputs=inputs, outputs=[gender_pred, final_age_pred], name='gender_age_pipeline')

整体准确率计算 #

自定义评估指标即可统计端到端准确率，规则为性别分类正确且对应年龄分支预测正确才算样本预测正确，实现代码如下：

def end2end_acc(y_true, y_pred):
    # y_true结构：[真实性别标签, 真实年龄标签]
    # y_pred结构：[预测性别输出, 预测年龄输出]
    gender_true = y_true[:, 0]
    age_true = y_true[:, 1]
    gender_pred = tf.cast(tf.greater(y_pred[:, 0], 0.5), tf.float32)
    age_pred = tf.cast(tf.argmax(y_pred[:, 1:], axis=1), tf.float32)
    
    gender_correct = tf.equal(gender_true, gender_pred)
    age_correct = tf.equal(age_true, age_pred)
    total_correct = tf.cast(tf.logical_and(gender_correct, age_correct), tf.float32)
    return tf.reduce_mean(total_correct)

训练时将该指标传入model.compile()的metrics参数即可自动统计。

内容的提问来源于stack exchange，提问作者Armağan