Keras多输入神经网络无论输入均输出Y_train平均值的问题求助

问题现象 #

用Python的Keras构建了一个融合数值输入和整数编码文本输入的神经网络，但无论输入是什么，模型始终输出同一个固定值。该值接近Y_train数组的平均值（Y_train已归一化，数值大多在0.4-0.6区间）：训练过程中损失能降至0.02左右的极小值，但输出完全不受输入影响；训练准确率保持恒定，仅能准确预测与平均值非常接近的样本，训练后模型输出几乎和Y_train的平均值完全一致。

已尝试的解决措施 #

• 测试3种不同优化器、2种损失函数，共训练6个模型：每个模型的固定输出值不同，但单个模型的输出仍完全不随输入变化
• 实现权重正则化与初始化、早停机制、Dropout层（用于防止过拟合）
• 对所有输入和输出进行缩放处理
• 调整模型复杂度，设置Zero偏置初始化、强偏置正则化，将激活函数替换为LeakyReLU（解决神经元死亡问题），多次调整学习率（范围从0.003到1e-6）

模型代码 #

def regDense(a):
    return layers.Dense(a, activation=LeakyReLU(), kernel_initializer=initializers_v2.HeNormal(),
                       kernel_regularizer=l2(0.001), bias_initializer=initializers_v2.Zeros(),
                       bias_regularizer=l1_l2(0.003, 0.02))


def regLSTM(a):
    return layers.LSTM(a, kernel_regularizer=l1_l2(0.0001, 0.0003),
                kernel_initializer=initializers_v2.GlorotNormal(),
                bias_initializer=initializers_v2.Zeros(),
                return_sequences=True,
                bias_regularizer=l1_l2(0.0002, 0.002))


num_inp = keras.Input(shape=(30, 3, 1), name='nums')
text_inp = keras.Input(shape=(30, 7, 3208), name='text')

embed = layers.Embedding(vocabsize, output_dim=152)(text_inp)
tlstm1 = layers.TimeDistributed(layers.TimeDistributed(regLSTM(256)))(embed)
tdrop1 = layers.Dropout(0.2)(tlstm1)
nlstm1 = layers.TimeDistributed(regLSTM(256))(num_inp)
ndrop1 = layers.Dropout(0.2)(nlstm1)


tlstm2 = layers.TimeDistributed(layers.TimeDistributed(regLSTM(256)))(tdrop1)
tdrop2 = layers.Dropout(0.2)(tlstm2)
nlstm2 = layers.TimeDistributed(regLSTM(256))(ndrop1)
ndrop2 = layers.Dropout(0.2)(nlstm2)


tlstm3 = layers.TimeDistributed(layers.TimeDistributed(regLSTM(256)))(tdrop2)
tdrop3 = layers.Dropout(0.2)(tlstm3)
ndense1 = regDense(213)(ndrop2)
ndrop3 = layers.Dropout(0.5)(ndense1)


tdense1 = regDense(200)(tdrop3)
tdrop4 = layers.Dropout(0.5)(tdense1)
ndense2 = regDense(170)(ndrop3)
ndrop4 = layers.Dropout(0.5)(ndense2)


tdense2 = regDense(144)(tdrop4)
tpool2 = layers.MaxPooling3D((1, 1, 401), padding='same')(tdense2)
trsp1 = layers.Reshape((30, 1152, 7))(tpool2)
tpool3 = layers.MaxPooling2D((1, 3), padding='same')(trsp1)
trsp2 = layers.Reshape((2688, 30))(tpool3)
tpool4 = layers.MaxPooling1D(7, padding='same')(trsp2)
trsp3 = layers.Reshape((30, 384))(tpool4)
tdrop5 = layers.Dropout(0.5)(trsp3)
ndense3 = regDense(128)(ndrop4)
nrsp = layers.Reshape((30, 384))(ndense3)
ndrop5 = layers.Dropout(0.5)(nrsp)


concat = layers.concatenate([tdrop5, ndrop5])
prc1 = regDense(384)(concat)
pdrop1 = layers.Dropout(0.5)(prc1)
prc2 = regDense(192)(pdrop1)
pdrop2 = layers.Dropout(0.5)(prc2)
prc3 = regDense(96)(pdrop2)
pdrop3 = layers.Dropout(0.5)(prc3)
prc4 = regDense(48)(pdrop3)
pdrop4 = layers.Dropout(0.5)(prc4)
prc5 = regDense(24)(pdrop4)
pdrop5 = layers.Dropout(0.3)(prc5)
prc6 = regDense(12)(pdrop5)
lastpool = layers.GlobalAveragePooling1D()(prc6)
last = layers.Dense(1, activation='relu', kernel_regularizer=l2(0.0008),
                    kernel_initializer=initializers_v2.HeNormal(),
                    bias_initializer=initializers_v2.Zeros(),
                    bias_regularizer=l1_l2(0.003, 0.02),
                    name='output')(lastpool)
model = keras.Model(inputs=[num_inp, text_inp], outputs=last)

model.compile(optimizer='adam', loss=losses.MeanSquaredError(), metrics=['accuracy'])

model.fit(
    {"nums": X1_train, "text": X2_train},
    {"output": y_train}, validation_data=({'nums':X1_test, 'text':X2_test}, y_test),
    epochs=9, callbacks=[callbacks.EarlyStopping(monitor='val_loss', patience=1, restore_best_weights=True), callbacks.LearningRateScheduler(schedule=scheduler, verbose=1)],
    batch_size=1)

补充：训练后权重与偏置情况 #

检查最近训练的模型参数，几乎所有层的权重为：

-4.1277455e-33  1.9695687e-33  4.1031190e-23 ... -1.5578026e-20
  -8.8301270e-33 -1.5537877e-20

偏置为：

-4.0955161e-05  4.0955088e-05 -4.0955132e-05 -4.0955088e-05
 -4.0955088e-05 -4.0955132e-05 -4.0955132e-05 -4.0955088e-05

内容的提问来源于stack exchange，提问作者flexorcist