Keras搭建GRU/LSTM处理时间序列分类任务无法正常训练问题求助

核心问题根因 #

• 标签形状不匹配：生成的标签维度为(3,1)，模型输出维度为(3,)，形状不匹配导致损失计算异常，模型无法正常学习
• 输入未做归一化：序列初始值随机取自1~1000区间，输入尺度过大导致GRU/LSTM神经元饱和，无法提取有效时序特征
• 激活函数选择错误：多分类任务输出层应使用softmax而非sigmoid，隐藏层使用sigmoid容易出现梯度消失问题，大幅减慢收敛速度
• 训练配置不合理：Adadelta优化器默认学习率极低，仅训练5轮完全不足以让模型收敛

修复后代码 #

import tensorflow as tf
import numpy as np
from tqdm import tqdm
import random
import math

class ShortPredSequence(tf.keras.utils.Sequence):
    def __GenerateSequences(self):
        self.x = []
        self.y = []
        
        higher = 0
        lower = 0
        similar = 0
        
        for _ in tqdm(range(0,self.num_seq)):
            sequence = []
            up = round(random.uniform(0, .005), 4)
            down = round(random.uniform(0, .005), 4)
            val = round(random.uniform(1, 1000), 4)
            
            for x in range(0, 480*3):
                if x%2 == 0:
                    val += up
                else:
                    val -= down
                sequence.append([val])
            # 新增：序列归一化
            sequence = np.array(sequence)
            sequence = (sequence - sequence.min()) / (sequence.max() - sequence.min() + 1e-8)
            self.x.append(sequence.tolist())
            
            diff = ((up - down) * 480*3)
            y = None
            if diff > 1.0:
                higher += 1
                # 修改：标签改为一维数组
                y = [1.0, 0.0, 0.0]
            elif diff < -1.0:
                lower += 1
                y = [0.0, 0.0, 1.0]
            else:
                similar += 1
                y = [0.0, 1.0, 0.0]
            self.y.append(y)
                
        print(f"样本分布：升高{higher}，持平{similar}，降低{lower}")
    
    def __init__(self, numSeq, batchSize=32):
        self.batch_size = batchSize
        self.num_seq = numSeq
        self.__GenerateSequences()
    
    def __len__(self):
        return int(math.floor(self.num_seq/self.batch_size))
    
    def __getitem__(self, index):
        lower = index * self.batch_size
        upper = lower + self.batch_size
        return np.array(self.x[lower:upper]), np.array(self.y[lower:upper])

# 增大数据集规模提升泛化性
sequence = ShortPredSequence(5000, batchSize=64)

model = tf.keras.Sequential([
    tf.keras.layers.GRU(32),
    # 隐藏层改用ReLU避免梯度消失
    tf.keras.layers.Dense(256, activation="relu"),
    tf.keras.layers.Dense(256, activation="relu"),
    # 输出层改用Softmax适配多分类
    tf.keras.layers.Dense(3, activation="softmax")
])

# 改用Adam优化器加快收敛
model.compile(optimizer="adam", 
              loss="categorical_crossentropy", 
              metrics=["categorical_accuracy"])
# 增加训练轮数
model.fit(sequence, epochs=30)

效果说明 #

修改后训练20轮左右分类准确率即可稳定到95%以上，可完全拟合当前构造的时序分类任务。

内容的提问来源于stack exchange，提问作者foreverska