Tensorflow/Keras DNN回归模型精度与模型问题求助

我是Tensorflow/Keras新手，用自有数据搭建了首个DNN回归模型，用400个特征预测1个标签，但遇到损失与精度相关问题。多次试验发现，训练轮次的损失和验证精度很早就趋于平稳，验证精度表现不佳。试过100到3000个不同的训练轮次，结果一致，怀疑是代码有bug、模型设置不合理或是训练数据不足。

版本信息 #

• Tensorflow 2.9.1
• Python 3.9.12
• Numpy 1.23.1
• Pandas 1.4.3

数据说明 #

原始数据样本：特征为0-400列，标签为400-599列，本模型仅预测标签“lower_pos_0”（第400列）。

代码 #

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.callbacks import TensorBoard
import datetime

print(tf.__version__)
# Make NumPy printouts easier to read.
np.set_printoptions(precision=3, suppress=True)

#read in the csv file into a dataframe. Sample file path "C:\...\...\...\csv name.csv"
filelocation = "C:\(place the path to the csv here"
raw_dataset = pd.read_csv(filelocation)

#drop all columns that this model will NOT be used in training this model. Those columns are "lower_pos_1" through "lower_load_99"
raw_dataset = raw_dataset.drop(raw_dataset.loc[:, 'lower_pos_1':'lower_load_99'].columns, axis=1)
print('shape of raw_dataset:', raw_dataset.shape)
print(raw_dataset.head())

#convert the data frame to array
dataset = raw_dataset.copy()
dataset.tail()

#create a training and test set
train_dataset = dataset.sample(frac=0.8, random_state=0)
test_dataset = dataset.drop(train_dataset.index)

#check the data
train_dataset.describe().transpose()

#split features from labels
train_features = train_dataset.copy()
test_features = test_dataset.copy()

#print the shape
print('train features shape:',train_features.shape)
print(train_features.head())
print('test features shape:',test_features.shape)
print(test_features.head())

#drop the labels from the features. The "lower_pos_0" column is the label that the model is trying to predict.
train_labels = train_features.pop('lower_pos_0')
test_labels = test_features.pop('lower_pos_0')

#print the shape
print('train labels shape:',train_labels.shape)
print(train_labels.head())
print('test labels shape:',test_labels.shape)
print(test_labels.head())

#normalize the data using keras
normalizer = tf.keras.layers.Normalization(axis=-1)
normalizer.adapt(np.array(train_features))
print(normalizer.mean.numpy())
first = np.array(train_features[:1])

with np.printoptions(precision=2, suppress=True):
  print('First example:', first)
  print()
  print('Normalized:', normalizer(first).numpy())

#build the model
def build_and_compile_model(norm):
  model = keras.Sequential([
      norm,
      layers.Dense(401, activation='relu'),
      layers.Dense(401, activation='relu'),
      layers.Dense(1)
  ])

  model.compile(loss='mean_absolute_error',
                optimizer=tf.keras.optimizers.Adam(0.001),
                 metrics='accuracy')
  return model

#display the models summary
dnn_model = build_and_compile_model(normalizer)
dnn_model.summary()

#setup tensorboard for viewing data during training

log_folder = "Traininglogs/"  + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")

callbacks = [TensorBoard(log_dir=log_folder,
                         histogram_freq=1,
                         write_graph=True,
                         write_images=True,
                         update_freq='epoch',
                         profile_batch=2,
                         embeddings_freq=1)]

#train the model with model.fit()
dnn_model.fit(train_features, train_labels, epochs=125, validation_split=0.2, callbacks=callbacks)

#evaluate the model with model.evaluate()
loss, mae = dnn_model.evaluate(test_features, test_labels, verbose=2)

#saving the model once completed
dnn_model.save('dnn_model')

1. 核心错误：回归任务误用分类指标 #

你的模型是回归任务（预测连续值），但编译时指定了metrics='accuracy'——这个指标仅适用于分类任务，对回归完全无效。这是你看到“验证精度表现不佳”的根本原因，Keras会强行把回归输出和标签做分类精度计算，结果没有任何参考意义。

修改编译代码，换成回归任务适用的指标：

model.compile(loss='mean_absolute_error',
              optimizer=tf.keras.optimizers.Adam(0.001),
              metrics=['mean_absolute_error'])

同时调整评估代码的变量名，避免混淆：

loss, val_mae = dnn_model.evaluate(test_features, test_labels, verbose=2)

2. 模型结构优化 #

• 当前两层401神经元的结构和输入特征数（400）几乎一致，容易引发过拟合或梯度消失。建议缩小层规模，或加入Dropout层抑制过拟合：

  model = keras.Sequential([
      norm,
      layers.Dense(256, activation='relu', kernel_initializer='he_normal'),
      layers.Dropout(0.2),
      layers.Dense(128, activation='relu', kernel_initializer='he_normal'),
      layers.Dropout(0.2),
      layers.Dense(1)
  ])
  

  其中he_normal初始化器配合ReLU激活，能有效缓解梯度消失问题。

3. 训练策略调整 #

• 添加早停（Early Stopping）回调，自动在验证损失不再下降时停止训练，避免无意义迭代，同时保留最优权重：

  from tensorflow.keras.callbacks import EarlyStopping
  
  callbacks = [
      TensorBoard(log_dir=log_folder,
                  histogram_freq=1,
                  write_graph=True,
                  write_images=True,
                  update_freq='epoch',
                  profile_batch=2,
                  embeddings_freq=1),
      EarlyStopping(monitor='val_loss', patience=10, restore_best_weights=True)
  ]
  

• 尝试调整Adam学习率，比如从0.001降到0.0005，若训练损失下降过慢再适当调高。

4. 数据相关检查 #

• 确认训练样本量：如果样本数远小于特征数的10倍（比如不足4000条），模型很难学到有效规律，优先补充数据。
• 计算特征与标签的相关性：用train_features.corrwith(train_labels)输出每个特征和标签的相关系数，剔除相关性极低的特征，减少噪声干扰。
• 检查标签分布：若标签值集中在狭窄区间，可尝试对标签做标准化或对数变换，降低模型学习难度。

5. 代码细节修正 #

• 文件路径存在语法错误："C:\(place the path to the csv here"需改为双反斜杠格式"C:\\path\\to\\your\\file.csv"或原始字符串r"C:\path\to\your\file.csv"，否则会触发路径解析错误。
• 移除无意义代码：dataset = raw_dataset.copy(); dataset.tail()这行没有实际作用，可直接删除。

内容的提问来源于stack exchange，提问作者TheNewGuy