如何训练具有单个数字输出的神经网络？答案

【问题标题】：How to train a neural network with a single digit output?如何训练具有单个数字输出的神经网络？
【发布时间】：2021-04-13 09:57:17
【问题描述】：

我的目标是使用 Tensorflow 构建一个神经网络，该网络接收一个 128 维向量作为输入并分配一个等级 (0.0-5.0)。

问题：无论我使用多少个 epoch，准确度都保持在 0.0

这是我目前得到的

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt
import pickle


#Load data frame containing training data
with open("label_train_vector.txt", "rb") as f: 
   df_train = pickle.load(f)


#Transform training data to correct format
feature_train_list = df_train['feature_vector'].to_list()
train_input = np.array(feature_train_list)
train_label_rating = df_train['rating'].to_numpy()/5


rating_model = keras.Sequential([
    
    keras.layers.Dense(64, activation="sigmoid", input_shape=(128,)),
    keras.layers.Dense(1, activation="sigmoid")
])

rating_model.compile(optimizer='rmsprop', loss="mse", metrics=["accuracy"])
rating_model.fit(train_input, train_label_rating, epochs=10)

这段代码的输出是

Epoch 1/10
104/104 [==============================] - 1s 992us/step - loss: 0.0280 - accuracy: 0.0000e+00
Epoch 2/10
104/104 [==============================] - 0s 979us/step - loss: 0.0140 - accuracy: 0.0000e+00
Epoch 3/10
104/104 [==============================] - 0s 994us/step - loss: 0.0111 - accuracy: 0.0000e+00
Epoch 4/10
104/104 [==============================] - 0s 929us/step - loss: 0.0093 - accuracy: 0.0000e+00
Epoch 5/10
104/104 [==============================] - 0s 934us/step - loss: 0.0082 - accuracy: 0.0000e+00
Epoch 6/10
104/104 [==============================] - 0s 938us/step - loss: 0.0080 - accuracy: 0.0000e+00
Epoch 7/10
104/104 [==============================] - 0s 929us/step - loss: 0.0079 - accuracy: 0.0000e+00
Epoch 8/10
104/104 [==============================] - 0s 936us/step - loss: 0.0074 - accuracy: 0.0000e+00
Epoch 9/10
104/104 [==============================] - 0s 930us/step - loss: 0.0074 - accuracy: 0.0000e+00
Epoch 10/10
104/104 [==============================] - 0s 923us/step - loss: 0.0073 - accuracy: 0.0000e+00

在训练过程中准确率没有变化，但我不知道哪里出了问题。

【问题讨论】：

这不是正确的设置。尽管sigmoid 给出的输出介于 0 和 1 之间。你将如何预测大于 1 的值？
我通过将评级标准化为 [0,1] 来调整您的观点，但同样的问题仍然存在。
如果这是一个回归问题（正如使用loss="mse" 所暗示的那样），准确性毫无意义，activation="sigmoid" 是错误的。此外，我们不使用sigmoid 激活中间层 - idownvotedbecau.se/noresearch

标签： python tensorflow machine-learning keras neural-network

【解决方案1】：

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt
import pickle


#Load data frame containing training data
with open("label_train_vector.txt", "rb") as f: 
   df_train = pickle.load(f)


#Transform training data to correct format
feature_train_list = df_train['feature_vector'].to_list()
train_input = np.array(feature_train_list)
train_label_rating = df_train['rating'].to_numpy()/5


rating_model = keras.Sequential([
    
    keras.layers.Dense(64, activation="relu", input_shape=(128,)),
    keras.layers.Dense(1, activation="sigmoid")
])

rating_model.compile(optimizer='rmsprop', loss="mse", metrics=["accuracy"])
rating_model.fit(train_input, train_label_rating, epochs=10)

可能会更好

【讨论】：