将Keras模型的输出重新缩放为原始比例 [英] Re-scaling outputs from a Keras model back to original scale

问题描述

我是神经网络的新手(只是免责声明).

I'm new to neural nets (just a disclaimer).

基于8个特征，我有一个预测混凝土强度的回归问题.我首先要做的是使用最小-最大归一化来重新缩放数据:

I have a regression problem of predicting the strength of concrete, based on 8 features. What I've done first, is rescaled the data using min-max normalization:

# Normalize data between 0 and 1
from sklearn.preprocessing import MinMaxScaler

min_max = MinMaxScaler()
dataframe2 = pd.DataFrame(min_max.fit_transform(dataframe), columns = dataframe.columns)

然后将数据帧转换为numpy数组，并将其拆分为X_train，y_train，X_test，y_test. 现在，这是网络本身的Keras代码:

then converted the dataframe into numpy array and split it into X_train, y_train, X_test, y_test. Now here is the Keras code for the network itself:

from keras.models import Sequential
from keras.layers import Dense, Activation

#Set the params of the Neural Network
batch_size = 64
num_of_epochs = 40
hidden_layer_size = 256

model = Sequential()
model.add(Dense(hidden_layer_size, input_shape=(8, )))
model.add(Activation('relu'))
model.add(Dense(hidden_layer_size))
model.add(Activation('relu'))
model.add(Dense(hidden_layer_size))
model.add(Activation('relu'))
model.add(Dense(1))
model.add(Activation('linear'))


model.compile(loss='mean_squared_error', # using the mean squared error function
              optimizer='adam', # using the Adam optimiser
              metrics=['mae', 'mse']) # reporting the accuracy with mean absolute error and mean squared error

model.fit(X_train, y_train, # Train the model using the training set...
          batch_size=batch_size, epochs=num_of_epochs,
          verbose=0, validation_split=0.1)

# All predictions in one array
predictions = model.predict(X_test)

问题:

1. 预测数组将具有缩放格式的所有值(0到1之间)，但是显然我需要将预测值设为真实值.如何将这些输出重新缩放为实际值?

1. predictions array will have all the values in the scaled format (between 0 and 1), but obviously I would need the predictions to be in their real values. How can I rescale those outputs back to the real values?

Min-Max或Z-Score标准化是否更适合回归问题?那这个批处理规范化"呢?

Is Min-Max or Z-Score standardization more appropriate for regression problems? What about this 'Batch-Normalization'?

谢谢

推荐答案

按照文档，MinMaxScaler类具有inverse_transform方法，该方法可以执行您想要的操作:

As per the doc, the MinMaxScaler class has an inverse_transform method which does what you want:

inverse_transform(X):根据feature_range撤消X的缩放比例.

inverse_transform(X): Undo the scaling of X according to feature_range.

这篇关于将Keras模型的输出重新缩放为原始比例的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！