使用tflearn进行回归的神经网络

Question

My question is about coding a neural network which does regression(and NOT classification) using tflearn.

Dataset:

fixed acidity  volatile acidity  citric acid  ...  alcohol  quality   
7.4             0.700            0.00         ...  9.4        5    
7.8             0.880            0.00         ...  9.8        5  
7.8             0.760            0.04         ...  9.8        5     
11.2            0.280            0.56         ...  9.8        6      
7.4             0.700            0.00         ...  9.4        5

I want to build a neural network which takes in 11 features (chemical values in wine) and outputs or predicts a score ie, quality(out of 10). I DON'T want to classify the wine like quality_1, quality_2,... I want the model to perform a regression function for my features and predict a value out of 10(could be even a float).

The quality column in my data only has values = [3, 4, 5, 6, 7, 8, 9]. It does not contain 1, 2, and 10.

Due to the lack in experience, I could only code a neural network that CLASSIFIES the wine into classes like [score_3, score_4,...] and I used one hot encoding to do so.

Processed Data:

Features:

[[  7.5999999    0.23         0.25999999 ...,   3.02999997   0.44
    9.19999981]
 [  6.9000001    0.23         0.34999999 ...,   2.79999995   0.54000002
   11.        ]
 [  6.69999981   0.17         0.37       ...,   3.25999999   0.60000002
   10.80000019]
 ..., 
 [  6.30000019   0.28         0.47       ...,   3.11999989   0.50999999
    9.5       ]
 [  5.19999981   0.64499998   0.         ...,   3.77999997   0.61000001
   12.5       ]
 [  8.           0.23999999   0.47999999 ...,   3.23000002   0.69999999
   10.        ]]

Labels:

[[ 0.  1.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  1.  0.  0.]
 ..., 
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  1. ...,  0.  0.  0.]]

Code for a neural network which CLASSIFIES into different classes:

import pandas as pd
import numpy as np
import tflearn
from tflearn.layers.core import input_data, fully_connected
from tflearn.layers.estimator import regression
from sklearn.model_selection import train_test_split


def preprocess():

    data_source_red = 'F:\Gautam\...\Datasets\winequality-red.csv'

    data_red = pd.read_csv(data_source_red, index_col=False, sep=';')

    data = pd.get_dummies(data, columns=['quality'], prefix=['score'])

    x = data[data.columns[0:11]].values
    y = data[data.columns[11:18]].values

    x = np.float32(x)
    y = np.float32(y)

    return (x, y)


x, y = preprocess()

train_x, test_x, train_y, test_y = train_test_split(x, y, test_size = 0.2)

network = input_data(shape=[None, 11], name='Input_layer')

network = fully_connected(network, 10, activation='relu', name='Hidden_layer_1')

network = fully_connected(network, 10, activation='relu', name='Hidden_layer_2')

network = fully_connected(network, 7, activation='softmax', name='Output_layer')

network = regression(network, batch_size=2, optimizer='adam', learning_rate=0.01)

model = tflearn.DNN(network)

model.fit(train_x, train_y, show_metric=True, run_id='wine_regression',
          validation_set=0.1, n_epoch=1000)

The neural network above is a poor one(accuracy=0.40). Moreover, it classifies the data into different classes. I would like to know how to code a regression neural network which gives a score out of 10 for the input features(and NOT CLASSIFICATION). I would also prefer tflearn as I'm quite comfortable with it.

Hopefully, a few alterations to my code is all I need. Thanks.

Answer 1

This is the line in your code which makes your network a classifier with seven categories, instead of a regressor:

network = fully_connected(network, 7, activation='softmax', name='Output_layer')

I don't use TFLearn any more, I have switched over to Keras (which is similar, and has better support). However, I will suggest that you want the following output layer instead:

network = fully_connected(network, 1, activation='linear', name='Output_layer')

Also, your training data will need to change. If you want to perform a regression, you want a one-dimensional scalar label instead. I assume that you still have the original data, which you say that you altered? If not, the UC Irvine Machine Learning Data Repository has the wine quality data with a single, numerical Quality column .