How to use datasets.fetch_mldata() in sklearn - Python?

mldata.org does not have an enforced convention for storing data or naming the columns in a data set. The default behavior of this function works well with most of the common cases mentioned below:

1. Data values stored in the column are 'Data’, and target values stored in the column are ‘label’.
2. The first column table stores target, and the second stores' data.
3. The data array is stored as features and samples and needed to be transposed to match the sklearn standard.

Fetch a machine learning data set, if the file does not exist, it is downloaded automatically from mldata.org.

sklearn.datasets package directly loads datasets using function: sklearn.datasets.fetch_mldata() 

Syntax: sklearn.datasets.fetch_mldata(dataname, target_name=’label’, data_name=’data’, transpose_data=True, data_home=None)

Parameters:

• dataname: (<str>) It is the name of the dataset on mldata.org, e.g: “Iris” , “mnist”, “leukemia”, etc.
• target_name: (optional, default: ‘label’) It accepts the name or index of the column containing the target values and needed to pass the default values of the label.
• data_name: (optional, default: ‘data’) It accepts the name or index of the column containing the data and needed to pass default values of data.
• transpose_data: (optional, default: True) The default value passed is true, and if True, it transposes the loaded data.
• data_home: (optional, default: None) It loads cache folder for the datasets. By default, all sklearn data is stored in ‘~/scikit_learn_data’ subfolders.

Returns: data, (Bunch) Interesting attributes are: ‘data’, data to learn, ‘target’, classification labels, ‘DESCR’, description of the dataset, and ‘COL_NAMES’, the original names of the dataset columns.

Let's see the examples:

Example 1: Load the ‘iris’ dataset from mldata, which needs to be transposed.

# import fetch_mldata function
from sklearn.datasets.mldata import fetch_mldata

# load data and transpose data
iris = fetch_mldata('iris', 
                    transpose_data = False)

# iris data is very large 
# so print the dataset shape
# print(iris)
print(iris.data.shape)

Output: 

(4,150)

Example 2: Load the MNIST digit recognition dataset from mldata. 

# import fetch_mldata function
from sklearn.datasets.mldata import fetch_mldata

# load data  
mnist = fetch_mldata('MNIST original')

# mnist data is very large
# so  print the shape of data
print(mnist.data.shape)

Output:

 (70000, 784)

Note: This post is according to Scikit-learn (version 0.19).