How to split/partition a dataset into training and test datasets for, e.g., cross validation?

What is a good way to split a NumPy array randomly into training and testing/validation dataset? Something similar to the cvpartition or crossvalind functions in Matlab.

If you want to split the data set once in two parts, you can use numpy.random.shuffle, or numpy.random.permutation if you need to keep track of the indices (remember to fix the random seed to make everything reproducible):

import numpy
# x is your dataset
x = numpy.random.rand(100, 5)
numpy.random.shuffle(x)
training, test = x[:80,:], x[80:,:]

or

import numpy
# x is your dataset
x = numpy.random.rand(100, 5)
indices = numpy.random.permutation(x.shape[0])
training_idx, test_idx = indices[:80], indices[80:]
training, test = x[training_idx,:], x[test_idx,:]

There are many ways other ways to repeatedly partition the same data set for cross validation. Many of those are available in the sklearn library (k-fold, leave-n-out, ...). sklearn also includes more advanced "stratified sampling" methods that create a partition of the data that is balanced with respect to some features, for example to make sure that there is the same proportion of positive and negative examples in the training and test set.

There is another option that just entails using scikit-learn. As scikit's wiki describes, you can just use the following instructions:

from sklearn.model_selection import train_test_split

data, labels = np.arange(10).reshape((5, 2)), range(5)

data_train, data_test, labels_train, labels_test = train_test_split(data, labels, test_size=0.20, random_state=42)

This way you can keep in sync the labels for the data you're trying to split into training and test.

Just a note. In case you want train, test, AND validation sets, you can do this:

from sklearn.cross_validation import train_test_split

X = get_my_X()
y = get_my_y()
x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
x_test, x_val, y_test, y_val = train_test_split(x_test, y_test, test_size=0.5)

These parameters will give 70 % to training, and 15 % each to test and val sets. Hope this helps.

As sklearn.cross_validation module was deprecated, you can use:

import numpy as np
from sklearn.model_selection import train_test_split
X, y = np.arange(10).reshape((5, 2)), range(5)
 
X_trn, X_tst, y_trn, y_tst = train_test_split(X, y, test_size=0.2, random_state=42)

You may also consider stratified division into training and testing set. Startified division also generates training and testing set randomly but in such a way that original class proportions are preserved. This makes training and testing sets better reflect the properties of the original dataset.

import numpy as np  

def get_train_test_inds(y,train_proportion=0.7):
    '''Generates indices, making random stratified split into training set and testing sets
    with proportions train_proportion and (1-train_proportion) of initial sample.
    y is any iterable indicating classes of each observation in the sample.
    Initial proportions of classes inside training and 
    testing sets are preserved (stratified sampling).
    '''

    y=np.array(y)
    train_inds = np.zeros(len(y),dtype=bool)
    test_inds = np.zeros(len(y),dtype=bool)
    values = np.unique(y)
    for value in values:
        value_inds = np.nonzero(y==value)[0]
        np.random.shuffle(value_inds)
        n = int(train_proportion*len(value_inds))

        train_inds[value_inds[:n]]=True
        test_inds[value_inds[n:]]=True

    return train_inds,test_inds

y = np.array([1,1,2,2,3,3])
train_inds,test_inds = get_train_test_inds(y,train_proportion=0.5)
print y[train_inds]
print y[test_inds]

This code outputs:

[1 2 3]
[1 2 3]

I wrote a function for my own project to do this (it doesn't use numpy, though):

def partition(seq, chunks):
    """Splits the sequence into equal sized chunks and them as a list"""
    result = []
    for i in range(chunks):
        chunk = []
        for element in seq[i:len(seq):chunks]:
            chunk.append(element)
        result.append(chunk)
    return result

If you want the chunks to be randomized, just shuffle the list before passing it in.

After doing some reading and taking into account the (many..) different ways of splitting the data to train and test, I decided to timeit!

I used 4 different methods (non of them are using the library sklearn, which I'm sure will give the best results, giving that it is well designed and tested code):

1. shuffle the whole matrix arr and then split the data to train and test
2. shuffle the indices and then assign it x and y to split the data
3. same as method 2, but in a more efficient way to do it
4. using pandas dataframe to split

method 3 won by far with the shortest time, after that method 1, and method 2 and 4 discovered to be really inefficient.

The code for the 4 different methods I timed:

import numpy as np
arr = np.random.rand(100, 3)
X = arr[:,:2]
Y = arr[:,2]
spl = 0.7
N = len(arr)
sample = int(spl*N)

#%% Method 1:  shuffle the whole matrix arr and then split
np.random.shuffle(arr)
x_train, x_test, y_train, y_test = X[:sample,:], X[sample:, :], Y[:sample, ], Y[sample:,]

#%% Method 2: shuffle the indecies and then shuffle and apply to X and Y
train_idx = np.random.choice(N, sample)
Xtrain = X[train_idx]
Ytrain = Y[train_idx]

test_idx = [idx for idx in range(N) if idx not in train_idx]
Xtest = X[test_idx]
Ytest = Y[test_idx]

#%% Method 3: shuffle indicies without a for loop
idx = np.random.permutation(arr.shape[0])  # can also use random.shuffle
train_idx, test_idx = idx[:sample], idx[sample:]
x_train, x_test, y_train, y_test = X[train_idx,:], X[test_idx,:], Y[train_idx,], Y[test_idx,]

#%% Method 4: using pandas dataframe to split
import pandas as pd
df = pd.read_csv(file_path, header=None) # Some csv file (I used some file with 3 columns)

train = df.sample(frac=0.7, random_state=200)
test = df.drop(train.index)

And for the times, the minimum time to execute out of 3 repetitions of 1000 loops is:

• Method 1: 0.35883826200006297 seconds
• Method 2: 1.7157016959999964 seconds
• Method 3: 1.7876616719995582 seconds
• Method 4: 0.07562861499991413 seconds

I hope that's helpful!

Split into train test and valid

x =np.expand_dims(np.arange(100), -1)


print(x)

indices = np.random.permutation(x.shape[0])

training_idx, test_idx, val_idx = indices[:int(x.shape[0]*.9)], indices[int(x.shape[0]*.9):int(x.shape[0]*.95)],  indices[int(x.shape[0]*.9):int(x.shape[0]*.95)]


training, test, val = x[training_idx,:], x[test_idx,:], x[val_idx,:]

print(training, test, val)

Here is a code to split the data into n=5 folds in a stratified manner

% X = data array
% y = Class_label
from sklearn.cross_validation import StratifiedKFold
skf = StratifiedKFold(y, n_folds=5)
for train_index, test_index in skf:
    print("TRAIN:", train_index, "TEST:", test_index)
    X_train, X_test = X[train_index], X[test_index]
    y_train, y_test = y[train_index], y[test_index]

Thanks pberkes for your answer. I just modified it to avoid (1) replacement while sampling (2) duplicated instances occurred in both training and testing:

training_idx = np.random.choice(X.shape[0], int(np.round(X.shape[0] * 0.8)),replace=False)
training_idx = np.random.permutation(np.arange(X.shape[0]))[:np.round(X.shape[0] * 0.8)]
    test_idx = np.setdiff1d( np.arange(0,X.shape[0]), training_idx)

Likely you will not only need to split into train and test, but also cross validation to make sure your model generalizes. Here I am assuming 70% training data, 20% validation and 10% holdout/test data.

Check out the np.split:

> If indices_or_sections is a 1-D array of sorted integers, the entries > indicate where along axis the array is split. For example, [2, 3] > would, for axis=0, result in > > ary[:2] ary[2:3] ary[3:]

t, v, h = np.split(df.sample(frac=1, random_state=1), [int(0.7*len(df)), int(0.9*len(df))]) 

I'm aware that my solution is not the best, but it comes in handy when you want to split data in a simplistic way, especially when teaching data science to newbies!

def simple_split(descriptors, targets):
    testX_indices = [i for i in range(descriptors.shape[0]) if i % 4 == 0]
    validX_indices = [i for i in range(descriptors.shape[0]) if i % 4 == 1]
    trainX_indices = [i for i in range(descriptors.shape[0]) if i % 4 >= 2]

    TrainX = descriptors[trainX_indices, :]
    ValidX = descriptors[validX_indices, :]
    TestX = descriptors[testX_indices, :]

    TrainY = targets[trainX_indices]
    ValidY = targets[validX_indices]
    TestY = targets[testX_indices]

    return TrainX, ValidX, TestX, TrainY, ValidY, TestY

According to this code, data will be split into three parts - 1/4 for the test part, another 1/4 for the validation part, and 2/4 for the training set.

Yet another pure numpy way to split the dataset. This solution is based on numpy.split which has already been mentioned before but I add this here for reference.

# Dataset
dataset = np.load(...)                      # Dataset of shape N x (d1 ... dM)

# Splitting and shuffling with indexes
idx = np.arange(len(dataset))               # Vector of dataset samples idx
id_train = int(len(idx) * 0.8)              # Train 80%
id_valid = int(len(idx) * (0.8 + 0.05))     # Valid 5%, Test 15%
train, valid, test = np.split(idx, (id_train, id_valid))

# Indexing dataset subsets
dataset_train = dataset[train]              # Train set
dataset_valid = dataset[valid]              # Valid set
dataset_test = dataset[test]                # Test set