How to use one hot encoding of string categorical features in keras?

Question

I am dealing with a binary classification problem. The output column of my dataset is already encoded in 0/1. The problem is that I have many categorical features (columns), which are strings and I would like to one-hot encode them.

I have 18 features (few features are integers and others are strings, the categorical ones) and 1 output column.

I tried this:

dataframe = pd.read_csv('basic_df_export.csv', sep=';', encoding = 'ISO-8859-1', header=None) 

dataset = dataframe.values
# split into input (X) and output (Y) variables
X = dataset[:,0:17]
Y = dataset[:,17]

# define example
encoded = to_categorical(X)
print(encoded)

but it doesn't work, throwing me this error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-41-318da09d7033> in <module>()
      9 
     10 # define example
---> 11 encoded = to_categorical(X)
     12 print(encoded)

~/anaconda/lib/python3.6/site-packages/keras/utils/np_utils.py in to_categorical(y, num_classes)
     21         is placed last.
     22     """
---> 23     y = np.array(y, dtype='int')
     24     input_shape = y.shape
     25     if input_shape and input_shape[-1] == 1 and len(input_shape) > 1:

ValueError: invalid literal for int() with base 10: 'photo'

Answer 1

For string data, use get_dummies() (from Pandas). to_categorical() takes integers as inputs.

There are two important differences between Keras: to_categorical() and Pandas: get_dummies().

---

Keras: to_categorical()

• to_categorical() takes integers as input (no strings allowed).
• to_categorical() generates dummies starting at 0 by default!

Looking at the help function:

print(help(to_categorical))

Says:

to_categorical(y, num_classes=None, dtype='float32')
    Converts a class vector (integers) to binary class matrix.

    E.g. for use with categorical_crossentropy.

    # Arguments
        y: class vector to be converted into a matrix
            (integers from 0 to num_classes).
        num_classes: total number of classes.
        dtype: The data type expected by the input, as a string
            (`float32`, `float64`, `int32`...)
...

So if your data is numeric (int), you can use to_categorical(). You can check if your data is an np.array by looking at .dtype and/or type().

import numpy as np
npa = np.array([2,2,3,3,4,4])
print(npa.dtype, type(npa))
print(npa)

Result:

int32 <class 'numpy.ndarray'>
[2 2 3 3 4 4]

Now you can use to_categorical():

from keras.utils import to_categorical
cat1 = to_categorical(npa)
print(cat1.dtype, type(cat1))
print(cat1)

Which yields a matrix:

float32 <class 'numpy.ndarray'>
[[0. 0. 1. 0. 0.]
 [0. 0. 1. 0. 0.]
 [0. 0. 0. 1. 0.]
 [0. 0. 0. 1. 0.]
 [0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 1.]]

Note that the matrix contains five columns (starting at zero up to four, which is my max. value in the np.array). The first two columns (representing 0 and 1 in the original data) are 0 in the whole matrix, because none of these values are found in the original data.

to_categorical() also takes input which is not explicitly defined as np.array. For instance the statements below would also be legal.

alt1 = to_categorical([0,0,1,1,2,2])
print(alt1.dtype, type(alt1))
print(alt1)

alt2 = to_categorical((0,0,1,1,2,2))
print(alt2.dtype, type(alt2))
print(alt2)

Because the range of values now is between 0 and 2, the result would look like:

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

---

Pandas: get_dummies()

When you have a Pandas df, you can convert some column to dummies using get_dummies(), regardless of the data type in the column. So it is also possible to convert a column of strings to dummies.

import pandas as pd
df = pd.DataFrame(data={'col1':["A", "A", "B", "B", "C", "C"]})
alt3 = pd.get_dummies(df['col1'])
print(type(alt3))

This gives:

<class 'pandas.core.frame.DataFrame'>
   A  B  C
0  1  0  0
1  1  0  0
2  0  1  0
3  0  1  0
4  0  0  1
5  0  0  1

Note that the result is (again) a Pandas df. So we need to convert it to a np.array.

alt3 = alt3.to_numpy()
print(alt3.dtype, type(alt3))
print(alt3)

This yields:

uint8 <class 'numpy.ndarray'>
[[1 0 0]
 [1 0 0]
 [0 1 0]
 [0 1 0]
 [0 0 1]
 [0 0 1]]

So that it is ready to be used with Keras.

Note that the matrix generated here does not (!) start at zero. Instead each distinct value in the chosen Pandas column gets it's own column in the dummy matrix.

Answer 2

Try:

X = dataset[:,0:17].astype(float).astype(int)

I think that if you have a string like '45.2', you will have to cast it as a floating-point first and from float you can cast them into integer.

I will be glad if an editor could corroborate/correct this answer.