onehotdecode

Decode probability vectors into class labels

Description

example

A = onehotdecode(B,classes,featureDim) decodes each probability vector in B to the most probable class label from the labels specified by classes. featureDim specifies the dimension along which the probability vectors are defined. The function decodes the probability vectors into class labels by matching the position of the highest value in the vector with the class label in the corresponding position in classes. Each probability vector in A is replaced with the value of classes that corresponds to the highest value in the probability vector.

example

A = onehotdecode(B,classes,featureDim,typename) decodes each probability vector in B to the most probable class label and returns the result with data type typename. Use this syntax to obtain decoded class labels with a specific data type.

Examples

collapse all

Use the onehotencode and onehotdecode functions to encode a set of labels into probability vectors and decode them back into labels.

Create a vector of categorical labels.

colorsOriginal = ["red" "blue" "red" "green" "yellow" "blue"];
colorsOriginal = categorical(colorsOriginal)
colorsOriginal = 1x6 categorical
red      blue      red      green      yellow      blue

Determine the classes in the categorical vector.

classes = categories(colorsOriginal);

One-hot encode the labels into probability vectors by using the onehotencode function. Encode the probability vectors into the first dimension.

colorsEncoded = onehotencode(colorsOriginal,1)
colorsEncoded = 4×6

0     1     0     0     0     1
0     0     0     1     0     0
1     0     1     0     0     0
0     0     0     0     1     0

Use onehotdecode to decode the probability vectors.

colorsDecoded = onehotdecode(colorsEncoded,classes,1)
colorsDecoded = 1x6 categorical
red      blue      red      green      yellow      blue

The decoded labels match the original labels.

Use onehotdecode to decode a set of probability vectors into the most probable class for each observation.

Create a set of 10 random probability vectors. The vectors express the probability that an observation belongs to one of five classes.

numObs = 10;
numClasses = 5;

prob = rand(numObs,numClasses);

tot = sum(prob,2);
prob = prob./tot;

Define the set of five classes.

classes = ["Red" "Yellow" "Green" "Blue" "Purple"];

Decode the probabilities into the most probable classes. The probability vectors are encoded into the second dimension, so specify the dimension containing encoded probabilities as 2. Obtain the most probable classes as a vector of strings.

result = onehotdecode(prob,classes,2,"string")
result = 10x1 string
"Red"
"Yellow"
"Yellow"
"Green"
"Yellow"
"Blue"
"Green"
"Yellow"
"Red"
"Red"

Input Arguments

collapse all

Probability vectors to decode, specified as a numeric array.

Values in B must be between 0 and 1. If a probability vector in B contains NaN values, the function decodes that observation to the class with the largest probability that is not NaN. If an observation contains only NaN values, the function decodes that observation to the first class label in classes.

Data Types: single | double

Classes, specified as a cell array of character vectors, a string vector, a numeric vector, or a two-dimensional character array.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | string | cell | char

Dimension containing probability vectors, specified as a positive integer.

Use featureDim to specify the dimension in B that contains the probability vectors. The function replaces each vector in B along the specified dimension with the element of classes in the same position as the highest value along the vector.

The dimension of B specified by featureDim must have length equal to the number of classes specified by classes.

Data type of decoded labels, specified as a character vector or a string scalar.

Valid values of typename are 'categorical', 'string', and numeric types such as 'single' and 'int64'. If you specify a numeric type, classes must be a numeric vector.

Example: 'double'

Data Types: char | string

Output Arguments

collapse all

Decoded class labels, returned as a categorical array, a string array, or a numeric array.