onehotdecode

Decode probability vectors into class labels

Since R2021b

Syntax

A = onehotdecode(B,classes,featureDim)

A = onehotdecode(B,classes,featureDim,typename)

Description

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.

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

One-Hot Encode and Decode Labels

Encode a set of labels into probability vectors, and then decode them back into labels.

Create a vector of categorical labels specifying color types.

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)

classes = 4x1 cell
    {'blue'  }
    {'green' }
    {'red'   }
    {'yellow'}

One-hot encode the labels into probability vectors by using the onehotencode function. Encode the labels into the first dimension, so that each row corresponds to a class and each column corresponds to a probability vector.

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

Decode the probability vectors by using the onehotdecode function.

colorsDecoded = onehotdecode(colorsEncoded,classes,1)

colorsDecoded = 1x6 categorical
     red      blue      red      green      yellow      blue 

The decoded labels match the original labels.

One-Hot Decode Dummy Variables

Create dummy variables, and then decode them back into the original data.

Create a column vector of categorical data specifying color types.

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

colorsOriginal = 6x1 categorical
     red 
     blue 
     red 
     green 
     yellow 
     blue 

Determine the classes in the categorical vector.

classes = categories(colorsOriginal);

Create dummy variables for each color type by using the dummyvar function.

dummyColors = dummyvar(colorsOriginal)

dummyColors = 6×4

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

Decode the dummy variables in the second dimension by using the onehotdecode function.

colorsDecoded = onehotdecode(dummyColors,classes,2)

colorsDecoded = 6x1 categorical
     red 
     blue 
     red 
     green 
     yellow 
     blue 

The decoded variables match the original color types.

Decode Probability Vectors into Most Probable Classes

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

prob = 10×5

    0.2938    0.0568    0.2365    0.2546    0.1582
    0.3895    0.4174    0.0154    0.0137    0.1641
    0.0427    0.3217    0.2854    0.0931    0.2573
    0.2878    0.1529    0.2943    0.0145    0.2505
    0.2640    0.3341    0.2834    0.0405    0.0780
    0.0422    0.0614    0.3280    0.3564    0.2120
    0.1078    0.1632    0.2876    0.2689    0.1725
    0.1940    0.3249    0.1392    0.1125    0.2293
    0.2356    0.1949    0.1613    0.2338    0.1745
    0.3345    0.3326    0.0593    0.0119    0.2616

Define the set of five classes.

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

Decode the probabilities into the most probable classes by using the onehotdecode function. The probability vectors are encoded into the second dimension (each column corresponds to a unique class), so specify the dimension containing encoded probabilities as 2. Obtain the most probable classes as a string vector.

result = onehotdecode(prob,classes,2,"string")

result = 10x1 string
    "Red"
    "Yellow"
    "Yellow"
    "Green"
    "Yellow"
    "Blue"
    "Green"
    "Yellow"
    "Red"
    "Red"

One-Hot Decode Classification Scores

Decode predicted class scores into predicted labels.

Load the fisheriris data set. Create X as a numeric matrix that contains four measurements for 150 irises. Create S as a vector of categorical labels that contains the corresponding iris species.

load fisheriris
X = meas;
S = categorical(species);

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

Y = onehotencode(S,2);

Compute the fitted coefficients of a simple linear classifier.

B = X\Y

B = 4×3

    0.0834    0.2117   -0.1481
    0.2533   -0.3059    0.1412
   -0.2270    0.1888    0.0181
   -0.0635   -0.5749    0.5873

Predict the class scores from the fitted coefficients, and ensure that the scores are in the range [0,1].

scores = X*B;
scores = min(1,max(0,scores));

Decode the predicted class scores into predicted labels by using the onehotdecode function. Then, create a confusion chart to compare the true labels S with the predicted labels label.

label = onehotdecode(scores,categories(S),2);
confusionchart(S,label)

Input Arguments

B — Probability vectors

numeric array

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 — Classes

cell array | string vector | numeric vector | character array

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

featureDim — Dimension containing probability vectors

positive integer

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.

typename — Data type of decoded labels

'categorical' (default) | character vector | string scalar

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

A — Decoded class labels

categorical array (default) | string array | numeric array

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

Version History

Introduced in R2021b