File Exchange

image thumbnail

Classic AdaBoost Classifier

version 1.5 (4.07 KB) by

Weak threshold classifier boosted to strong Classifier with Adaboost

4.61538
13 Ratings

143 Downloads

Updated

View License

This a classic AdaBoost implementation, in one single file with easy understandable code.

The function consist of two parts a simple weak classifier and a boosting part:
The weak classifier tries to find the best threshold in one of the data dimensions to separate the data into two classes -1 and 1
The boosting part calls the classifier iteratively, after every classification step it changes the weights of miss-classified examples. This creates a cascade of "weak classifiers" which behaves like a "strong classifier"
.
Training mode:
  [estimateclass,model]=adaboost('train',datafeatures,dataclass,itt)
Apply mode:
 estimateclass=adaboost('apply',datafeatures,model)
 
inputs/outputs:
   datafeatures : An Array with size number_samples x number_features
   dataclass : An array with the class off all examples, the class
                 can be -1 or 1
   itt : The number of training iterations
   model : A struct with the cascade of weak-classifiers
   estimateclass : The by the adaboost model classified data
            
.
Please leave a comment, if you like the code, find a bug or have a suggestion.

Comments and Ratings (26)

PLZ PROVIDE adBoost CLASSIFICATION EXAMPLES CODES

fafz1203

jayapriya p

good work .... can u provide us code for integral image

Anuja Kelkar

Error using accumarray
First input SUBS must contain positive integer subscripts.

Error in adaboost>WeightedThresholdClassifier (line 125)
h1f=accumarray([p1f(:) i1(:)],repmat(w1(:),ndims,1),[ntre ndims],[],0);

Error in adaboost (line 50)
            [estimateclass,err,h] =
            WeightedThresholdClassifier(datafeatures,dataclass,D);

Please help! I don't understand what's wrong!
 

BUSI HANUSRI

good work..can you provide us a code for adaboost estimator in ofdm

ankit dixit

ankit dixit (view profile)

Nice work dude, keep it up :-)

KimHan

KimHan (view profile)

in adaboost.m, there is a line with "...wrongly classified samples will have more weight", but this weight D is never used in the code.

  Hi Dirk-Jan Kroon,

In file ADABOOST_tr.m, there is a line as follow:

% The weight of the turn-th weak classifier
adaboost_model.weights(turn) =
 log10((1-error_rate)/error_rate);

The weight should be changed to

 adaboost_model.weights(turn) =
 log((1-error_rate)/error_rate);

i.e. not base 10 logarithm but natural logarithm.

Nikita

Nikita (view profile)

I get an OutOfMemory error, when Matlab is passing cumsum() function in

adaboost>WeightedThresholdClassifier (line 126)

while i have 8GB of RAM availble and Training set contains only 6,5 milions entries of double. Is it possible to replace the cumsum() function with sum() for example?

Hello Dirk-Jan Kroon,
Can you please let me know how can I use SVM as weak classifier? .Can you tell me how to put the Adaboost algorithm connect with the SVM ?
Thanks

lu li

lu li (view profile)

Thank you! It's very helpful and instructive.

mehregan

Hi Dirk-Jan Kroon, please can you help me
THIS CODE is great
and I need a paper to explain this code beacuse I do not underestand some parts of this code.
Thanks a lot
mehrzad

Nooshin

Can you please let me know how can I use SVM as weak classifier?
Thanks

Nooshin

Hi Dirk
Thanks for your code.
I had a question. For each subject I have a feature vector including 144 features. Class labels are 0 and 1. The number of subjects are 20. I used your code and accuracy was so low, while I used SVM's accuracy is high. What is its reason? Do you think too many features can affect the accuracy of adaboost? please let me know. Thanks a lot

Fa Fa

Fa Fa (view profile)

Hi Dirk-Jan Kroon, please can you help me,
i have 1000 faceimages and 1000 background and i have a histogram of each image, i have also 512 lookup table from 000000000 to 111111111 integer feature.i created the feature of the trainingimages with census transform.how can i save for every lookup table a value with adaboost?if the value of lookup table prefer to be face or non-face.
thanks

Andreas

Hi Dirk,
Are there any requirements on the features (such as no NaNs? numerical scaling? equal numbers in each class?). The internal variable p2c at line 117 has runs of large and NaN values, causing accumarray and the WeightedThresholdClassifier function to crash. Thanks - Andreas

Hi!
Can i use this code instead of Haar feature based AdaBoost cascade classifiers?
Actually i am not understanding the concept behind Adaboost and its very much compulsory for me to implement it in just 10 days.If i have to use the Adaboost cascade classifiers on the basis of the brake lights and taillights of the vehicles on the roads in evening hours then what will be parameters of this function?waiting for reply...please help

Su Dongcai

Su Dongcai (view profile)

Neatly codes, very nice demonstration

jinkyu do

sudheesh p

i am not able to run the adaboost m-file, it shows some eeror at the statement switch (mode)asking about error at mode and in the example m-file only the training pattern is visible

CarloG

CarloG (view profile)

AMVR

AMVR (view profile)

Great contribution, Dirk-Jan.

This is just a suggestion, but have you thought about adding probability calibration to the output of the classifier? e.g. logistic correction, platt scaling or isotonic regression (see "Obtaining Calibrated Probabilities from Boosting", by Alexandru Niculescu-Mizil and Rich Caruana for more context).

h gd

h gd (view profile)

Easy to understand

David

David (view profile)

Xu Cui code?

Hi Dirk-Jan Kroon,

Thank you for the nice comments making the code easily understandable.

Can I do feature selection using your code? I am new at Adaboost so am not able to figure out how can this be done. Can you please help me?

Updates

1.5

Fixed boundary bug

1.4

Speed improvement (Replaced loops by 1D indexing and bsxfun operations.)
The function now limits features of the test data to the outer-boundaries of training data.

1.3

Changed bug : ndims(datafeatures)to size(datafeatures,2)

1.2

Solved division by zero, causing NaN

1.1

Changed Screenshot and example figure

MATLAB Release
MATLAB 7.10 (R2010a)

Download apps, toolboxes, and other File Exchange content using Add-On Explorer in MATLAB.

» Watch video