Hi there

I am new with prtools. I am trying to create a dataset and train adaboost but i keep getting this error:
??? Error using ==> vertcat
CAT arguments dimensions are not consistent.

Error in ==> affine at 3
R = [cos(r) -sin(r) 0;

Error in ==> nmc at 52
W = affine([R -R],[offset -offset],a,getlablist(a));

Error in ==> map at 115
d = feval(mapp,a,pars{:});

Error in ==> E:\Tkaravides\prtools_ac\prtools\@mapping\mtimes.p>mtimes at 15

Error in ==> adaboostc at 84
w = b*clasf; % train weak classifier

It seems that I don’t give enough data. I used struct(A) to check my dataset settings and I found some strange results with ident and featdom but I can’t solve it with setident or something

Regards

Dear Thomasson,

could you post the code that leads to your error?

Pavel

Dear Pavel

my code is simple. I create a matrix Data <324x76500double> and a label column of <324x1double> then A=dataset(Data,labs);
and w=adaboostc(A)

with weakc it did not work but with treec it did. I think I have to see what treec does, and see if it suits me.

Thanks a lot

Thomasson

Dear Thomasson,

what a feature size! :-) I can train the adaboost using weakc on data with the similar number of samples and features.
One thing I noticed in your error message above: Don’t you have your own affine function on the Matlab path? The PRTools affine does not have any [cos(r) -sin(r) code in it:

Can you try ‘which affine’ to see what routine is actually called?

Pavel

Dear Thomasson,

Your dataset has a very high dimension (76500). Thereby linear classifiers generated by
weakc may have a problem. Treec makes a decision tree by thresholding the most informative
features recursively (and splitting the training set accordingly). This is still feasible
for high dimensions. You may also try stumpc, which is a weak version of treec.

Best regards,

Bob Duin

Dear Pavel

well it seems that since I have specified many file paths in my directory it uses mfiles from other parts of my project. this one is one that i use to collect SIFT features. Do you think I will have a problem with that? Thanks for the information. I have so many features because I use pixel values that I get from Harr like feature filtering. I even need more…

Thomasson

Dear Bob Duin

thanks for the reply. I am wondering if I can observe which these most informative features are, so I can use less features, because with that feature vector I have memory problems

regards
Thomasson

Dear Thomasson,

regarding the path, I think you do get the error when training the PRTools adaboost with weak base classifier simply because it requires PRTools affine.m mapping and you already have other affine.m on your path. Either rename your affine function or separate the feature computation and training steps (and keep only PRTools on Matlab path when you’re training, not your other code).

If you want to use adaboost to select the Haar-like features similarly to the Viola & Jones face detection algorithm, maybe just use decision stump base classifier (stumpc) as Bob suggested. Stumpc selects the feature based on some criterion and trains a threshold for that feature:

Hope it helps,

Pavel