sample_embedding_folder2/0765687.txt

Ticket Name: CCS/TDA2: Adaboost train

Query Text:
Part Number: TDA2 Tool/software: Code Composer Studio I used github.com/.../acf-jacinto train my data , when I used small amount of data to train,the model no problem.But if the positive image large (more than 10000), the adaboost going to happen stopping eraly. I changed the code, if not used cellSum to deal with the feature channel,the above situation would not happened,but the train weights cannot used in TDA2X object detecion .How can I deal with it or how train large dataset ?

Responses:
I did not exactly understand the problem, but can you add one line after the line: github.com/.../acfJacintoTrainTest.m Please add the following and then do training: opts.nPos=10000; %num positive to be collected

Thanks for you reply. Add the opts.nPos can't solve the problem. I use opts.pPyramid.pChns.pFastMode.enabled=1; %default: 0 Adaboost early stopping,train log as follows Sampling windows completed=100.0% [elapsed=35.4s / remaining~=0.0s] Sampled 30000 windows from 2048 images. Done sampling windows (time=37s). Extracting features... done (time=14s). Training AdaBoost: nWeak= 32 nFtrs=2560 pos=10150 neg=30000 i= 16 alpha=5.000 err=0.000 loss=9.03e-34 stopping early Done training err=0.0000 fp=0.0000 fn=0.0000 (t=0.3s). Done training stage 0 (time=119s). --------------------------------------------------------------------------- Training stage 1 Sampling windows completed=100.0% [elapsed=123.8s / remaining~=0.0s] Sampled 10000 windows from 10176 images. Done sampling windows (time=126s). Extracting features... done (time=13s). Training AdaBoost: nWeak=128 nFtrs=2560 pos=10150 neg=30000 i= 16 alpha=5.000 err=0.000 loss=1.80e-35 stopping early But if used opts.pPyramid.pChns.pFastMode.enabled=0; %default: 0 Adaboost train OK. On the pFastMode.enabled = 1, The adaboost train easy to appear with early stopping on my training sample. How do I handle this situation?(I check the code in chnsCompute.m,it use cellSum to deal with channel feature,Maye be caused it?)

I am not sure about the reason for this issue as I have not faced this with the datasets that I have tried. Question 1: As I understand, this issue happens only if your positive samples are more than 10000. Can you try to manually select a smaller set of positive samples, so that this issue is not there? Is this an acceptable solution. Question 2: As I understand, you are able to train properly if you set: opts.pPyramid.pChns.pFastMode.enabled=0; Have you tried using the output descriptor (trained with this modification) in TDA2x object detection?

Hi @user5920690 , We haven't heard back from you, I'm assuming you were able to resolve your issue. If not, just post a reply below (or create a new thread if the thread has locked due to time-out). Regards, Yordan