My training has already been running for more than a week and now I don't want the number of stages I am training it for. I want to change the number of stages from 25 to 17-18. Do I need to retrain the whole classifier from the beginning? Please explain why.
Asked
Active
Viewed 743 times
-1
-
This question appears to be off-topic because it is about machine learning. – Aurelius Jun 10 '14 at 18:48
-
It is related to training haar classifiers using traincascade available in OpenCV therefore, I dont see anything wrong with the tags. – user3626948 Jun 11 '14 at 02:16
2 Answers
1
"Do I need to retrain the whole classifier from the beginning" - no.
you can ctrl-c it at any time. you will be loosing the currently trained stage though.
if you stop at say, stage 17, you can re-run your cmd with numstages 17, this will generate a cascade.xml, that you can use for testing.
later you can restart it with numstages 25, to calculate the missing 17-25 stages.
HAAR training takes very long. you probably should start with a LBP cascade first, even if you want a HAAR one later. the LBP one will be slightly less accurate, but trains in minutes instead of weeks. so if you're unsure about the fitness of your data, try that first. if LBP does not work at all, HAAR never will.
berak
- 39,159
- 9
- 91
- 89
0
EDIT: My original answer was completely wrong. And I have just realised why.
Yes you do need to retrain, however not to produce the same output as before. If you rerun trancascades across the same folder as before, but change the number of stages to 17 it will pick up the already generated classifier stage .xml files and create you a new cascade.xml file which will only be aware of the 17 stages. You Can then remove the additional stages.
on an aside
Running anything over 19 stages will likely take far longer than you have.
GPPK
- 6,546
- 4
- 32
- 57