0

I saw this question about improving SVM speed. The RandomForestClassifier works amazingly fast (~140 seconds on dataset of 50000 32X32 RGB images. Just to compare, with regular SVC it took me ~140 for 2500 !! images).

The problem with it is, that it has sub-optimal accuracy (<40% where the regular SVM with much less samples gave 50%). Is it possible to use it with linear / polynomial kernel like the regular SVC(C = 0.01, kernel='poly', degree = 3))?

CIsForCookies
  • 12,097
  • 11
  • 59
  • 124
  • Is it worth asking this at https://dsp.stackexchange.com/ ? – Élie Dec 27 '17 at 20:44
  • SVMs work best with normalized data. Is your data normalized/ standardized? – Vivek Kumar Dec 28 '17 at 05:29
  • Hmm... I don't really know what to answer. My data is decriptors (using sift) of images taken from cifar10 challange. Does that count as random? – CIsForCookies Dec 28 '17 at 10:10
  • First, try increasing `n_estimators` (default is 10 I think) and see if it improves cross validation accuracy (more estimators will slow down training however). Second, try playing around with `max_features`. – joergd Dec 28 '17 at 20:44

0 Answers0