yyh0806's profile - activity

I want to try opencv model in (KNN,Neural,Bayes,SVM) and decide which to use.I find Neural,Bayes have update paramters, just _neural.load("neural.xml") and _neural.train(train_data, train_class, Mat(), Mat(), params, CvANN_MLP::UPDATE_WEIGHTS); I don't know how to upload the code.

I trained model by my samples. I want to updated model using the new training data instead of retrain it. Because it takes too long time.I test this method by some samples. I guess that if I train the model by a full samples or train it by half samples , updated it by the other half, it will be the same result. But I was wrong.I don't understand why and I want to know that my idea is right or not.

Thanks for your answer.I trained A-Z 0-9 for 180 samples and accuracy is 90%. If I train 36 samples as one group, updated it by parameters "update" in train() 4 times(the same sample).accuracy is under 50%. How can I train the samples one by one and get a good result.

I try to add new train data to an already trained model. In neural, I find that it can be trained by " UPDATE_WEIGHTES",however it doesn't work well. I want to recognize 0-9, use 180 samples train it, its ok. when I seperate samples to 5, train it by UPDATE_WEIGHTES 4times , its going so far.

I have tried Bayes, KNN, Neural. Now I don't understand what does the 'update' in train() used for. sorry for my poor english