ANN_MLP huge training sets

asked 2017-12-11 15:27:58 -0600

1170 ●4 ●18

updated 2017-12-11 15:30:20 -0600

I am having difficulty understanding the artificial neural network in OpenCV.

I found a code sample that trains the ANN to perform the XOR operation. The code looks something like this:

TermCriteria termCrit = TermCriteria(TermCriteria::Type::COUNT + TermCriteria::Type::EPS, 10000, 0.000001);
mlp->setTermCriteria(termCrit);

Ptr<TrainData> trainingData = TrainData::create(inputTrainingData, SampleTypes::ROW_SAMPLE, outputTrainingData);

mlp->train(trainingData);

.. which, as I understand, trains the network for up to 10000 times while looking for a maximum error of 0.000001.

This code says to me that all of the training data must be made available to the ANN before training occurs. Is this true, and if so, what if the training data consists of thousands of images?

I tried the following code, but it appears that the training is reset whenever mlp->train() is called:

TermCriteria termCrit = TermCriteria(TermCriteria::Type::COUNT + TermCriteria::Type::EPS, 1, 0.000001);
mlp->setTermCriteria(termCrit);

Ptr<TrainData> trainingData = TrainData::create(inputTrainingData, SampleTypes::ROW_SAMPLE, outputTrainingData);

for(size_t i = 0; i < 10000; i++)
mlp->train(trainingData);

Is there no way to continue the training instead of resetting it each time?

answered 2017-12-11 15:36:44 -0600

LBerger
9317 ●2 ●20 ●88 http://www.traimaocv.fr

updated 2017-12-11 15:37:55 -0600

use flag UPDATE_WEIGHTS in train method

edit flag offensive delete link

Comments

Thank you! That seems to do the trick as long as I call train without UPDATE_WEIGHTS first, then do the rest of the iterations with UPDATE_WEIGHTS:

Ptr<TrainData> trainingData = TrainData::create(inputTrainingData, SampleTypes::ROW_SAMPLE, outputTrainingData);

mlp->train(trainingData);

for(int i = 1; i < 10000; i++)
    mlp->train(trainingData, ANN_MLP::TrainFlags::UPDATE_WEIGHTS);

sjhalayka ( 2017-12-11 15:43:02 -0600 )edit

@LBerger -- When doing training based on an image, I was wondering if you'd recommend converting the many row 2D image Mats to single row 1D Mats? Thanks again!

sjhalayka ( 2017-12-11 20:36:39 -0600 )edit

use reshape method

LBerger ( 2017-12-12 03:49:07 -0600 )edit

Thank you so much! I will have reshaped them to 1D, and will have num_columns*4 input neurons; one input neuron per r, g, b, and a, times the number of pixels.

sjhalayka ( 2017-12-12 09:14:51 -0600 )edit

Don't forget data type must be CV_32FC or CV_64FC

LBerger ( 2017-12-12 12:39:21 -0600 )edit

OK. Thank you for the reminder. As well, the network's input values and output values must each be within [0, 1]. The output values can be (precisely) 0 or 1, more like it -- "the output [can] range from [-1.7159, 1.7159]": https://docs.opencv.org/3.0-beta/modu...

Edit: It turns out that you cannot train for values like 1 when using UPDATE_WEIGHTS. See: http://answers.opencv.org/question/18...

sjhalayka ( 2017-12-12 13:34:39 -0600 )edit

No. I haven't got enough time to test your code there is no relation between UPDAE_WEIGHTS and values. Problem is in predict. If you want to know values in output layers you must get response in predict function

LBerger ( 2017-12-14 07:40:48 -0600 )edit

add a comment

ANN_MLP huge training sets

1 answer

Comments

Links

Question Tools

Stats

Related questions

ANN_MLP huge training sets edit

1 answer

Comments

Links

Question Tools

Stats

Related questions

ANN_MLP huge training sets