@iarganda I trained a random forest classifier with the weka segmentation plugin. Is there a way of applying the classifier to a whole series of images without loading them into memory all at once? Also, how much effort would it be to port the feature calculation into the imglib2 realm? I have an
InstanceView that extends AbstractConvertedRandomAccessible< RealComposite< R >, Instance > and a CompositeInstance that wraps around RealComposite that I can use for predictions using an existing classifier. I do not know, however, how to fill the RandomAccessibleInterval< R > with features that are required for an existing classifier before passing it to my InstanceView.
@iarganda I am trying to load a trained classifier, what am I doing wrong? Minimum working (or failing) example:
final WekaSegmentation seg = new WekaSegmentation();
final String cp = "/data/hanslovskyp/hk-test/ds-8/2017-02-21-7_40.model";
seg.loadClassifier( cp );Output (stdout & error):
Field of view: max sigma = 16.0, min sigma = 0.0
Membrane thickness: 1, patch size: 19
Read class name: class 1
Read class name: class 2
Error while adjusting data!
java.lang.NullPointerException
trainableSegmentation.WekaSegmentation.adjustSegmentationStateToData(WekaSegmentation.java:4141)
trainableSegmentation.WekaSegmentation.loadClassifier(WekaSegmentation.java:712)
de.hanslovsky.zspacing.spark.experiments.mask.weka.Dummy.main(Dummy.java:12)
at trainableSegmentation.WekaSegmentation.adjustSegmentationStateToData(WekaSegmentation.java:4141)
at trainableSegmentation.WekaSegmentation.loadClassifier(WekaSegmentation.java:712)
at de.hanslovsky.zspacing.spark.experiments.mask.weka.Dummy.main(Dummy.java:12)
@hanslovsky 3D data?
No, 2D
Model as well as image to which you want to apply it?
Yes. I trained the model using the
Trainable Weka Segmentation plugin
So that should be 2D, right?
The images to which I want to apply the classifier are 2D as well.
But loading fails already when loading the classifier with the example above.
The exception happens in
this.featureStackArray.getEnabledFeatures() and this.featureStackArray is null.
at
WekaSegmentation.java:4141
WekaSegmentation() doesn't initialize featureStackArray
Stumbled over that before but never reported that behavior ( :-1: for me)
I switched to using the
WekaSegmentation(ImagePlus) constructor
Cool, thanks for the hint! Will that calculate features that I potentially might not need?
That's actually a good question..
@hanslovsky FYI, I haven't really noticed @iarganda very active in Gitter. He is very active on the ImageJ forum though.
I will use your trick for the short term. Right now I do not care too much about wasted feature computation. I will start a github issue.
Ok,thanks for the hint
@hanslovsky Sounds good! Also a small remark on feature computations: there is some work in Ops happening. Maybe you could ask over in the imagej-ops channel and ping @dietzc
that's dietzc ...
Great, I'll contact him.
@hanslovsky ...on the forum. :wink:
Or GitHub issue, I suppose. But public.
Yes, will do.
Thanks.
:plus1:
(@stelfrich do you know you can edit your gitter posts within a short time after posting to fix small errors, like mentioning the wrong person?)
(I know @imagejan. Seems like my fingers were way faster than my brain though)
I just opened the issue:
fiji/Trainable_Segmentation#35
fiji/Trainable_Segmentation#35
Sorry to jump late into the conversation, but yes, the empty constructor needs the training image to be set in order for the rest of methods to work.
@hanslovsky I will check if I can implement a workaround
@iarganda Cool, thank you. So if I use the
ImagePlus constructor or setTrainingImage, will WekaSegmentation compute any features or does this happen only when I call applyClassifier? I would like to avoid redundant or unncessary feature comptuations.
@hanslovsky only if you call applyClassifier, no worries!
Cool. I will add this part of the question to the github issue so other people can see it as well.
great, thanks!