search

pelkmanslab / TissueMAPS-OLD

Old TissueMAPS code (until 2016); please use pelkmanslab/TissueMAPS instead.
http://github.com/pelkmanslba/TissueMAPS
0 stars 0 forks source link

Classification parameters accessible from the gui #1

Open alecrimi opened 7 years ago

alecrimi commented 7 years ago

Some parameters for doing classification are hardcoded into the code and not accessible to users. The GUI should be expanded to allow the user to set manually those.

hackermd commented 7 years ago

I have optimised the grid search for the crossvalidation in tmlib.tools.base.Classifier.classifiy_supervised(). This improved classification accuracy (based on visual inspection). Here is a result for "randomforest" classifier:

screen shot 2017-01-25 at 18 49 00

For the "svm" classifier, I used values for penalty parameters as recommended by the paper "A practical guide to support vector classification". This looks good as well.

hackermd commented 7 years ago

I wouldn't expose the parameters in the user interface. They are not very intuitive and a cross-validation of a wide range of parameter values should do the trick. For an extensive cross-validation, it's important to have enough data points. So generally speaking, the more labels the better.

Tools can, in addition to a labeled layer, return a tmlib.models.result.Plot. We may want classifiers to return users additional graphical information about the classification result, such as a ROC curve.

alecrimi commented 7 years ago

I disagree, if a user is using a tool, (s)he is supposed to know what he is doing and (s)he should also be allowed to tune for the experiments. Let' s see if the results are satisfactory for a wide range of users, we leave it for now with the grid search and we don't add parameters.