Regression refactoring : TrainVectorRegression

Summary

This MR introduces a new application, TrainVectorRegression for training a regression machine learning model from vector data, in the same fashion as TrainVectorClassifier

Rationale

See #1799 (closed)

This MR includes two major changes:

• The application TrainVectorBase (the base class for TrainVectorClassifier), is now template on TInputValue (features) and TOutputValues (class). Before the MR it only works on float as feature type and int as class type (classification case), but now it can also be used for other type like float,float (regression case).

• A new application TrainVectorRegression deriving from TrainVectorBase

Tests

A test has been added for the new application, using a rf classifier as regression algorithm. In the end all regression algorithm should be tested, but I think we can do that in the (future) TrainImagesRegression, to keep the same testing strategy as for classification.

Additional notes

This is not exactly the workflow described in the issue, because I don't think the first step is relevent (removing sampling from TrainRegrssion), as TrainRegression will be deprecated at the end of the refactoring.

The next step of the refactoring is to create a TrainImagesRegression application: it could be a composite application that chains ImageEnvelope to create a polygon on the extent of the image, SampleSelection to select random points over this polygon, SampleExtraction to extract feature and predictor values over two input images and finally TrainVectorRegression to extract the model (this is the workflow used in the KMeansClassification composite application), what do you think ? Anyway I think this is out of the scope of this MR.

In the issue we talked about CSV input compatibility. It is hard to add it in TrainVectorBase because of the design of the application, the best way (given the design of the learning applications) would probably to create a new application TrainCSVBase inheriting from LearningApplicationBase doing the CSV input reading, and then create a TrainCSVRegression from it, and maybe also a TrainCSVClassifier. But it there really a need for such functionality ?

Copyright

The copyright owner is CNES and has signed the ORFEO ToolBox Contributor License Agreement.

---

Check before merging:

• All discussions are resolved
• At least 2 votes from core developers, no vote.
• The feature branch is (reasonably) up-to-date with the base branch
• Dashboard is green
• Copyright owner has signed the ORFEO ToolBox Contributor License Agreement
• Optionally, run git diff develop... -U0 --no-color | clang-format-diff.py -p1 -i on latest changes and commit

changed milestone to %7.0.0

added app refactoring labels

added 54 commits

• 8e06c83c...0df44b31 - 51 commits from branch develop
• b948c575 - Merge branch 'develop' into regression_refactoring
• 3e071bca - ENH: use mse instead of model as text baseline for regression
• 5b5c03d6 - BUG: add additional mse baseline for OpenCV 3

Compare with previous version

added 2 commits

• 40e15141 - ENH: replace GetfieldAs... by GetValue<..>
• 2fabfa17 - ENH: use reference instead of pointer in mse computation

Compare with previous version

changed title from Regression refactoring to Regression refactoring : TrainVectorRegression

changed the description

Status: need review and vote

Also need opinions on the the workflow for the next step of the refactoring (composite application for trainImagesRegression, and no CSV application, see the additional notes of the Merge Request)

added CNES backlog label

added 1 deleted label

added 1 commit

• 1d3baef0 - DOC: typo in TrainVectorRegression

Compare with previous version

pushed a minor typo fix

added 14 commits

• 1d3baef0...cb2f6d1b - 10 commits from branch develop
• b0e27774 - Merge branch 'develop' into regression_refactoring
• c90877f5 - COMP: remove SetDocName
• 12ec94ae - STY: run clang format
• 5e79e868 - COMP: resolve merge conflict

Compare with previous version

resolved all discussions

merged

mentioned in commit 78d7f56c

Just one comment about TrainImageRegression: I think it would be a limitation of usage to generate the vector support from sampling though ImageEnveloppe : one might want to perform regression only for a given class or subset of the image.

So how you see it, the application takes as input a shapefile containing the vector support, a predictor image and a list of feature images ? or a shapefile containing the vector support and the predictor as features, and a list of feature images ?

There are several possibility regarding the IOs of this application, and we should agree on which one to use before starting the implementation.

The idea behind the use of ImageEnveloppe was to provide an application that works with feature and predictor images as input (and no vector data), which is how kmeansClassification works (this app could be called TrainVectorUnsupervised if we had more than one unsupervised algorithm), and is also the API of the current TrainRegression application.

I think predictor and feature shall remain images, but allowing to restrict the pixels used for regression using a vector data instead of forcing the use of the image enveloppe would be great.

I can imagine some use cases that would be very hard to implement without this, for instance if you want to train the regression algorithm only on a certain category of vegetation.

mentioned in merge request !496 (merged)

mentioned in issue #1682 (closed)