Image transformers that do not use domain transformations

[markotoplak]

Some image transformation methods need the whole image because they do both parameter estimation (=learning) and actual transformation (using learned parameters) in a single operation. Therefore, when transforming, instances are not independent and thus this can not be solved with domain transformation in Orange.

[markotoplak]

@borondics and @ngergihun, does this solve problems due to chunking?

This is the "latter option" from yesterdays email - which I believe is actually the only possible solution for some methods and thus unavoidable.

[ngergihun]

Hi @markotoplak !
It seems to work. The polynomial fit works just fine. However, the line leveling seems to do the leveling along the opposite direction of the image as expected. I think the data is reshaped in the opposite way than pySNOM transformer expects it.

Do you think, that this approach work for the referencing?

[markotoplak]

Yes, this approach would also work for references, with the only addition that the reference would need to get restructured in the same way.

[markotoplak]

@ngergihun, so what is the structure that pySNOM expects? If the first (numpys) axis Y or X?

[ngergihun]

@markotoplak if I swap the attr_x and attr_y in PreprocessImageOpts2DOnlyWhole like:

hypercube, _, indices = get_ndim_hyperspec(
                data, (image_opts["attr_y"], image_opts["attr_x"])
            )

Then it work all fine for me!

[markotoplak]

@ngergihun, thanks for the comment, I fixed accordingly and added a comment about it.

[ngergihun]

@markotoplak in the meanwhile I tried to make the referencing work. Now it is a kind of very simple approach that I reashpe the reference.X to the same shape as the image from the data

class SelfRef(PreprocessImageOpts2DOnlyWhole):
    def __init__(self, reference):
        self.reference = reference.X

    def transform_image(self, image):
        self.reference = np.reshape(self.reference,np.shape(image))
        d = SelfReference(referencedata=self.reference).transform(image)
        return d

I don't know it is a good approach but for now, it works. Should I make a PR from that version at all?

[markotoplak]

@ngergihun, can you open a PR with your modifications? Thanks!

What you suggest can't work in a general case. It probably works because the data in the table is ordered in just the right way (which is a tough assumption).

The same image reshaping operations as for the data will also have to be executed for the reference. I can do that part. what you can help me with is provide me the data where this works with your current modification, so it will be easier to ensure I did not break anything.

[ngergihun]

Yes, I agree, that it is generally a bad solution and only work under very strict circumstances. I open the PR and send some data.

[ngergihun]

@markotoplak Another thing came to my mind. So in the new PreprocessImageOpts2DOnlyWhole class the image is passed to the transform_image(self, image) method. When we reimplement the method in a preprocessor then we can access the reshaped image but cannot access the attributes of data where we store the datatype of the image data. Is there an easy solution, to also pass the attributes?

[markotoplak]

@ngergihun, thanks for the comment. I'll add another parameter, data.