Remove any non-scipy stack dependencies from notebooks

[karlnapf]

Many of the notebooks have dependencies that are not from the scipy stack (i.e. numpy, scipy, pandas, etc). Those regularly cause a lot of headaches in our build infrastructure. This task is to go through the notebooks (grep for import), and try to get rid of those imports, and use the available tools to do the job.

Example:

• Image in the PCA notebook
• networkx in the FGM and neural net digits and RBM notebooks

[rahul13ramesh]

Hey, I am new to shogun and would like to take up this issue. Here are a few points that I had queries about

• Most imports(in ipython notebooks) use modshogun. Should this be changed to shogun as per #3898?

• The following imports require removal

  • sklearn
  • networkx
  • Image(Pillow)
  • cv2

  The issue is that some of the functionality cannot be replicated using the scipy stack. For example. cv2, has some image manipulations( cv2.approxPolyDP) not easily replicable in scipy.ndimage. Another example is the usage of networkx.draw(), which has no equivalent, in the scipy-stack. How do I address this issue?

• Some notebooks use the import from shogun.Features import RealFeatures which fails to execute. Am I right in believing this needs to be changed?

I can submit a PR, addressing the above points, but I am unclear on how to approach the second query raised.

[karlnapf]

Hi!

Thanks for taking this up, really appreciated!

• modshogun -> shogun definitely! It should actually be import shogun as sg and then use sg.Bla() rather than importing everything into the namespace.
• we don't want to use pylab but rather implicit imports such as import matplotlib.pyplot as plt
• if you cannot replicate things with the standard scipy stack, make the parts that depend on non-standard packages optional, i.e. the notebook should execute smoothly, with no errors raised @vigsterkr comments?