FIX Correctly handle NaNs in VertexRGB and VolumeRGB

[mvdoc]

When creating a new VertexRGB object with vertices containing NaNs, the values of those vertices would be cast to uint8 and transformed to 0. However, the alpha value would be set to 1, causing the Nan vertices to appear black. This PR fixes this issue by automatically setting a correct alpha of 0 for NaN values. It partially addresses #455.

Minimal example:

import cortex
import numpy as np
import matplotlib.pyplot as plt

# get an roi to mask out
roi = cortex.get_roi_verts("fsaverage", roi="IPS0")["IPS0"]

# Make a VertexRGB object
c = np.linspace(0, 1, 327684)
rgb = np.array([c, np.zeros_like(c), np.zeros_like(c)])
rgb[:, roi] = np.nan
vtx = cortex.VertexRGB(*rgb, "fsaverage")

cortex.webgl.show(vtx)
cortex.quickshow(vtx, with_colorbar=False, with_curvature=True);plt.show()

Before:

After:

[TomDLT]

This is definitely better than before.

Two nitpicks:

• if self.red is updated after initialization, self.alpha is not updated.
• if self.alpha is given as an array, it ignores NaNs in the data.

We could address both issues with the slightly more complicated solution:

    def __init__(...):
        ...
        self.alpha = alpha

    @property
    def alpha(self):
        """Compute alpha transparency"""
        alpha = self._alpha
        if alpha is None:
            alpha = np.ones(self.red.vertices.shape)
            alpha = Vertex(alpha, self.red.subject, vmin=0, vmax=1)
        if not isinstance(alpha, Vertex):
            alpha = Vertex(alpha, self.red.subject)
        
        rgb = np.array([self.red.data, self.green.data, self.blue.data])
        mask = np.isnan(rgb).any(axis=0)
        alpha._data[mask] = alpha.vmin
        return alpha

    @alpha.setter
    def alpha(self, alpha):
        self._alpha = alpha

[mvdoc]

Done. I copied your suggestion as it was, with a small change:

# alpha._data[mask] = alpha.vmin
alpha._data[..., mask] = alpha.vmin

If alpha is set to None, the data of the alpha channel for a VertexRGB object is 2D. If instead a Vertex object is explicitly passed as alpha, then the data of the alpha channel is 1D.

[mvdoc]

Done. I copied your suggestion as it was, with a small change:

# alpha._data[mask] = alpha.vmin
alpha._data[..., mask] = alpha.vmin

If alpha is set to None, the data of the alpha channel for a VertexRGB object is 2D. If instead a Vertex object is explicitly passed as alpha, then the data of the alpha channel is 1D.

Nevermind. Vertex.vertices always returns a 2D array, and that's why alpha became a 2D array when initialized with None. Now it should be OK.

[TomDLT]

Thanks for adding it to VolumeRGB too. Then I think we should refactor it into DataviewRGB, to avoid code duplication. It seems that self.red.vertices and self.red.volume could be merged with self.red.data.

edit: Nevermind, creating a Volume/Vertex for alpha might be hard to merge into one code. We might as well duplicate the code.

[mvdoc]

Then I think we should refactor it into DataviewRGB, to avoid code duplication. It seems that self.red.vertices and self.red.volume could be merged with self.red.data.

Yeah, I started doing that, but then realized that DataviewRGB should be refactored heavily to make it a bit more robust. Now DataviewRGB tries to access properties that do not exist yet (e.g., .red, which only exists for the subclasses). Also, as you discovered there is not a lot of consistency between VertexRGB and VolumeRGB, so I had to duplicate the code...

I also had to modify VolumeRGB.color_voxels. It was casting nans to 0 without setting an appropriate alpha. I'll merge once I make sure that it is working...