Building a general numpy array index dynamically

Question

Basic Background

I am working on some code to simplify training SVMs for data with varying numbers of features, and visualizing the decision boundaries of these SVMs using "slices" specified by the user. If I have n features in my data set and m samples, I generate an (n+1)-dimensional meshgrid where each slice along the first index is an m x m x ... mesh of dimension n. I can then use my SVM to classify each data point in my meshgrid.

What I want to do next is plot a slice of these results in any two dimensions specified by the user. I have code that plots what I want when the data only has two features, but as soon as I add a third I start running into indexing problems.

Problem Statement

Say I have a three-dimensional matrix, predictions, and I want to plot these predictions over all of the values in my meshgrid mesh associated with index0=0 and index1=1, as well as the training data in those dimensions. I can do this with a function call like:

import matplotlib.pyplot as plt
plt.contourf(mesh[index0,:,:,0], mesh[index1,:,:,0], pred[:,:,0])
plt.scatter(samples[:,index0], samples[:,index1], c=labels)
plt.show()

What I want to know is how I can build my indexing arrays dynamically so that if index0=0 and index1=1, we get the above code, but if index0=1 and index1=2, we would get:

plt.contourf(mesh[index0,0,:,:], mesh[index1,0,:,:], pred[0,:,:])

and if index0=0 and index1=2, we would get:

plt.contourf(mesh[index0,:,0,:], mesh[index1,:,0,:], pred[:,0,:])

How can I go about building these dynamically? Is there a better way to go about this in general for cases where I may not know ahead of time how many features the data is going to have?

More Details

I attempted something like:

mesh_indices0 = [0]*len(mesh.shape)
mesh_indices0[0] = index0
mesh_indices0[index0+1] = ':'    # syntax error: I cannot add this dynamically
mesh_indices0[index1+1] = ':'    # same problem

I also tried going at it from the opposite direction with mesh_indices = [:]*len(mesh.shape), but that is invalid syntax as well. I thought about trying something like:

mesh_indices[index0+1] = np.r_[:len(samples[:, 1])]

where samples is my m x n set of observations. This seemed really clunky to me though, so I figured there had to be a better way.

perimosocordiae · Accepted Answer · 2014-06-03 20:40:32Z

1

I'm not sure that I fully understand what you're trying to do, but if you want to manipulate slices, you should use the python slice object:

mesh[index0,0,:,:]

is equivalent to:

mesh[index0,0,slice(0,mesh.shape[2]),slice(0,mesh.shape[3])]

Also note that you can index with a list or tuple of slices and indices:

inds = (index0, 0, slice(0,mesh.shape[2]), slice(0,mesh.shape[3]))
mesh[inds]

Putting it all together, you can make a list of :-equivalent slice objects, then replace the appropriate one with your concrete index. Or, go the other way:

mesh_indices = [0]*len(mesh.shape)
mesh_indices[0] = index0
mesh_indices[index0+1] = slice(0, mesh.shape[index0+1])
mesh_indices[index1+1] = slice(0, mesh.shape[index1+1])

answered Jun 3, 2014 at 20:40

perimosocordiae

17.9k14 gold badges64 silver badges76 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Engineero Over a year ago

That seems to do exactly what I want it to. Thank you so much!

Collectives™ on Stack Overflow

Building a general numpy array index dynamically

Basic Background

Problem Statement

More Details

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

Basic Background

Problem Statement

More Details

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related