Transforming data and broadcasting

[vascotenner]

Quite often, I want to substract a mean background from an Image or a Curve. This mean background is often calculated by reducing an image twice. Is it possible to substract a single number from an Image?

In numpy this would work like:
image = img_raw - bg.mean()

Or more complicated:

image.shape  = (10,20,300,300)
dark.shape  = (1,20,300,300)
image -= dark.mean(axis=-1)[...,np.newaxis]

I tried to accomplish this first in 1D (a Curve minus a number), but this is also not very easy.

Is this covered in the upcomming first part of the transformation tutorial?

[philippjfr]

I'd recommend just doing the following:

img = hv.Image(np.random.rand(10,10))
img.clone(img.data - img.data.mean())

or on a HoloMap:

imgmap = hv.HoloMap({i: hv.Image(np.random.rand(10,10)) for i in range(3)})
imgmap.map(lambda x: x.clone(x.data - x.data.mean()), [hv.Image])

for Curves this is more difficult because of the way the data is stored (we're hoping to improve that in future), I can imagine something like this being supported:

curve = hv.Curve(np.random.rand(10))
curve.clone({'y': curve.data['y'] - curve.data['y'].mean()})

[jlstevens]

Perhaps there is a new way to do it using the Data API? Otherwise, if we feel this is resolved or don't have any new suggestions, we should close this issue.

[philippjfr]

Hmm, I've considered whether Columns should support a setitem so you could do something like:

curve = hv.Curve(np.random.rand(10))
curve['y'] = curve['y'] - curve['y'].mean()
# Or even
curve['y'] -= curve['y'].mean()

We also still need to add the columns method which returns the Columns dictionary, that would at least allowRight now you can do:

curve = hv.Curve(np.random.rand(10))
columns = curve.columns()
columns['y'] = columns['y'] - columns['y'].mean()
curve.clone(columns)

[philippjfr]

Dataset.transform was merged in #4080 which allows specifying transforms. In future we may also add Dataset.assign.