Improve plotting for learners

[basnijholt]

(original issue on GitLab)

opened by Adriaan (@AdriaanRol) at 2018-05-10T19:07:21.297Z

Currently all plots (and live plots) are done using holoviews. There are some points where this could be improved.

• The 2D learner (and possibly the 1D leaner too) works with functions that return a vector of values. However when calling the plot method it will only plot (and live plot) the first element of this array even though all are present in learner.data
• There is no colorbar to quantify what z-value that was extracted.
• There are no sensible labels and units added to the axes, it may be possible to optionally specify this to the learners.

Below is an image of a leaner that

• uses a function that returns a vector
• should have labels and units on the x-axis
• has no colorbar

[basnijholt]

originally posted by Bas Nijholt (@basnijholt) at 2018-05-11T12:25:33.706Z on GitLab

You can easily add a colorbar by using the magics that HoloViews provides:

%%opts Image [colorbar=True]
learner.plot().redim(x='your_blabla_axis', y='this is the new y')

This answers two of your points.

• However, uses a function that returns a vector

How would you like to plot a vector?

[basnijholt]

originally posted by Adriaan (@AdriaanRol) at 2018-05-11T12:44:52.027Z on GitLab

%%opts Image [colorbar=True]

I tried that command ( googled around holoviews myself) but it messes up the aspect ratio of the image. In the example dataset I show, the y-axis goes from 0 to .5 while the x-axis (which is in degrees) goes from 10 to 100. As there is an order of magnitude difference between the quantities (and even more if I put values in 1e9 Hz on the x-axis). This can be solved in the matplotlib using the aspect keyword (only found that out today), however I think it makes sense to default to the auto aspect for adapative as well. The reason is that in the interpolation, everything is rescaled to threat both the x and y range on equal footing, it makes sense to have the default plotting behave in the same way.

learner.plot().redim(x='your_blabla_axis', y='this is the new y')

Thanks, I was not familiar with the redim method. Would be cool if it was also unit aware (e.g. converting Hz to GHz/MHz when appropriate) but I understand that is on the holoviews roadmap.
However rather than adding this each time I specify a plot method, I think it makes sense to pass this to the learner or make the learner extract this from the passed function/object in some way. As we are not redimensioning the data in the holomap but it is a unit/label that was always there.

How would you like to plot a vector?

I would like to plot them as independent quantities.
For our experiments (and also some simulations I am doing now), it is common to extract multiple quantities of interest for each datapoint that is measured/simulated. As such I would like to plot them as independent quantities e.g. hm = hm_0 + hm_1.

Let me know what you think.

[basnijholt]

originally posted by Bas Nijholt (@basnijholt) at 2018-05-11T15:57:16.902Z on GitLab

I tried that command ( googled around holoviews myself) but it messes up the aspect ratio of the image. In the example dataset I show, the y-axis goes from 0 to .5 while the x-axis (which is in degrees) goes from 10 to 100. As there is an order of magnitude difference between the quantities (and even more if I put values in 1e9 Hz on the x-axis). This can be solved in the matplotlib using the aspect keyword (only found that out today), however I think it makes sense to default to the auto aspect for adapative as well. The reason is that in the interpolation, everything is rescaled to threat both the x and y range on equal footing, it makes sense to have the default plotting behave in the same way.

I actually think the aspect ratios problem that you are having is a HoloViews bug, I've seen it before. I will report it to the developers.

I would like to plot them as independent quantities.
For our experiments (and also some simulations I am doing now), it is common to extract multiple quantities of interest for each datapoint that is measured/simulated. As such I would like to plot them as independent quantities e.g. hm = hm_0 + hm_1.

I could extend the plotting function to return a holoviews.HoloMap (which returns the plot with a slider), then you could simply call the object like learner.plot().layout() and it will all show up in subplots.

[basnijholt]

originally posted by Joseph Weston (@jbweston) at 2018-07-09T14:43:42.207Z on GitLab

I think it makes sense to pass this to the learner or make the learner extract this from the passed function/object in some way.

IMO the only solution would be to have these as parameters to plot. I would like to have the learners concentrate solely on learning, rather than carrying around a bunch of metadata too.

[basnijholt]

originally posted by Joseph Weston (@jbweston) at 2018-07-09T14:47:17.938Z on GitLab

How would you like to plot a vector?

I would like to plot them as independent quantities.
For our experiments (and also some simulations I am doing now), it is common to extract multiple quantities of interest for each datapoint that is measured/simulated. As such I would like to plot them as independent quantities e.g. hm = hm_0 + hm_1.

I could equally well imagine that one could want to plot a vector field. IMO there is probably not a single sensible default.

We could consider providing alternative plotting functions that work with the various learners.

[basnijholt]

originally posted by Bas Nijholt (@basnijholt) at 2018-07-18T17:39:39.447Z on GitLab

Can this issue be closed now that gitlab:!62 is merged?

The other questions have been answered.