Fantastic, thanks both. Hoping next time I can contribute a fix as well as an issue ticket!

Hi @shoyer , here's an attempt: https://github.com/pydata/xarray/pull/794

Thanks @fmaussion for that example.

I would tentatively call this is a bug. At the very least, I don't think there's a good reason why this shouldn't work.

I think we need a slightly more robust solution to these lines: https://github.com/pydata/xarray/blob/v0.7.1/xarray/plot/plot.py#L380-L382

For example, it might be a better idea to look at darray[xlab].dims rather than xlab itself. Any interest in putting together a PR?

Here is a mwe:

``` python import xarray as xr import numpy as np

temp = xr.DataArray(np.random.randn(3, 5), dims=['time', 'x'], coords={'x': [1,2,3,4,5], 'time': [1,2,3]}) temp.coords['rtime'] = temp.time + 3

for some reason this works

temp.plot(x='x', y='rtime');

this does not

temp.plot(x='rtime', y='x');

and this works again

temp.T.plot(x='rtime', y='x'); ```

the question (I guess) is: is it legal in xarray to use a non-linked coordinate name as argument for a plot? For example @jamesp wants to plot the following DataArray:

python <xarray.DataArray 'temperature' (time: 3, x: 5)> ... Coordinates: * x (x) int64 1 2 3 4 5 * time (time) int64 1 2 3 r int64 2 rtime (time) float64 0.5 1.0 1.5

By doing:

python .plot(x='x', y='rtime');

instead of the more conventional:

python .plot(x='x', y='time');

The dimension is still time but the coordinates of the plot (for the axis) is rtime. It works if you don't change the order of the dims, but it does not work if you do (hence the transpose in my workaround).

To be clear, by "Can you do this in pure matplotlib instead of xarray plotting?", I meant could you please demonstrate that here so I can understand exactly what you were hoping to see? Bug reports that show what you expect are more useful than those that only show what went wrong :).

I think the axes on the pdf are wrong as I used integers in the coordinate definition and Python 2 to generate the example. Redefining time as [1.0, 2,0, 3.0] gives a better plot. I should have done this in my example, sorry about that.

I can do this in matplotlib and will do so, I guess I was little surprised that (a) it worked at all, and (b) it only worked in one orientation.

I would expect the plots to look exactly like time vs. x, except the time axis is normalised by 'r'.

My hope was that this would work, so I could do something like:

temp.plot(x='rtime', y='x', col='r')

and get an array of plots, each with their own timescales along the x-axis. For my analysis, time is seconds is not really that interesting, it's how it maps to the rotation timescale that matters.

I realise this is probably not a normal use case, so thanks to both of you for taking the time to reply to my question!

On 11 Mar 2016, at 16:58, Stephan Hoyer notifications@github.com wrote:

Honestly I'm a little surprised this doesn't error: dat.temperature.sel(r=2).mean(’y’).plot(x=’x’, y=’rtime’). The plot from your PDF doesn't seem to have the right y labels (rtime ranges from 0 to 2, but the highest label on the plot is 1).

What would you expect these plots to look like? Can you do this in pure matplotlib instead of xarray plotting? I would guess the result you're looking for may be possible but it's not immediately obvious to me what plot should result from this code.

— Reply to this email directly or view it on GitHub.

Honestly I'm a little surprised this doesn't error: dat.temperature.sel(r=2).mean(’y’).plot(x=’x’, y=’rtime’). The plot from your PDF doesn't seem to have the right y labels (rtime ranges from 0 to 2, but the highest label on the plot is 1).

What would you expect these plots to look like? Can you do this in pure matplotlib instead of xarray plotting? I would guess the result you're looking for may be possible but it's not immediately obvious to me what plot should result from this code.

That works, thanks. You say that it only worked by chance, is there a "proper" xarray-way that I should be doing this?

I think that the first case works by chance, since the data dimensions are still 'time', 'x'. But this workaround should work:

python dat.temperature.sel(r=2).mean('y').T.plot(x='rtime', y='x')

(note the transpose)

yes, in some sense, but it doesn't need to infer. I'm explicitly telling the plot command I want to use 'x' and 'rtime' as the axial coordiates in both examples in my notebook.

It does the right thing when y=rtime; it gives an error when x=rtime.

I guess that what you are asking is that xarray should infer from y='x' that the other dimension must be time, and then scale with your new rtime coordinates?

Aha, I think I just got it. I deleted my previous answer, as I don't know how to answer your question :flushed:

my use case is that rtime is kind of like normalised time. rtime is implicitly linked to temperature via time and r, no?

I'm running many simulations with varying rotation period r. time is just clock time in seconds; what I want to plot is time in terms of revolutions i.e. time/r.

It works just fine with rtime along the y-axis, but I would like it along the x and there it fails as the length of the coordinate seems to have changed.

Just out of curiosity, what is the use case here? In your example the coordinate rtime has a shape (time: 3, r: 3) ant it is not linked to any dimension of the variable that you are plotting. What do you expect the plot to do?