issue_comments

I think pcolormesh is slightly more likely to yield correct and consistent results for plotting different types of data than our current hybrid of imshow/contourf. In my experience, pcolormesh also works better with cartopy. For default behavior, I would lean towards being correct rather than fast. Of course, waiting ~0.5 seconds per plot does start to add up when you're faceting, but it's not much more trouble to type .imshow.

On Wed, Oct 7, 2015 at 7:04 PM, Joe Hamman notifications@github.com wrote:

This is a good point. For me, its not a deal breaker though and the benefits of using pcolormesh all the time (simpler code, broad applicability) are worth the wait.

Reply to this email directly or view it on GitHub: https://github.com/xray/xray/pull/608#issuecomment-146393127

LGTM

I'm OK with changing the default to yincrease=True. We might consider making this a globally settable option, if, for example, we start to get lots of people familiar with the other convention (popular for images) using xray.

I'm conflicted about what the default xincrease/yincrease keyword values should be. In practice, it seems like some datasets need them, and some don't. But it also depends on the specific application.

@jhamman could you please explain your reasoning a little more for why we need this switch? Do you have a comprehensive sense of the ways that people organize their data cubes?

With climate data, it does seem pretty common that datasets have decreasing latitude values (e.g., 90, 80, 70, ...). This works well with the default behavior of imshow, but not pcolormesh (unless yincrease=True). I suppose this is your motivation?

Originally, we only had imshow and contourf. We picked imshow as the default because it shows individual pixels. But pcolormesh also meets that criteria, and has the convenience of also working on non-evenly spaced grid.

The only reason to possibly not pick pcolormesh is that it has a different convention from imshow in terms of the image orientation (top-left vs bottom-left origin). Given that xray is largely used for data with meaningful coordinate labels (rather than images), I think it's probably a better choice to use the bottom-left origin as a default. It composing better, for example, with 1D plots.

OK, works for me :)

On Wed, Oct 7, 2015 at 2:40 PM, Joe Hamman notifications@github.com wrote:

I started to go down the road of removing the squeeze you were talking about but _infer_interval_breaks got in my way (when len(coord) == 1). I think this is an important corner case but this PR is getting pretty cluttered so maybe it is best to save that for another.

— Reply to this email directly or view it on GitHub https://github.com/xray/xray/pull/608#issuecomment-146341080.

so you would have a DataArray of shape (1, 5) be plotted via pcolormesh (by default)?

Either that, or we should extract the row and col arguments first, to ensure they aren't squeezed out.

On the other hand, it's not so terrible to encourage people use to explicit plot method in applications instead of the generic .plot, so maybe this isn't entirely necessary.

While you're at it -- maybe the default .plot method shouldn't start with darray.squeeze()? That sort of behavior starts to make things behavior to predict. In particular, faceting should still work consistently even if one of the dimensions only has length 1.

How about making the default pcolormesh since that works in all cases?

:+1:

Can we make it so the generic .plot(x='lon', y='lat') works? Right now it defaults to imshow is the dimensions are equally spaced but that logic isn't probably right anymore if we allow different x/y labels: https://github.com/xray/xray/blob/cb4e4138fdb52078d27341dadf0feedb15e1de80/xray/plot/plot.py#L161

Maybe instead of the complex fallback logic to handle specifying x and/or y, we should rewrite this to only handle two cases: 1. Both x and y are provided explicitly. They can be any coordinate variables. 2. Neither x nor y is provided, so use dims to provide default values.

Something like this...

python def _infer_xy_labels(plotfunc, darray, x, y): if x is None and y is None: if darray.ndim != 2: raise ValueError('must be 2d') x, y = darray.dims elif x is None or y is None: raise ValueError('cannot supply only one of x and y') elif any(k not in darray.coords for k in (x, y)): raise ValueError('x and y must be coordinate variables') return x, y