Ternary contour plot #1413

emmanuelle · 2019-01-27T22:16:15Z

This is a work in progress version of contour plots in ternary diagrams. I'm opening the PR even if it's too early to merge or do a thorough feedback, but I'd appreciate some quick feedback to know whether this is a good direction to go.

TODO
[x] Docstrings for helper functions.
[x] Improve coding style of helper functions in places.
[x] Decide whether other arguments (kw) need to be exposed in create_ternaryplot
[x] Write more tests, in particular when errors are raised.

Other questions : where in the doc can I write a tutorial about this new function?

emmanuelle · 2019-01-27T22:17:06Z

@empet @jackparmer

jackparmer · 2019-01-27T22:23:14Z

Other questions : where in the doc can I write a tutorial about this new function?

https://github.com/plotly/documentation/blob/source-design-merge/_posts/python/README.md

@Michael Babyn or @priyatharsan (Slack handles) can help you with contributing questions here. The pages that we want to replace/revamp are:

https://plot.ly/python/ternary-contour/
https://plot.ly/python/ternary-scatter-contour/

jackparmer · 2019-01-27T22:26:43Z

@nicolaskruchten can we please include some or all of these example datasets in px:
https://github.com/nicholasehamilton/ggtern/tree/master/data
@empet used some of these recently in her ternary plots: https://plot.ly/~empet#/
Wonder if there are other "classic" datasets for oil & gas and financial services that we should be including in px.

jonmmease · 2019-01-28T11:29:52Z

plotly/figure_factory/_ternarycontour.py

+
+    # Annotations for ticklabels on side 0
+    annotations.extend([dict(showarrow=False,
+                             text=f'{ticklabel[j]}',


Unfortunately we can't use f-strings in released code yet because we still support Python 2.7 and 3.5. (Note that we do use them in the code generation logic, but this isn't part of the released bundle).

jonmmease · 2019-01-28T11:31:05Z

This looks really cool @emmanuelle! I'll look through the code in more detail soon but the overall concept and structure look fine, so feel free to keep on working through your TODO list.

ALso, as you make more examples, please add them as comments here in the PR so that folks can get a feel for how it works. Thanks!

emmanuelle · 2019-01-28T22:01:41Z

Thanks for the info and feedback @jackparmer and @jonmmease . I plan to finish working on my todo list tomorrow morning (Montréal time).

emmanuelle · 2019-01-29T14:53:46Z

This PR is ready for review! @jonmmease

jonmmease · 2019-01-29T17:12:35Z

Great, thanks @emmanuelle. I'll do a thorough review this evening (EST), and then hopefully we can get this merge in tomorrow for 3.6.0.

emmanuelle · 2019-01-29T21:45:41Z

Awesome, thanks @jonmmease !

jonmmease · 2019-01-29T21:03:58Z

plotly/figure_factory/_ternarycontour.py

+                                   coloring=coloring,
+                                   smoothing=smoothing)
+    side_trace, tick_trace = _styling_traces_ternary(x_ticks, y_ticks)
+    fig = go.FigureWidget(data=[contour_trace,  tick_trace, side_trace],


Currently all of the other figure factories return a go.Figure instance rather than a FigureWidget. For consistency this should return a Figure instance by default.

My eventual goal is for all of the figure factories to support an output_type argument that can be set to 'dict', 'Figure', or 'FigureWidget' to control how the figure is returned (See #1079). For this PR feel free to just always return a Figure, or to add the output_type argument that defaults to Figure.

jonmmease · 2019-01-30T00:30:33Z

Hi @emmanuelle, I had one comment above about the return type, but otherwise your implementation looks great!

I would like to take a step back, however, and discuss this general approach a bit. Here's my summary of how this works now (please correct me if I misstate something)

The input grid of 3D Barycentric (ternary) coordinates is transformed into a mesh of 2D cartesian coordinates
This mesh is interpolated/upsampled into a 2d Cartesian grid.
The interpolated grid points that fall outside of the ternary triangle are removed
This remaining grid is used to create a standard Cartesian plotly contour trace type.
The ternary axes and ticks and labels are drawn with lines and annotations.

Here are a few hangups I have with this approach:

Aesthetically, it leads to this jagged fill behavior when you zoom in.
Plotly.js supports a nice ternary axis system that support ternary specific pan/zoom (https://plot.ly/python/ternary-plots/), so its a shame to not be able to take advantage of that.
This approach won't be very composable if users want to layer scatter points or lines on top of the contours.

I'm not sure that it's practical, but I'd like to at least discuss an alternative approach. Could we do something like:

(as above)
(as above)
(as above)
Rather than construct a contour trace type, use scikit-image to find the contour lines (http://scikit-image.org/docs/dev/auto_examples/edges/plot_contours.html)
Transform the contour lines back into Barycentric coordinates
For each contour line, create a scatterternary trace with mode: 'lines' (https://plot.ly/python/ternary-scatter-contour/#create-ternary-contour-plot).

I think this would work fine for contours as lines, and I think we could probably even get the filled contour appearance by setting the fill to some combination of 'toself' and 'tonext'.

With this approach, it would be possible for a user to use the figure factory to build the initial contour plot, but then they could then add their own scatterternary traces on top of it.

What do you think @emmanuelle @jackparmer ?

One additional thought is the question of whether we also want this figure factory to support computing the kernel density directly from a collection of input points in Barycentric coordinates. Something analagous to the histogram2dcontour trace type (https://plot.ly/python/2d-histogram-contour/)

emmanuelle · 2019-01-30T12:59:04Z

Great suggestion, thank you @jonmmease. I'll definitely give it a try, it will have to wait tomorrow though... If you need this figure factory for the release and you're satisfied with the API, one solution is to merge this implementation and try another implementation later on.

Regarding your other points, the input coordinates do not have to lie on a grid, they can be any set of points in barycentric coordinates (but the interpolation will be inside their convex hull only).

jonmmease · 2019-01-30T14:04:51Z

Hi @emmanuelle, yeah I know this suggestion is a lot of change. Sorry I didn't process the approach enough when you first posted the WIP PR.

I'm definitely open to merging this as is and revisiting it later. If my suggested approach works out, I think we should be able to keep the same figure factory API.

@jackparmer what would you like to do? I think the near-term options are:

merge as-is and release 3.6.0 in the next day or two, then choose whether or not to investigate a new approach after the release.
Release 3.6.0 without this and then focus on a new approach right after the release.
Hold 3.6.0 while we investigate a new approach.

jackparmer · 2019-01-30T14:13:20Z

I think that merging (and releasing 3.6.0) as-is makes sense, then pursuing the approach that you outlined that does not use the cartesian plotly.js system (only barycentric). As you pointed out, it would be awesome if we could simulate smooth contours with many toself or tonext fills.

@empet , in the meantime, would you be interested in helping to see if Jon's approach (or something similar) is viable for ternary contours (see above)?

Rather than construct a contour trace type, use scikit-image to find the contour lines (http://scikit-image.org/docs/dev/auto_examples/edges/plot_contours.html)
Transform the contour lines back into Barycentric coordinates
For each contour line, create a scatterternary trace with mode: 'lines' (https://plot.ly/python/ternary-scatter-contour/#create-ternary-contour-plot).

nicolaskruchten · 2019-01-30T14:15:05Z

Yeah it would be best if we could use native ternary axes!

…

On Wed, Jan 30, 2019 at 07:59 Emmanuelle Gouillart ***@***.***> wrote: Great suggestion, thank you @jonmmease <https://github.com/jonmmease>. I'll definitely give it a try, it will have to wait tomorrow though... If you need this figure factory for the release and you're satisfied with the API, one solution is to merge this implementation and try another implementation later on. Regarding your other points, the input coordinates do not have to lie on a grid, they can be any set of points in barycentric coordinates (but the interpolation will be inside their convex hull only). — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1413 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAMbA04cW5oguNdtp5_7iTLl4Q4uFUyLks5vIZcYgaJpZM4aU-J_> .

empet · 2019-01-30T14:23:29Z

My comments on @jonmmease raised questions:

if Plotly go.Contour allowed x and y as 2d arrays, with their values not necessarily ordered , then we could generate a ternary contour without jagged sides, via an ilr-transformation, typical for compositional data.
To use the Plotly ternary axes provided for go.Scatterternary, we should know the cartesian coordinates of the reference triangle vertices.
As long as Plotly provides contour lines in cartesian coordinates only via go.Contour, why should we call a function from scikit-image for ternary contour lines? Obviously we can use plt.contour, too

empet · 2019-01-30T22:41:11Z

@jonmmease, @jackparmer
The function skimage.measure.find_contours is of no use for ternary contour lines.
First drawback: it works only for a single level, not a list of levels http://scikit-image.org/docs/dev/api/skimage.measure.html#skimage.measure.find_contours.

Second: We have 3 meshgrids x, y, z to get the contour lines, but skimage function uses only z-values.
So it returns the coordinates of contour-lines with respect to a cartesian reference system unrelated to the system of our x, y.

In the example at the link provided by @jonmmease, http://scikit-image.org/docs/dev/auto_examples/edges/plot_contours.html
the x coordinate in meshgrid runs in the interval [-pi, pi], i.e. [-3.14, 3.14], and the contours[0] has x-coordinates in the interval [0, 9.939027340869417].

emmanuelle · 2019-01-31T12:00:05Z

@empet we can still use the skimage function by iterating over contour values, and rescaling contour values by the grid length. I'm giving it a try at the moment, I think it's doable. More soon!

emmanuelle · 2019-01-31T12:01:17Z

@jonmmease please tell me if I should still merge this PR, another one with the more native ternary scatter should be ready this afternoon.

empet · 2019-01-31T12:06:11Z

@emmanuelle I did it here https://plot.ly/~empet/15111, applying an isometric-log-ratio transformation to ternary points (that maps data isometrically on 2d-space), extracting contour lines via plt.contour() , that are mapped back to points on ternary contour lines. plt.contour() accepts x, y, z as arguments and a list of levels. No rescaling is needed.

emmanuelle · 2019-01-31T12:08:49Z

@empet awesome! How can I see the code generating this plot? My problem with plt.contour was that in a notebook it generated a matplotlib figure with the contour plot.

empet · 2019-01-31T12:35:07Z

@emmanuelle I'll upload the notebook on Plotly cloud in a few minutes and post here the link. Now I'm cleaning the code.

empet · 2019-01-31T12:55:08Z

@emmanuelle This is the link to the J. notebook https://plot.ly/~empet/15113

I learned about ilr - transformation from this paper: http://ima.udg.es/~jamf/Updating%20on%20the%20KDE%20for%20Compositional%20Data_MChM_definitive.pdf
Using the method presented here I estimated the density of some ternary data https://plot.ly/~empet/15092

empet · 2019-01-31T13:04:37Z

@emmanuelle The above method to plot the contour lines works only for proportions at the moment.

jonmmease · 2019-01-31T13:35:37Z

@jonmmease please tell me if I should still merge this PR, another one with the more native ternary scatter should be ready this afternoon.

@emmanuelle I was thinking of merging today, but I'm fine giving it another day or two since it sounds like you and @empet are making good progress! Why don't you open a new PR if you come up with an alternative approach, that way this one is still here as-is in case we decide to merge this now and update to a new approach later.

Thanks!

empet · 2019-01-31T17:16:33Z

@emmanuelle I updated the notebook with a short presentation of the ilr-transform method.

emmanuelle · 2019-02-01T07:35:42Z

Thank you very much @empet !

so that we don't have to support them initially when we convert this figure factory to represent contours using scatterternary traces.

jonmmease · 2019-02-01T11:15:12Z

@emmanuelle, I wanted to make two more changes here before merging for 3.6.0

I renamed create_ternarycontour as create_ternary_contour to be a bit more consistent with the other figure factories.
I removed the reversescale, smoothing, and showlabels options so that we don't need to support these in Updated method for ternary contour figure factory #1418 before replacing this implementation with the scatterternary implementation.

Does this look alright to you?

emmanuelle · 2019-02-01T11:21:43Z

@jonmmease yes it will be difficult to support showlabels, good idea to remove it. reversescale is easy to implement but I don't think it's a killer feature. OK for me!

jonmmease · 2019-02-01T11:27:34Z

reversescale is easy to implement but I don't think it's a killer feature

Cool, definitely feel free to bring it back in #1418 🙂

emmanuelle added 2 commits January 26, 2019 23:00

poc of ternarycontour

c6744a0

ternary contour figure factory

82b600d

jackparmer requested a review from jonmmease January 27, 2019 22:23

Changed string formatting to be compatible with Python <3.6

0c02beb

jonmmease reviewed Jan 28, 2019

View reviewed changes

emmanuelle added 4 commits January 28, 2019 17:35

Changed string format to be compatible with Python < 3.6

358b376

Improved documentation

4245423

corrected bug, typo in kw argument

bcd35e8

fixed encoding error

a32f58f

added tests for ternary contour plot

3d8f9bf

emmanuelle changed the title ~~[WIP] Ternary plot~~ Ternary plot Jan 29, 2019

jonmmease added this to the v3.6.0 milestone Jan 29, 2019

updated docstring

581b93c

jonmmease suggested changes Jan 29, 2019

View reviewed changes

FigureWidget -> Figure for consistency with other figure factories

d80e070

jonmmease approved these changes Jan 31, 2019

View reviewed changes

emmanuelle mentioned this pull request Jan 31, 2019

Updated method for ternary contour figure factory #1418

Merged

jonmmease added 2 commits February 1, 2019 06:03

ternarycontour -> ternary_contour

9f6aee1

Remove the reversescale, smoothing, and showlabels options

bc26b99

so that we don't have to support them initially when we convert this figure factory to represent contours using scatterternary traces.

jonmmease merged commit 3908788 into plotly:master Feb 1, 2019

jonmmease changed the title ~~Ternary plot~~ Ternary contour plot Feb 1, 2019

nicolaskruchten mentioned this pull request Feb 1, 2019

Ternary density contours plotly/plotly.js#3503

Closed

benjaminsabatini mentioned this pull request Jul 22, 2019

ternary_contour plotly/plotly.R#1583

Open

archmoj mentioned this pull request Jul 9, 2020

Polar contour plot / Contourpolar request plotly/plotly.js#4992

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Ternary contour plot #1413

Ternary contour plot #1413

emmanuelle commented Jan 27, 2019 •

edited

Loading

emmanuelle commented Jan 27, 2019

jackparmer commented Jan 27, 2019

jackparmer commented Jan 27, 2019 •

edited

Loading

jonmmease Jan 28, 2019

jonmmease commented Jan 28, 2019

emmanuelle commented Jan 28, 2019

emmanuelle commented Jan 29, 2019

jonmmease commented Jan 29, 2019

emmanuelle commented Jan 29, 2019

jonmmease Jan 29, 2019

jonmmease commented Jan 30, 2019

emmanuelle commented Jan 30, 2019

jonmmease commented Jan 30, 2019

jackparmer commented Jan 30, 2019 •

edited

Loading

nicolaskruchten commented Jan 30, 2019 via email

empet commented Jan 30, 2019

empet commented Jan 30, 2019

emmanuelle commented Jan 31, 2019

emmanuelle commented Jan 31, 2019

empet commented Jan 31, 2019

emmanuelle commented Jan 31, 2019

empet commented Jan 31, 2019

empet commented Jan 31, 2019

empet commented Jan 31, 2019

jonmmease commented Jan 31, 2019

empet commented Jan 31, 2019

emmanuelle commented Feb 1, 2019

jonmmease commented Feb 1, 2019

emmanuelle commented Feb 1, 2019

jonmmease commented Feb 1, 2019

Ternary contour plot #1413

Ternary contour plot #1413

Conversation

emmanuelle commented Jan 27, 2019 • edited Loading

emmanuelle commented Jan 27, 2019

jackparmer commented Jan 27, 2019

jackparmer commented Jan 27, 2019 • edited Loading

jonmmease Jan 28, 2019

Choose a reason for hiding this comment

jonmmease commented Jan 28, 2019

emmanuelle commented Jan 28, 2019

emmanuelle commented Jan 29, 2019

jonmmease commented Jan 29, 2019

emmanuelle commented Jan 29, 2019

jonmmease Jan 29, 2019

Choose a reason for hiding this comment

jonmmease commented Jan 30, 2019

emmanuelle commented Jan 30, 2019

jonmmease commented Jan 30, 2019

jackparmer commented Jan 30, 2019 • edited Loading

nicolaskruchten commented Jan 30, 2019 via email

empet commented Jan 30, 2019

empet commented Jan 30, 2019

emmanuelle commented Jan 31, 2019

emmanuelle commented Jan 31, 2019

empet commented Jan 31, 2019

emmanuelle commented Jan 31, 2019

empet commented Jan 31, 2019

empet commented Jan 31, 2019

empet commented Jan 31, 2019

jonmmease commented Jan 31, 2019

empet commented Jan 31, 2019

emmanuelle commented Feb 1, 2019

jonmmease commented Feb 1, 2019

emmanuelle commented Feb 1, 2019

jonmmease commented Feb 1, 2019

emmanuelle commented Jan 27, 2019 •

edited

Loading

jackparmer commented Jan 27, 2019 •

edited

Loading

jackparmer commented Jan 30, 2019 •

edited

Loading