Changed interpolation type to make numeric range consistent with Plotly#2712
Changed interpolation type to make numeric range consistent with Plotly#2712himkt merged 1 commit intooptuna:masterfrom
Conversation
Codecov Report
@@ Coverage Diff @@
## master #2712 +/- ##
=======================================
Coverage 91.67% 91.68%
=======================================
Files 140 140
Lines 11538 11539 +1
=======================================
+ Hits 10578 10579 +1
Misses 960 960
Continue to review full report at Codecov.
|
nzw0301
left a comment
There was a problem hiding this comment.
Thank you for giving comprehensive examples. Looks great to me!
|
@c-bata Can you give your thoughts on this? |
|
@Crissman Could you unassign me? I need to work on the review of optuna-dashboard project for a while. |
|
Probably got slipped away, I just unassigned you @c-bata. And thanks for checking. |
himkt
left a comment
There was a problem hiding this comment.
Thank you for tackling the issue. Let me approve.
I confirmed the range of objective values is corrected.
(Although we may still have to sophisticate the interpolation algorithm, it is another point.)
For further development, I think there is room to improve xrange and yrange of the plot.
The data point near (y, z) = (-10, -100) appears in the plotly's output but not in matplotlib's.
| matplotlib | plotly |
|---|---|
 |
 |
Motivation
The range of objective in
optuna.visualization.matplotlib.plot_contourshould have a similar range tooptuna.visualization.plot_contour.Description of the changes
I read the provided link for interpolation of irregular space data and tried the CubicInterpolation code [interpolating z], still got the same results (incorrect numeric range). The Linear interpolation and the option
linearin thegriddata functionwork and corrects the numeric range as well, in the same way, so I went with linear interpolation.Before_Cubic_Interpolation:
After_Linear_Interpolation:
Plotly Results:
As for the colour scale, I have addressed that issue in this PR: #2711
Final Results after colour scale and interpolation correction after merging both the PRs:
closes #2595