[MRG] Handle warnings #222
Conversation
|
Is there a way to see the output of the notebooks from the circle CI run? Reproducing the matplotlib warning locally will be tricky as you only see if the very first time you import matplotlib. |
Current coverage is 81.83% (diff: 100%)@@ master #222 diff @@
==========================================
Files 18 18
Lines 892 892
Methods 0 0
Messages 0 0
Branches 0 0
==========================================
Hits 730 730
Misses 162 162
Partials 0 0
|
|
Voila: https://circleci.com/gh/scikit-optimize/scikit-optimize/100#artifacts/containers/0 how cool is that? If we made the links relative you could navigate the whole website for the build. Tres cool. Unfortunately it seems to suggest that the warnings are still there. 😕 |
| @@ -43,6 +43,10 @@ export SKOPT_HOME=$(pwd) | |||
| conda install --yes jupyter | |||
| pip install pdoc==0.3.2 pygments | |||
|
|
|||
| # importing matplotlib once builds the font caches. This avoids | |||
| # having warnings in our example notebooks | |||
| python -c "import matplotlib.pyplot as plt" | |||
Remove deprecation warnings by changing our scoring parameter names. Try and handle the warning from matplotlib about building the font cache.
Importing matplotlib for the first time creates warnings about the font cache. To avoid those appearing in our example notebooks we import it before running the notebooks
|
Hmm, look at the last plot. What happened? bad initial random state? |
|
Good question. I can reproduce it locally. # Plot GP(x) + concours
gp = res.models[12]
y_pred, sigma = gp.predict(x, return_std=True)gives a flat prediction, but but already looks a bit weird. Something strange about the kernel we use? |
|
From looking at commits related to the kernel choice the only idea I have is #209 ? |
|
Uggh. not sure what I broke :/ |
|
I figured out the bug. Thinking of an efficient way to fix it. |
|
The fix is more subtle, decribed as TODO item no 2 in #225. (and may or may not have a huge impact) However, the bottom line is that you need more samples such that the noise prediction becomes stabilized. A better function approximation is obtained after 50 iterations (or even before around 25). Here is a plot of the attribute |
|
I will try and make a summary to see if I understand what is happening: the kernel has a parameter that represents the amount of noise (the I don't quite understand why it causes the prediction (mean of the GP) to become flat in the example though. |
|
Another question: should we merge this with the justification that this was already a problem before this PR and is being addressed in #225? |
|
Yes we can merge this now |
|
@betatim Yes, that is right. My assumption is that the prediction is wrong because keeping the noise term (or the WhiteKernel) at prediction time gives an estimate of |
Fixes #221
Remove deprecation warnings by changing our scoring parameter
names.
Try and handle the warning from matplotlib about building the
font cache.