[MRG+1] EHN add decimals parameter for export_graphviz #8698

glemaitre · 2017-04-04T14:31:16Z

Reference Issue

What does this implement/fix? Explain your changes.

Introduce a parameter decimals to control decimal precision when displaying threshold, impurity, and value parameters in export_graphviz.

Any other comments?

MechCoder · 2017-04-21T19:09:41Z

sklearn/tree/export.py

@@ -397,6 +405,14 @@ def recurse(tree, node_id, criterion, parent=None, depth=0):
            return_string = True
            out_file = six.StringIO()

+        if isinstance(decimals, Integral):
+            if decimals < 0:
+                raise ValueError("'decimal' should be greater or equal to 0."


nitpick: This should be decimal(s)

MechCoder · 2017-04-21T19:09:41Z

sklearn/tree/tests/test_export.py

+        dot_data = StringIO()
+        export_graphviz(clf, out_file=dot_data, decimals=decimals)
+        # check value
+        for finding in finditer("nvalue = \d+\.\d+", dot_data.getvalue()):


Why (n) value?

my mistake, I did not notice it was a \n

MechCoder · 2017-04-21T19:09:41Z

sklearn/tree/tests/test_export.py

+    clf = DecisionTreeRegressor(criterion="friedman_mse", random_state=0,
+                                max_depth=1)
+    clf.fit(X, y_reg)
+    for decimals, nb_len in zip((4, 3), (5, 4)):


Sorry for being dumb, but what is nb_len? Or why are you checking that there should be 3 decimals when there can be a maximum of 4.

this is the opposite. The length will be decimals + 1 since I will try to match a .. But this is true that nb_len is useless.

MechCoder · 2017-04-21T19:09:41Z

sklearn/tree/tests/test_export.py

+        export_graphviz(clf, out_file=dot_data, decimals=decimals)
+        # check value
+        for finding in finditer("nvalue = \d+\.\d+", dot_data.getvalue()):
+            assert_equal(len(search("\.\d+", finding.group()).group()), nb_len)


I'm pretty sure this should be assert_less(...., decimals)

It could be less_equal. But with this random seed, we should get exactly decimals + 1. Making it less_equal can make the test too much permissive

MechCoder · 2017-04-21T19:16:12Z

sklearn/tree/tests/test_export.py

+    y_reg = rng.random_sample((5, ))
+
+    # regression case
+    clf = DecisionTreeRegressor(criterion="friedman_mse", random_state=0,


Can you add a test for a classifer as well?

MechCoder

That's it from me.

MechCoder · 2017-04-22T23:16:40Z

sklearn/tree/tests/test_export.py

+    clf.fit(X, y_reg)
+    for decimals in (4, 3):
+        dot_data = StringIO()
+        export_graphviz(clf, out_file=dot_data, decimals=decimals)


You can use out_file=None directly since #7390

MechCoder · 2017-04-22T23:16:40Z

sklearn/tree/tests/test_export.py

+        # check impurity
+        for finding in finditer("friedman_mse = \d+\.\d+",
+                                dot_data.getvalue()):
+            assert_equal(len(search("\.\d+", finding.group()).group()),


I agree that assert_less makes this more permissive. In that case, this warrants a comment that this is particular for this random seed as you mention or you can use toy data where you know that the number of decimals in impurity etc are greater than the provided value.

MechCoder · 2017-04-22T23:16:40Z

sklearn/tree/tests/test_export.py

+    clf.fit(X, y_cla)
+    for decimals in (4, 3):
+        dot_data = StringIO()
+        export_graphviz(clf, out_file=dot_data, decimals=decimals)


Setting proportion=True over here might be a good idea.

MechCoder · 2017-04-22T23:17:24Z

LGTM pending whatsnew

glemaitre · 2017-04-23T09:13:15Z

@MechCoder Done

MechCoder

LGTM

glemaitre · 2017-04-25T20:47:43Z

@raghavrv ping Could you give a look?

raghavrv

Thanks!! A few questions. Otherwise LGTM

+1 for merge after refactoring the tests into a loop (if possible)

raghavrv · 2017-04-27T17:31:17Z

sklearn/tree/tests/test_export.py

+    y_reg = rng.random_sample((5, ))
+
+    # regression case
+    clf = DecisionTreeRegressor(criterion="friedman_mse", random_state=0,


raghavrv · 2017-04-27T17:31:17Z

sklearn/tree/export.py

@@ -72,7 +74,7 @@ def export_graphviz(decision_tree, out_file=SENTINEL, max_depth=None,
                    feature_names=None, class_names=None, label='all',
                    filled=False, leaves_parallel=False, impurity=True,
                    node_ids=False, proportion=False, rotate=False,
-                    rounded=False, special_characters=False):
+                    rounded=False, special_characters=False, decimals=3):


Should it be precision or should we leave it as decimals to conform to the assert_* syntax?

raghavrv · 2017-04-27T17:31:17Z

sklearn/tree/export.py

@@ -142,6 +144,10 @@ def export_graphviz(decision_tree, out_file=SENTINEL, max_depth=None,
        When set to ``False``, ignore special characters for PostScript
        compatibility.

+    decimals : int, optional (default=3)
+        The number of decimal reported of the ``impurity``,


Desired precision in the values of impurity, threshold and value attributes of each node.

raghavrv · 2017-04-27T17:31:17Z

sklearn/tree/tests/test_export.py

+            assert_equal(len(search("\.\d+", finding.group()).group()),
+                         decimals + 1)
+
+    # classification case


Is it possible to make a loop for reg/classification?

jmschrei · 2017-04-28T02:52:57Z

I can see a case being made for precision since that's what numpy uses in set_printoptions. I am neutral on it, as this isn't a big enough feature to warrant nit-picking. Once the checks are fixed, I am +1 as well.

glemaitre · 2017-04-28T13:28:02Z

I first thought about decimals as in around, but the point of @jmschrei and @raghavrv is pertinent.

I made the changes

jnothman · 2017-04-29T09:27:01Z

pandas used precision too

…

On 28 Apr 2017 11:28 pm, "Guillaume Lemaitre" ***@***.***> wrote: I first thought about decimals as in around, but the point of @jmschrei <https://github.com/jmschrei> and @raghavrv <https://github.com/raghavrv> is pertinent. I made the changes — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#8698 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz6zfH6wJfIUtfo63vikBLU1-MOdJtks5r0elkgaJpZM4My8fh> .

glemaitre · 2017-04-29T11:37:09Z

@jnothman ok so definitely precision makes sense. Then, it should be OK to be merged.

raghavrv · 2017-04-29T13:15:00Z

Thanks a lot... I'm merging this in!!

ryanrozich · 2017-04-29T14:51:02Z

Thanks!

…8698) * EHN add decimals parameter for export_graphviz * FIX address comments * TST add test for classification * TST/FIX address comments * FIX comments raghav

EHN add decimals parameter for export_graphviz

Loading status checks…

528f8d1

MechCoder reviewed Apr 21, 2017

View changes

glemaitre added 2 commits Apr 22, 2017

FIX address comments

Loading status checks…

d70e169

TST add test for classification

Loading status checks…

370cc09

MechCoder reviewed Apr 22, 2017

View changes

glemaitre added 2 commits Apr 23, 2017

Merge remote-tracking branch 'origin/master' into is/8662

Loading status checks…

bcdd689

TST/FIX address comments

Loading status checks…

498137d

MechCoder approved these changes Apr 23, 2017

View changes

MechCoder changed the title ~~[MRG] EHN add decimals parameter for export_graphviz~~ [MRG+1] EHN add decimals parameter for export_graphviz Apr 23, 2017

Merge branch 'master' into is/8662

Loading status checks…

1ceabc2

raghavrv suggested changes Apr 27, 2017

View changes

FIX comments raghav

Loading status checks…

063ffda

scikit-learn / scikit-learn

[MRG+1] EHN add decimals parameter for export_graphviz #8698

[MRG+1] EHN add decimals parameter for export_graphviz #8698

glemaitre commented Apr 4, 2017

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

MechCoder left a comment

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

MechCoder commented Apr 22, 2017

glemaitre commented Apr 23, 2017

MechCoder left a comment

glemaitre commented Apr 25, 2017

raghavrv left a comment

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

jmschrei commented Apr 28, 2017

glemaitre commented Apr 28, 2017

jnothman commented Apr 29, 2017

glemaitre commented Apr 29, 2017

raghavrv commented Apr 29, 2017

ryanrozich commented Apr 29, 2017

scikit-learn / scikit-learn

Sponsor scikit-learn/scikit-learn

[MRG+1] EHN add decimals parameter for export_graphviz #8698

[MRG+1] EHN add decimals parameter for export_graphviz #8698

Conversation

glemaitre commented Apr 4, 2017

Reference Issue

What does this implement/fix? Explain your changes.

Any other comments?

This comment has been minimized.

MechCoder Apr 21, 2017 Member

This comment has been minimized.

MechCoder Apr 21, 2017 • edited Member

This comment has been minimized.

glemaitre Apr 22, 2017 Author Contributor

This comment has been minimized.

MechCoder Apr 21, 2017 Member

This comment has been minimized.

glemaitre Apr 22, 2017 Author Contributor

This comment has been minimized.

MechCoder Apr 21, 2017 Member

This comment has been minimized.

glemaitre Apr 22, 2017 Author Contributor

This comment has been minimized.

MechCoder Apr 21, 2017 Member

This comment has been minimized.

raghavrv Apr 27, 2017 Member

MechCoder left a comment

This comment has been minimized.

MechCoder Apr 22, 2017 Member

This comment has been minimized.

MechCoder Apr 22, 2017 • edited Member

This comment has been minimized.

MechCoder Apr 22, 2017 Member

MechCoder commented Apr 22, 2017

glemaitre commented Apr 23, 2017

MechCoder left a comment

glemaitre commented Apr 25, 2017

raghavrv left a comment

This comment has been minimized.

raghavrv Apr 27, 2017 Member

This comment has been minimized.

raghavrv Apr 27, 2017 Member

This comment has been minimized.

raghavrv Apr 27, 2017 Member

This comment has been minimized.

raghavrv Apr 27, 2017 Member

jmschrei commented Apr 28, 2017

glemaitre commented Apr 28, 2017

jnothman commented Apr 29, 2017

glemaitre commented Apr 29, 2017

raghavrv commented Apr 29, 2017

ryanrozich commented Apr 29, 2017

MechCoder Apr 21, 2017
Member

MechCoder Apr 21, 2017 •

edited

Member

glemaitre Apr 22, 2017
Author Contributor

MechCoder Apr 21, 2017
Member

glemaitre Apr 22, 2017
Author Contributor

MechCoder Apr 21, 2017
Member

glemaitre Apr 22, 2017
Author Contributor

MechCoder Apr 21, 2017
Member

raghavrv Apr 27, 2017
Member

MechCoder Apr 22, 2017
Member

MechCoder Apr 22, 2017 •

edited

Member

MechCoder Apr 22, 2017
Member

raghavrv Apr 27, 2017
Member

raghavrv Apr 27, 2017
Member

raghavrv Apr 27, 2017
Member

raghavrv Apr 27, 2017
Member