WIP: Implement more continuous log CDF functions #2048

domenzain · 2017-04-18T11:10:46Z

Implement several log CDF functions

The distributions concerned are:

Pareto
Flat
Wald
Exponential

Each implementation is then tested against SciPy's own implementation.

The reasoning behind this is that a Flat distribution is a Uniform distribution with bounds stretching symmetrically towards negative and positive infinity. Or a Normal distribution with infinite standard deviation. Both of these give values of 1/2 for any finite point in the CDF. At negative infinity, the CDF is 0 and at positive infinity it is 1.

Since the normal CDF is used in both Wald and Normal it is refactored into a function but not exported

domenzain · 2017-04-18T12:02:59Z

This is done in the context of #1867 and #1864.
When it is complete, there should be an adequate CDF for all continuous distributions.

springcoil

Looks cool to me. Good contributions. If you could add a few tests for infinite values that'd be cool. Otherwise we can leave it for another commit.

springcoil · 2017-04-20T08:06:29Z

pymc3/tests/test_distributions.py

@@ -390,6 +390,8 @@ def test_discrete_unif(self):

    def test_flat(self):
        self.pymc3_matches_scipy(Flat, Runif, {}, lambda value: 0)
+        # TODO: Check infinite values?


I would check infinite values. Seems a good check.

I've pushed a quick explicit couple of tests.

springcoil · 2017-04-20T08:09:14Z

I guess we need to wait for Beta to get this to merge. But good work @domenzain :)

fonnesbeck · 2017-04-22T18:09:47Z

pymc3/distributions/continuous.py

+        lam = self.lam
+        alpha = self.alpha
+
+        value = value - alpha


use increment value += alpha

ferrine

Does it all work with tensors, not scalars, we need more tests for that

ferrine · 2017-04-22T19:46:12Z

pymc3/distributions/continuous.py

-            tt.log1p(-tt.erfc(z / tt.sqrt(2.)) / 2.)
-        )
+
+# TODO: Decide where to put this, or inline it everywhere it is used...


You can move it to dist_math

ferrine · 2017-04-22T19:47:27Z

pymc3/distributions/continuous.py

+            "Accurately computing log(1-exp(-|a|)) Assessed by the Rmpfr package"
+        """
+        lam = self.lam
+        a = lam * float(value)


What if value is tensor? Mb tt.as_tensor(value) fits better there

Changing this. I'm using floatX and as_tensor instead.

ferrine · 2017-04-22T19:51:18Z

pymc3/distributions/continuous.py

@@ -181,6 +182,17 @@ def random(self, point=None, size=None, repeat=None):
    def logp(self, value):
        return tt.zeros_like(value)

+    def logcdf(self, value):
+        return tt.switch(


Do you check vector value for this op?

I think all of these are element-wise operations.
When a tensor is passed as argument they would behave as expected:

>>> tt.eq(np.array([1, 2, np.nan, -np.inf, np.inf]), np.inf).eval() array([False, False, False, False, True], dtype=bool)

ferrine · 2017-04-22T19:52:43Z

pymc3/distributions/continuous.py

-        )
+
+# TODO: Decide where to put this, or inline it everywhere it is used...
+def normallogcdf(value, mu=0, sd=1):


Does it work with ndimentional tensors?

I'm unsure about broadcasting when the tensors will not be shaped identically.
But otherwise everything is element-wise ok.

…hods

domenzain

Latest review taken into account and integrated in commits where applicable.

domenzain · 2017-04-20T09:10:07Z

pymc3/distributions/continuous.py

-            tt.log1p(-tt.erfc(z / tt.sqrt(2.)) / 2.)
-        )
+
+# TODO: Decide where to put this, or inline it everywhere it is used...


I don't know where to put this. I'm guessing there must be a more logical place than here.

domenzain · 2017-04-20T09:12:59Z

pymc3/tests/test_distributions.py

@@ -390,6 +390,8 @@ def test_discrete_unif(self):

    def test_flat(self):
        self.pymc3_matches_scipy(Flat, Runif, {}, lambda value: 0)
+        # TODO: Check infinite values?


I've pushed a quick explicit couple of tests.

domenzain · 2017-04-24T09:21:19Z

pymc3/distributions/continuous.py

@@ -181,6 +182,17 @@ def random(self, point=None, size=None, repeat=None):
    def logp(self, value):
        return tt.zeros_like(value)

+    def logcdf(self, value):
+        return tt.switch(


I think all of these are element-wise operations.
When a tensor is passed as argument they would behave as expected:

>>> tt.eq(np.array([1, 2, np.nan, -np.inf, np.inf]), np.inf).eval() array([False, False, False, False, True], dtype=bool)

domenzain · 2017-04-24T09:21:39Z

pymc3/distributions/continuous.py

-            tt.log1p(-tt.erfc(z / tt.sqrt(2.)) / 2.)
-        )
+
+# TODO: Decide where to put this, or inline it everywhere it is used...


domenzain · 2017-04-24T09:22:56Z

pymc3/distributions/continuous.py

-        )
+
+# TODO: Decide where to put this, or inline it everywhere it is used...
+def normallogcdf(value, mu=0, sd=1):


I'm unsure about broadcasting when the tensors will not be shaped identically.
But otherwise everything is element-wise ok.

domenzain · 2017-04-24T09:23:30Z

pymc3/distributions/continuous.py

+            "Accurately computing log(1-exp(-|a|)) Assessed by the Rmpfr package"
+        """
+        lam = self.lam
+        a = lam * float(value)


Changing this. I'm using floatX and as_tensor instead.

domenzain · 2017-04-24T12:09:33Z

I've removed Beta from this pull request, seeing that @fonnesbeck is working on it directly on the target branch.

fonnesbeck · 2017-04-24T12:37:35Z

@domenzain I was giving it a try, but have not finished yet.

Keep in mind, everyone, that this is being merged into a feature branch, so there will be another opportunity to review it before it is merged into master.

fonnesbeck · 2017-04-24T16:47:07Z

I will merge this into my feature branch.

domenzain added 3 commits April 18, 2017 12:59

Implement log CDF for Pareto distribution

467c2f9

Implement log CDF for Wald distribution

8f341dc

Since the normal CDF is used in both Wald and Normal it is refactored into a function but not exported

domenzain changed the title ~~Implement more continuous log CDF functions~~ [WIP] Implement more continuous log CDF functions Apr 18, 2017

domenzain changed the title ~~[WIP] Implement more continuous log CDF functions~~ WIP: Implement more continuous log CDF functions Apr 18, 2017

Implement log CDF for Exponential distribution

fb229bf

springcoil reviewed Apr 20, 2017

View reviewed changes

Test for Flat distribution infinite cases individually

8dd6e79

fonnesbeck reviewed Apr 22, 2017

View reviewed changes

pymc3/distributions/continuous.py Outdated

lam = self.lam

alpha = self.alpha

value = value - alpha

Copy link

Member

fonnesbeck Apr 22, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

use increment value += alpha

Use decrement

da16cce

ferrine requested changes Apr 22, 2017

View reviewed changes

domenzain added 3 commits April 24, 2017 11:12

Put log CDF for Normal in dist_math for reuse

3dc5573

Use tensor floatX

e09b250

Merge branch 'cdf_methods' of github.com:domenzain/pymc3 into cdf_met…

d63e7f5

…hods

domenzain commented Apr 24, 2017

View reviewed changes

Merge branch 'cdf_methods' into cdf_methods

9ebaa17

fonnesbeck merged commit 2df35b3 into pymc-devs:cdf_methods Apr 24, 2017

domenzain mentioned this pull request Apr 26, 2017

Add CDF methods to continuous distributions #2073

Closed

domenzain deleted the cdf_methods branch April 26, 2017 16:52

This was referenced Oct 30, 2017

Implement incomplete beta function #2678

Closed

Implement continuous distribution CDF methods #2688

Merged

Uh oh!

WIP: Implement more continuous log CDF functions #2048

WIP: Implement more continuous log CDF functions #2048

Uh oh!

Conversation

domenzain commented Apr 18, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

domenzain commented Apr 18, 2017

Uh oh!

springcoil left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

springcoil commented Apr 20, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ferrine left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

domenzain left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

domenzain commented Apr 24, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fonnesbeck commented Apr 24, 2017

Uh oh!

fonnesbeck commented Apr 24, 2017

Uh oh!

Uh oh!

domenzain commented Apr 18, 2017 •

edited

Loading

domenzain commented Apr 24, 2017 •

edited

Loading