cast Distribution Inputs to floatX instead of tensor_variable #2233

hvasbath · 2017-05-28T12:26:55Z

Removes slowdown again:
#1824
Running the test from that thread:
[v8, v7, v6, v5, v4, v3, v2, v1]
('time for initialising population', 1.4203970432281494) previously 20s +

Also fixes AttributeError in multivariate.py line 250.

Was not sure about the Mixture Distribution as there are a lot of theano and numpy operators mixed in the random method generation. Someone having insight should fix that one ...

junpenglao · 2017-05-28T12:43:14Z

Could you test if it pass test in https://github.com/pymc-devs/pymc3/blob/master/pymc3/tests/test_distributions.py#L827-L838 locally? The _repr_latex is using the Name from Theano tensor.

hvasbath · 2017-05-28T13:21:47Z

Yep that one fails! How to deal with that? You could simply use the Distribution name and attribute names?

junpenglao · 2017-05-28T13:49:41Z

I think we should change the function https://github.com/pymc-devs/pymc3/blob/master/pymc3/util.py#L79

junpenglao · 2017-05-28T14:14:45Z

This might work:

def get_variable_name(variable):
     """Returns the variable data type if it is a constant, otherwise
     returns the argument name.
     """
     try:
        name = variable.name
     except AttributeError:
        name = None
     if name is None:
        if hasattr(variable, 'get_parents'):
            try:
                names = [get_variable_name(item) for item in variable.get_parents()[0].inputs]
                return 'f(%s)' % ','.join([n for n in names if isinstance(n, str)]) 
            except IndexError:
                pass
        try:
            value = variable.eval()
        except AttributeError:
            value = variable
        if not value.shape:
            return asscalar(value)
        return 'array'
    return name

for some reason I cannot check out your branch. Also any idea why Travis is not building @twiecki ?

fonnesbeck · 2017-05-28T18:51:33Z

Yes, go for it.

ColCarroll

can't figure out why it isn't building -- can you either make some of the suggested changes and push, or just git commit --allow-empty and then push?

ColCarroll · 2017-05-28T18:46:06Z

pymc3/distributions/multivariate.py

@@ -247,7 +247,7 @@ def _repr_latex_(self, name=None, dist=None):
        mu = dist.mu
        try:
            cov = dist.cov
-        except AttributeErrir:


ColCarroll · 2017-05-28T19:03:15Z

pymc3/distributions/multivariate.py

-        self.k = tt.as_tensor_variable(shape)
-        self.a = a = tt.as_tensor_variable(a)
+        self.k = floatX(shape)
+        self.a = a = floatX(a)


total nit, and not necessary, but might be nicer to just use self.a below, and stop referring to a (same with Sigma and nu.

hvasbath · 2017-05-29T08:18:54Z

The test failed, because you need scipy 0.18 otherwise there is no skewnorm. The requirements have to be updated!
Also I get tons of PEP8 errors too long lines... for now I ignored all of them, as you dont seem to care either?

hvasbath · 2017-05-29T10:58:02Z

Travis still not starting ...?
Locally test_distributions passes now. However in MvStundentT there is strange behavior. There n should be an int by nature, but the test fails if it is not a float. l. 651. But I have not enough insight in the test and no time to investigate that quite complex structure. So I cast it to float and edit the latex test ...
Also I introduce an intX function. If we use float32 to run on the GPU all Ints should be cast to int16, because interaction of int32 and above with float32 results in float64, what we dont want...

junpenglao · 2017-05-29T11:41:55Z

https://travis-ci.org/pymc-devs/pymc3/requests

abuse detected: known offender (request looked fishy)

What have you done @hvasbath 😂!

hvasbath · 2017-05-29T11:49:54Z

WTH???? I am sitting in Saudi Arabia-maybe my IP is suspicious ... :)
K thanks @junpenglao I did that lets see...

junpenglao · 2017-05-29T11:51:15Z

You are on the naughty list 😂!
You should email [email protected] travis-ci/travis-ci#6942

hvasbath · 2017-05-29T16:09:39Z

Had to connect git with travis to resolve that- apparently this was a false positive of their malware detection ...

hvasbath · 2017-05-30T08:12:59Z

OMG all these failed tests! Where do all these float64 casts come from? I have no time right now to look into all of these tests and whats going wrong, some of them as well assume the variables to be tensors ... not sure what are your oppinions?

junpenglao · 2017-05-30T08:24:26Z

I dont think it's that bad, you just need to recast some of the parameters into tensor.

hvasbath · 2017-05-30T08:38:11Z

But this is the point in the first place! We do not want to do that in order to have the random method perform fast ...

twiecki · 2017-05-30T08:44:33Z

That is surprising, @hvasbath I would imagine that the a single issue underlies most of the failing tests, so perhaps just look into one and I would hope that it would resolve the other ones.

junpenglao · 2017-05-30T08:46:52Z

But this is the point in the first place! We do not want to do that in order to have the random method perform fast ...

I know, as long as they are not tensor in the random method right?
I fixed some test, still failing in

pymc3/tests/sampler_fixtures.py:130: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pymc3/tests/sampler_fixtures.py:155: in make_step
    mu, stds, elbo = pm.advi(n=50000)
pymc3/variational/advi.py:164: in advi
    uw_i, elbo_current = f()
/usr/local/lib/python3.5/dist-packages/theano/compile/function_module.py:898: in __call__
    storage_map=getattr(self.fn, 'storage_map', None))
/usr/local/lib/python3.5/dist-packages/theano/gof/link.py:325: in raise_with_op
    reraise(exc_type, exc_value, exc_trace)
../../../.local/lib/python3.5/site-packages/six.py:685: in reraise
    raise value.with_traceback(tb)

…nto random_slowdown Conflicts: pymc3/distributions/discrete.py

hvasbath · 2017-06-11T12:15:44Z

Rebased the PR, fixed tests in variational_inference.
Something like float32 -1 will result in float64! Please keep that in mind while coding ...I had to fix a lot of these things in the tests. I wonder why the tests didnt fail before? I guess it works with tensors as theano can detect this and autocasts?
Any comments? Maybe I got something totally wrong here...

hvasbath · 2017-06-12T06:26:43Z

This likely only needs the new fix from @junpenglao included, then also these last tests should pass. However, I cannot rebase again without getting hundreds of conflicts. Can anybody say why?
I tried in this branch
git rebase master
When I did it yesterday it gave only few conflicts that were resolvable by hand. However, now at every line that I changed here, it reports a conflict ...

hvasbath · 2017-06-12T06:36:39Z

Again what learned :) .. Merge did the trick and all the mess went away.

junpenglao · 2017-06-12T06:40:25Z

pymc3/distributions/continuous.py

-        self.alpha = alpha = tt.as_tensor_variable(alpha)
-        self.m = m = tt.as_tensor_variable(m)
+        self.alpha = floatX(alpha)
+        self.m = floatX(m)

        self.mean = tt.switch(tt.gt(alpha, 1), alpha *


shouldnt here be self.alpha instead of just alpha

Yep missed that!

junpenglao · 2017-06-12T06:43:22Z

pymc3/distributions/discrete.py

@@ -438,17 +441,17 @@ class DiscreteUniform(Discrete):

    def __init__(self, lower, upper, *args, **kwargs):
        super(DiscreteUniform, self).__init__(*args, **kwargs)
-        self.lower = tt.floor(lower).astype('int32')
-        self.upper = tt.floor(upper).astype('int32')
+        self.lower = intX(np.floor(lower))


Will it work if lower being a tensor? Is intX(lower) sufficient?

Having theano tensor as distribution bound inputs? Is that intended? I should have known that in advance! If yes we can forget about this whole PR ...

Yeah theano tensor could be distribution bound.

In this case can you not do something similar to Uniform(Continuous):

self.lower = floatX(lower) self.upper = floatX(upper)

hvasbath · 2017-06-12T07:05:35Z

Oh no due to the rebase and then the merge there are duplicated commits now! Argh ... How to fix that?

twiecki · 2017-06-12T07:24:22Z

@hvasbath Looks OK to me. We will do a squash-merge in any case so it doesn't really matter.

hvasbath · 2017-06-12T07:59:25Z

ok then I will close the other PR again.

hvasbath · 2017-06-12T10:36:37Z

Continued in the other cleaned up PR!

cast Distribution Inputs to floatX instead of tensor_variable

71eb51c

ColCarroll reviewed May 28, 2017

View reviewed changes

remove double assignments, intX casting, get name function

57bd152

junpenglao mentioned this pull request May 29, 2017

Convert zero-inflated distributions to Mixture subclasses #2231

Closed

travis update

363a4ad

updated username locally

0c7cdb9

fix failed test

4d7b9de

junpenglao mentioned this pull request May 31, 2017

Fix issue #2247 #2250

Merged

hvasbath added 5 commits June 11, 2017 10:05

cast Distribution Inputs to floatX instead of tensor_variable

2c160a0

remove double assignments, intX casting, get name function

e456841

WIP continued floatX casting and rebase

3d7fb99

Merge branch 'random_slowdown' of https://github.com/hvasbath/pymc3 i…

66f702c

…nto random_slowdown Conflicts: pymc3/distributions/discrete.py

fixes tests in variational_inference

b90e39b

fix test_distributions_random

966e5e1

Merge branch 'master' into random_slowdown

44c6b52

junpenglao reviewed Jun 12, 2017

View reviewed changes

hvasbath mentioned this pull request Jun 12, 2017

Random slowdown clean #2296

Closed

hvasbath closed this Jun 12, 2017

hvasbath deleted the random_slowdown branch July 2, 2017 09:50

Uh oh!

cast Distribution Inputs to floatX instead of tensor_variable #2233

cast Distribution Inputs to floatX instead of tensor_variable #2233

Uh oh!

Conversation

hvasbath commented May 28, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

junpenglao commented May 28, 2017

Uh oh!

hvasbath commented May 28, 2017

Uh oh!

junpenglao commented May 28, 2017

Uh oh!

junpenglao commented May 28, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fonnesbeck commented May 28, 2017

Uh oh!

ColCarroll left a comment

Choose a reason for hiding this comment

Uh oh!

ColCarroll May 28, 2017

Choose a reason for hiding this comment

Uh oh!

ColCarroll May 28, 2017

Choose a reason for hiding this comment

Uh oh!

hvasbath commented May 29, 2017

Uh oh!

hvasbath commented May 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

junpenglao commented May 29, 2017

Uh oh!

hvasbath commented May 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

junpenglao commented May 29, 2017

Uh oh!

hvasbath commented May 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hvasbath commented May 30, 2017

Uh oh!

junpenglao commented May 30, 2017

Uh oh!

hvasbath commented May 30, 2017

Uh oh!

twiecki commented May 30, 2017

Uh oh!

junpenglao commented May 30, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hvasbath commented Jun 11, 2017

Uh oh!

hvasbath commented Jun 12, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hvasbath commented Jun 12, 2017

Uh oh!

junpenglao Jun 12, 2017

Choose a reason for hiding this comment

Uh oh!

hvasbath Jun 12, 2017

Choose a reason for hiding this comment

Uh oh!

junpenglao Jun 12, 2017

Choose a reason for hiding this comment

Uh oh!

hvasbath Jun 12, 2017

Choose a reason for hiding this comment

Uh oh!

junpenglao Jun 12, 2017

Choose a reason for hiding this comment

Uh oh!

hvasbath commented Jun 12, 2017

Uh oh!

twiecki commented Jun 12, 2017

Uh oh!

hvasbath commented May 28, 2017 •

edited

Loading

junpenglao commented May 28, 2017 •

edited

Loading

hvasbath commented May 29, 2017 •

edited

Loading

hvasbath commented May 29, 2017 •

edited

Loading

hvasbath commented May 29, 2017 •

edited

Loading

junpenglao commented May 30, 2017 •

edited

Loading

hvasbath commented Jun 12, 2017 •

edited

Loading