Cannot compute test value because of floatX() in Uniform

[michaelosthege]

When Uniform RVs are instantiated through the pmodel.Var("name", Uniform.dist(...))-API, they can break if the lower/upper bounds are some magic numbers (such as 0).

The following code demonstrates that this is because of the floatx(.) conversion which is only done in the Uniform.__init__:

import unittest
import theano
import theano.tensor as tt
import pymc3
import pymc3.distributions.continuous


class Uniform(pymc3.Continuous):
    R"""..."""

    def __init__(self, lower=0, upper=1, transform='interval',
                 *args, **kwargs):
        if transform == 'interval':
            transform = pymc3.distributions.continuous.transforms.interval(lower, upper)
        super(Uniform, self).__init__(transform=transform, *args, **kwargs)

        #self.lower = lower = pymc3.distributions.continuous.floatX(tt.as_tensor_variable(lower))
        #self.upper = upper = pymc3.distributions.continuous.floatX(tt.as_tensor_variable(upper))
        self.lower = lower = tt.as_tensor_variable(lower)
        self.upper = upper = tt.as_tensor_variable(upper)
        self.mean = (upper + lower) / 2.
        self.median = self.mean

    def random(self, point=None, size=None, repeat=None):
        return pymc3.distributions.continuous.Uniform.random(self, point, size, repeat)

    def logp(self, value):
        return pymc3.distributions.continuous.Uniform.logp(self, value)

    def _repr_latex_(self, name=None, dist=None):
        return pymc3.distributions.continuous.Uniform._repr_latex_(self, name, dist)

Uniform_type = Uniform         # this is all fine
#Uniform_type = pymc3.Uniform    # this breaks

class TestDistAPI(unittest.TestCase):

    def test_1(self):
        print("test 1")
        u = Uniform_type.dist(lower=0.0, upper=1.4, testval=0.5)
        with pymc3.Model() as pmodel:
            pmodel.Var("u", u)
        return

    def test_2(self):
        print("test 2")
        u = Uniform_type.dist(lower=0.01, upper=1.4, testval=0.5)
        with pymc3.Model() as pmodel:
            pmodel.Var("u", u)
        return

    def test_3(self):
        print("test 3")
        with pymc3.Model() as pmodel:
            Uniform_type("u", lower=0, upper=1.4, testval=0.5)
        return


if __name__ == "__main__":
    print(theano.config.compute_test_value)
    ## >>> 'off'
    unittest.main(exit=False)

The execution order of the three test methods is relevant! This is probably because if test_3 runs first, theano caches the 0-constant and the other tests work fine, but if test_1 runs first, it breaks. test_2 shows that some constants don't cause trouble.

The failures can NOT be prevented via theano.config.compute_test_value (it already is off in my case).

Why is there the floatX(.) conversion in continuous.py#L158 ?
Can we remove it? (It's also appears in L109 by the way.)

[junpenglao]

Thanks for the detail investigation @michaelosthege.
What os you are using? I am on Linux and cannot reproduce the error (all test run fine)

[junpenglao]

The failures can NOT be prevented via theano.config.compute_test_value (it already is off in my case).

Actually for me it failed if theano.config.compute_test_value='off', but works fine if theano.config.compute_test_value='raise'

[michaelosthege]

I am testing on three setups and have not made observations that would suggest it being OS-specific:

• Python 3.5 on Windows 10 x64
• Python 3.5 on Ubuntu on Windows 10 (Windows Subsystem for Linux)
• Python 3.5 on Ubuntu 17.04 (freshly installed a few hours ago)

This issue may be related: #562
In the example I posted there, I tried to set theano_config of the pymc3.Model() because that one actually overrides whatever theano.config you set elsewhere.

I just ran this on the recently installed Ubuntu 17.04 and it promptly failed:

import pymc3

u = pymc3.DiscreteUniform.dist(0, 4, testval=3)
g = pymc3.Gamma.dist(alpha=12, beta=2, testval=8.2)
with pymc3.Model() as pmodel:
    pmodel.Var("u", u)
    pmodel.Var("g", g)

[junpenglao]

It run fine with me...

pmodel.test_point
Out[14]: {'g_log__': array(2.1041341542702074), 'u': array(3)}

I am on pymc3 master, and

CPython 3.5.2
IPython 6.1.0

numpy 1.13.1
pandas 0.20.3
theano 0.10.0dev1.dev-d51233517debf3a1231a33058d65e8f969db923d

compiler   : GCC 5.4.0 20160609
system     : Linux
release    : 4.4.0-96-generic
machine    : x86_64
processor  : x86_64
CPU cores  : 8
interpreter: 64bit

[michaelosthege]

On an empty environment, I did pip install pymc3 and then pip install git+https://github.com/pymc-devs/pymc3 -U --no-deps to switch to master.

(venv35) xxxxxx@xxxxxx:~$ pip list
certifi (2016.2.28)
cycler (0.10.0)
enum34 (1.1.6)
h5py (2.7.1)
joblib (0.11)
matplotlib (2.0.2)
numpy (1.13.1)
pandas (0.20.3)
patsy (0.4.1)
pip (9.0.1)
pymc3 (3.1)
pyparsing (2.2.0)
python-dateutil (2.6.1)
pytz (2017.2)
scipy (0.19.1)
setuptools (36.4.0)
six (1.11.0)
Theano (0.9.0)
tqdm (4.17.1)
wheel (0.29.0)
(venv35) xxxxxx@xxxxxx:~$ python
Python 3.5.4 |Continuum Analytics, Inc.| (default, Aug 14 2017, 13:26:58)
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>>

Nevertheless, if I run the snipped from my previous post as test.py:

(venv35) xxxxxx@xxxxxx:~$ python test.py
Traceback (most recent call last):
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/theano/gof/op.py", line 625, in __call__
    storage_map[ins] = [self._get_test_value(ins)]
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/theano/gof/op.py", line 581, in _get_test_value
    raise AttributeError('%s has no test value %s' % (v, detailed_err_msg))
AttributeError: Elemwise{Cast{int32}}.0 has no test value
Backtrace when that variable is created:

  File "test.py", line 3, in <module>
    u = pymc3.DiscreteUniform.dist(0, 4, testval=3)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/distributions/distribution.py", line 47, in dist
    dist.__init__(*args, **kwargs)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/distributions/discrete.py", line 443, in __init__
    self.upper = tt.floor(upper).astype('int32')


During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "test.py", line 6, in <module>
    pmodel.Var("u", u)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/model.py", line 752, in Var
    total_size=total_size, model=self)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/model.py", line 1123, in __init__
    self.logp_elemwiset = distribution.logp(self)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/distributions/discrete.py", line 464, in logp
    return bound(-tt.log(upper - lower + 1),
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/theano/tensor/var.py", line 147, in __sub__
    return theano.tensor.basic.sub(self, other)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/theano/gof/op.py", line 639, in __call__
    (i, ins, node, detailed_err_msg))
ValueError: Cannot compute test value: input 0 (Elemwise{Cast{int32}}.0) of Op Elemwise{sub,no_inplace}(Elemwise{Cast{int32}}.0, Elemwise{Cast{int32}}.0) missing default value.
Backtrace when that variable is created:

  File "test.py", line 3, in <module>
    u = pymc3.DiscreteUniform.dist(0, 4, testval=3)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/distributions/distribution.py", line 47, in dist
    dist.__init__(*args, **kwargs)
  File "/home/xxxxxx/miniconda3/envs/venv35/lib/python3.5/site-packages/pymc3/distributions/discrete.py", line 443, in __init__
    self.upper = tt.floor(upper).astype('int32')

But it is definitely important that you run it on a fresh session, because if theano already created a similar constant tensor before, it is cached. (This may also explain why unittests rarely catch these issues.)