Set theano config in model context

[aseyboldt]

This should fix #566. I'm not sure this is the best way to do this, it might be better to include the theano context in Context, but this gets a bit complicated because we need to pass arguments though metaclasses and multiple inheritance.
I removed sum_div_dimshuffle_bug. Couldn't even find a reference to this in the theano config. I guess this was removed at some point?
I set compute_test_value to 'ignore' in some functions about the hessian, because I had some trouble with those (see Theano/Theano#5907).

[twiecki]

Nice, this one was bothersome for quite some time, thanks for taking this on. I like the approach.

[ferrine]

I strongly support this PR. That's the thing that always annoys

[ferrine]

No check on exit. Do you really need it?

[ferrine]

If config is set in __new__ this code can be omitted. Attribute will be already set

[ferrine]

Good point having this check. I would also force testvalue in dict

[aseyboldt]

You mean adding the testval option even when a config dict is specified? I guess we could add it if it is not already in there, I'm not sure about this though.

[ferrine]

You do not need it. Init is done in modelcontext

[aseyboldt]

Yes, I think I do. The problem is that subclasses of Model are allowed to add variables in __init__, so __enter__ is called before Model.__init__ using some metaclass magic. But maybe it would be better to move this to the metaclass instead of doing it in init.

[ferrine]

Hm I see the problem now. Yes, It can be done in __new__(here), that's the best place for setting config.

[aseyboldt]

change_flags from theano can't be pickled, because of the usage of nested functions. I wrote a similar function in pymc3.theanof that doesn't have this problem, but I had to use some more internal functionality from theano. If nobody has objections to that, this should be ready to merge (if the tests...)

[aseyboldt]

@ferrine The config is set in __new__ now.

[ferrine]

Let's wait for tests

[twiecki]

==================================== ERRORS ====================================
__________________ ERROR collecting pymc3/tests/test_step.py ___________________
self = <pymc3.tests.test_step.TestStepMethods object at 0x7ff75c0b4b90>
    def test_step_elliptical_slice(self):
        start, model, (K, L, mu, std, noise) = mv_prior_simple()
        unc = noise ** 0.5
        check = (('x', np.mean, mu, unc / 10.),
                 ('x', np.std, std, unc / 10.))
        with model:
            steps = (
                EllipticalSlice(prior_cov=K),
                EllipticalSlice(prior_chol=L),
            )
        for step in steps:
>           trace = sample(5000, step=step, start=start, model=model, random_seed=1)
pymc3/tests/test_step.py:241: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pymc3/sampling.py:260: in sample
    return sample_func(**sample_args)
pymc3/sampling.py:275: in _sample
    for it, strace in enumerate(sampling):
../../../miniconda2/envs/testenv/lib/python2.7/site-packages/tqdm/_tqdm.py:833: in __iter__
    for obj in iterable:
pymc3/sampling.py:377: in _iter_sample
    point = step.step(point)
pymc3/step_methods/arraystep.py:119: in step
    apoint = self.astep(bij.map(point), *inputs)
pymc3/step_methods/elliptical_slice.py:91: in astep
    chol = draw_values([self.prior_chol])
pymc3/distributions/distribution.py:190: in draw_values
    param, point=point, givens=givens.values()))
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
param = Cholesky{lower=True, destructive=False, on_error='raise'}.0
point = None, givens = []
    def draw_value(param, point=None, givens=()):
        if hasattr(param, 'name'):
            if hasattr(param, 'model'):
                if point is not None and param.name in point:
                    value = point[param.name]
                elif hasattr(param, 'random') and param.random is not None:
                    value = param.random(point=point, size=None)
                else:
                    value = param.tag.test_value
            else:
                input_pairs = ([g[0] for g in givens],
                               [g[1] for g in givens])
    
                value = _compile_theano_function(param,
                                                 input_pairs[0])(*input_pairs[1])
        else:
            value = param
    
        # Sanitising values may be necessary.
        if hasattr(value, 'value'):
            value = value.value
        elif hasattr(value, 'get_value'):
            value = value.get_value()
    
        if hasattr(param, 'dtype'):
            value = np.atleast_1d(value).astype(param.dtype)
        if hasattr(param, 'shape'):
            try:
                shape = param.shape.tag.test_value
            except:
                shape = param.shape
>           if len(shape) == 0 and len(value) == 1:
E           TypeError: object of type 'TensorVariable' has no len()
pymc3/distributions/distribution.py:253: TypeError

[aseyboldt]

@twiecki The problem were the tests generated by yield. The generators were executed before the fixture that sets compute_test_value. I turned them into ordinary tests, as yield-tests are deprecated anyway. This also speeds up test discovery.

[twiecki]

Sounds good.

[twiecki]

Thanks @aseyboldt, great to finally have this issue resolved.

[krishnbera]

I am trying to set theano.config.compute_test_value="warn" but I don't know how to specify this in PyMC3.
I tried using pymc3.theanof.set_theano_conf({'compute_test_value':'warn'}) but that doesn't work. After setting this, when I print the value of compute_test_value, it returns raise. What is the correct way of specifying it?