garch example fails

[mikemac8888]

Running examples/garch_example.py fails with error:

Traceback (most recent call last):
  File "garch_example.py", line 40, in <module>
    beta1 = BoundedNormal('beta1', 0, sd=1e6)
  File "/Users/**/anaconda3/envs/py35/lib/python3.5/site-packages/pymc3/distributions/continuous.py", line 1102, in __call__
    *args, **kwargs)
  File "/Users/**/anaconda3/envs/py35/lib/python3.5/site-packages/pymc3/distributions/distribution.py", line 27, in __new__
    return model.Var(name, dist, data)
  File "/Users/**/anaconda3/envs/py35/lib/python3.5/site-packages/pymc3/model.py", line 288, in Var
    transform=dist.transform)
  File "/Users/**/anaconda3/envs/py35/lib/python3.5/site-packages/pymc3/model.py", line 689, in __init__
    transformed_name = "{}_{}_".format(name, transform.name)
AttributeError: 'int' object has no attribute 'name'

Python 3.5.2
pymc3 (3.0rc2)

[twiecki]

There's actually two issues at play:

1. Normal needs mu=0 explicitly.
2. That leads to the problem of passing a tensor to Bounded which infers the bounds using np.isinf() which chokes on the tensor. @ColCarroll any ideas?

[twiecki]

Also not sure why the tests pass.

[springcoil]

This is interesting - did I write this example originally?

[springcoil]

Is this related? #1170 I think it might be - because the sample is taking advi normally, and Bound might not be working properly. Any ideas @ColCarroll ?

[springcoil]

I can confirm that changing the sampling to nuts for example makes no difference.

[ColCarroll]

the stuff in the examples folder is not tested -- we moved some of these to the test suite, and recommend moving the others to full fledged notebooks. i'm looking at why this doesn't run, though!

[ColCarroll]

Actually, there are a few funny things going on here -- I haven't touched much of the transform code, but here's what I've got:

-- As pointed out, you have to supply mu= in a few places to get things running
-- You also have to change all the checks in Bounded to use tt.isinf instead of np.isinf
-- The example hits Bounded.__init__ with lower = -inf and upper being a theano tensor, but falls through all three checks (we should change those transform checks to if...elif...elif...else)
-- the call to self.__dict__.update(locals()) means the "default" transform is the string "infer"

[springcoil]

Can you refactor this a bit then Colin? I think it's clear what the fix is now.

…

On Mon, Dec 12, 2016 at 11:47 PM, Colin ***@***.***> wrote: Actually, there are a few funny things going on here -- I haven't touched much of the transform code, but here's what I've got: -- As pointed out, you have to supply mu= in a few places to get things running -- You also have to change all the checks in Bounded to use tt.isinf instead of np.isinf -- The example hits Bounded.__init__ with lower = -inf and upper being a theano tensor, but falls through all three checks (we should change those transform checks to if...elif...elif...else) -- the call to self.__dict__.update(locals()) means the "default" transform is the string "infer" — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1491 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8DiOGfMsap1psX4F-K9VKPlqGjE_oVks5rHdz2gaJpZM4Kl-gL> .

-- Peadar Coyle Skype: springcoilarch www.twitter.com/springcoil peadarcoyle.wordpress.com

[ColCarroll]

Oy, it is a little more complicated than all that -- I was fooled because tt.isinf is truthy for any input. I think it can work for this example still -- I'll take a look.

[kris-singh]

@twiecki @ColCarroll i think i figured it out. we could check if the type(float) is float or not. Then use np.isinf for the float variable.
But i am having trouble understanding this,

1. in the given example upper evaluates to (TensorConstant{1} - alpha1) the alpha1 value will available only when we sample the from the normal distribution. So how can we compare default >= upper .
2. is default also a theano variable

is this making any sense?

[kris-singh]

if isinstance(upper, tt.TensorVariable):
                upperisinf = tt.isinf(upper).eval()
            else:
                upperisinf = np.isinf(upper)

            if isinstance(lower, tt.TensorVariable):
                lowerisinf = tt.inf(lower).eval()
            else:
                lowerisinf = np.isinf(lower)

            if not lowerisinf and not upperisinf:
                self.transform = transforms.interval(lower, upper)
                if default <= lower or default >= upper:
                    self.testval = 0.5 * (upper + lower)

            if not lowerisinf and upperisinf:
                self.transform = transforms.lowerbound(lower)
                if default <= lower:
                    self.testval = lower + 1

            if lowerisinf and not upperisinf:
                self.transform = transforms.upperbound(upper)
                if default >= upper:
                    self.testval = upper - 1

Error: theano.gof.fg.MissingInputError: An input of the graph, used to compute Elemwise{sub,no_inplace}(DimShuffle{x}.0, alpha1), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.

[kris-singh]

Is there any code that is actually compiling the theano function. I tried to look for it in the sampling code but coudn't find it. Could anybody help.

[kris-singh]

this gist might help clarifying what i am asking.
https://gist.github.com/kris-singh/218542ddab4b62248731926c22da27b5

[twiecki]

@kris-singh The code is not run like python code. Maybe reading a bit on the theano docs clears things up: http://deeplearning.net/software/theano/

[kris-singh]

@twiecki if you could take a look at the gist. you would understand my doubt better.

[twiecki]

@kris-singh I looked at the gist. What output do you expect?

[kris-singh]

@twiecki since we get the theano variables lets say var1 and var2 and var3 = var1*var2. theano would have only symbolic variable. we need to supply inputs and then compile the var3 to get its output. where in the pymc3 code is this happening.

[kris-singh]

@twiecki something like this

import theano.tensor as T
from theano import function
a = T.dscalar('a')
b = T.dscalar('b')
c = a*b
f = function([a.b],c)
f(2,3)

[twiecki]

That all happens in model.py.

[twiecki]

https://github.com/pymc-devs/pymc3/blob/master/pymc3/model.py#L594

[kris-singh]

@twiecki and the inputs to the these functions are provided using the sample function right?

[twiecki]

Exactly, in step.

[kris-singh]

any ideas on how to solve this. I am stuck. if use something like this tt.le(default,upper). should the upper values from model.var. or how ???

[kris-singh]

can anybody help

[twiecki]

This is quite a tricky issue. Can we evaluate the tensor as it is passed in? We don't need to do the inference during sampling, only during instantiation.

[twiecki]

Oh, that's what you are doing above with isinf. I think we can do the same with lower and upper, no?

[kris-singh]

@twiecki but while doing this with isinf also there is a theano error. The inputs are not provided. So we need to figure out how we are going to provide the input values to these variables.

[twiecki]

Ah, I think if a tensor is passed in we can safely assume that it's not inf.

[twiecki]

@kris-singh I think we just need to remove the checks if a tensor is passed. We will assume that it's not inf and not set a test_val (and not check for its value). Does that make sense?

[kris-singh]

yes it does. but who to evaluate the tensor like upper when doinging default >= upper ?

[twiecki]

just not set it.

…

On Feb 19, 2017 7:43 AM, "Kris Singh" ***@***.***> wrote: yes it does. but who to evaluate the tensor like upper when doinging default >= upper ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1491 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AApJmEfRK2hEUZcqY7YpB0ZnVSPjFV3cks5rd-ShgaJpZM4Kl-gL> .

[twiecki]

or better yet use the .tag.test_value

…

On Feb 19, 2017 8:31 AM, "Thomas Wiecki" ***@***.***> wrote: just not set it. On Feb 19, 2017 7:43 AM, "Kris Singh" ***@***.***> wrote: > yes it does. but who to evaluate the tensor like upper when doinging > default >= upper ? > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#1491 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AApJmEfRK2hEUZcqY7YpB0ZnVSPjFV3cks5rd-ShgaJpZM4Kl-gL> > . >