nevermind - I realized that I should not even have a custom Op for
something that is already a theano computation graph - because then
optimizations don't work anymore.
thanks for the help!
Am Mittwoch, 2. August 2017 16:31:57 UTC+2 schrieb Michael Osthege:
> Your suggestion put me on the right track and I'm halfway there.
> I actually have two Ops - one that's doing a perform() using Python
> functions (I got this to work) and another that I can make purely from
> theano functions:
> def make_node(self, y0, theta, dt, n):
> t_y0 = tt.as_tensor_variable(y0)
> t_theta = tt.as_tensor_variable(theta)
> t_dt = tt.as_tensor_variable(dt)
> t_n = tt.as_tensor_variable(n)
> Y_hat, updates = theano.scan(fn=self.dydt,
> outputs_info =[{'initial':t_y0}],
> sequences=[theano.tensor.arange(t_dt,
> dt*(t_n+1), t_dt)],
> non_sequences=[t_dt, t_theta],
> n_steps=n-1)
>
> Y_hat = tt.concatenate((t_y0[None,:], Y_hat)) # scan does not
> return y0...
> Y_hat = tt.transpose(Y_hat) # return as
> (len(y0),n)
>
> apply_node = theano.Apply(self,
> [t_y0, t_theta], # symbolic
> inputs: y0 and theta
> [Y_hat]) # symbolic
> outputs: Y_hat
> return apply_node
>
> The computation graph leading to Y_hat is fine, but Apply() raises an
> exception: "All output variables passed to Apply must belong to it."
> because Y_hat.owner is already defined.
>
> From my understanding I should not use perform() here because the inputs
> and outputs to the actual computation should not happend via numpy.
> Where is the right place to put this sub-graph in the Op implementation?
>
>
>
>
> Am Dienstag, 1. August 2017 19:25:34 UTC+2 schrieb nouiz:
>
>> Just don't use otypes, itypes. Implement the make_node() method. It isn't
>> hard. We won't extend shortly as_op or otypes/itypes.
>>
>> If you have question about make_node(), you can ask them.
>>
>> Fred
>>
>> On Tue, Aug 1, 2017 at 12:33 PM 'Michael Osthege' via theano-users <
>> [email protected]> wrote:
>>
>>> sorry to dig up this old thread, but I am also working with pymc3 and
>>>>> have a related problem:
>>>>>
>>>>
>>> I am trying to create custom Ops for integrating an ODE model. I can
>>> already do it with as_op, but that can't be pickled leading to problems
>>> with parallelization in pymc3.
>>>
>>> I followed the theano documentation to implement a custom Op, but I
>>> noticed a problem with the otypes. The *Ops otypes is a list of
>>> dvector, but the length of that list can change with the Op parameters*.
>>> But the itypes/otypes are not instance-attributes but class attributes. So
>>> theoretically I *can't have multiply custom Ops of the same type that
>>> have different itypes/otypes*, right?
>>>
>>> Also, do you have any idea how I could circumvent this?
>>>
>>> cheers
>>>
>>> --
>>>
>>> ---
>>> You received this message because you are subscribed to the Google
>>> Groups "theano-users" group.
>>> To unsubscribe from this group and stop receiving emails from it, send
>>> an email to [email protected].
>>> For more options, visit https://groups.google.com/d/optout.
>>>
>>
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
