Add wrapper for theano as_op

[karlnapf]

Hi there,

is there a way to make pymc3 use custom density functions that are not theano compatible?
Say I have some black box c-code to evaluate my log-pdf, and now I want to use this with one of pymc3's gradient free samplers such as metropolis.

Currently, trying this gives me things like object of type 'TensorVariable' has no len() when trying to use my own python code to evaluate a log-pdf (which assumes the input is a numpy array).

Thanks for your help and for the great package!

[twiecki]

Hi Heiko,

Seems to be a dupe of #507. E.g.

@theano.compile.ops.as_op(itypes=[t.lscalar, t.dscalar, t.dscalar],otypes=[t.dvector])
def rate(switchpoint,early_mean, late_mean):
    ''' Concatenate Poisson means '''
    out = empty(years)
    out[:switchpoint] = early_mean
    out[switchpoint:] = late_mean
    return out

Let me know if that works or doesn't work for you.

[twiecki]

Closing, feel free to reopen if this doesn't work for you.

[karlnapf]

Hi Thomas,

thanks for the help, that fixed the problem. Amazing! Now I can do pseudo-marginal MCMC models using PyMC, which is something I am working on and would like to use PyMC for.

For others, here is how I did it:
http://nbviewer.ipython.org/gist/karlnapf/8ed7de163eed39e733da

[twiecki]

Cool, thanks for sharing!

[twiecki]

I also wonder if we should create a wrapper around @theano.compile.ops.as_op(itypes=[T.dscalar],otypes=[T.dscalar]) because it requires another call to pm.DensityDist('y', logp). Something along the lines of:

@CustomDensity(itypes=[T.dscalar],otypes=[T.dscalar])
def MyCustomDist():...

with model:
    my_dist = MyCustomDist('y', ...)

CC @fonnesbeck @jsalvatier

[jsalvatier]

Yes, that's true. I wonder if it should be T.dvector though, since that
seems like a more common case. I like this idea.

On Tue, Aug 19, 2014 at 4:55 AM, Thomas Wiecki [email protected]
wrote:

I also wonder if we should create a wrapper around
@theano.compile.ops.as_op(itypes=[T.dscalar],otypes=[T.dscalar]) because
it requires another call to pm.DensityDist('y', logp). Something along
the lines of:

@CustomDensity(itypes=[T.dscalar],otypes=[T.dscalar])def MyCustomDist():...
with model:
my_dist = MyCustomDist('y', ...)

CC @fonnesbeck https://github.com/fonnesbeck @jsalvatier
https://github.com/jsalvatier

—
Reply to this email directly or view it on GitHub
#584 (comment).

[fonnesbeck]

I like it too. In favor of anything that hides the Theano complexity from most users.

[jsalvatier]

Theano/Theano#3094 issue should make it easier to make as_op make theano ops.

[fonnesbeck]

Is this still something we worry about? Its pretty easy to create custom distributions now.

[twiecki]

Yeah, I think it's fine, haven't heard many complaints.