Cost function specification

[mzargham]

I had been writing neural nets code without Theano and recently converted to specifying the experiments using theanets because i want to use Theano to experiment with different cost functions at the output.

I am running a regression model with 96 output variables where Euclidean Distance (~Mean Square Error) is not a very good measure of estimatation accuracy. Since you are built on top of Theano it seems like this cost function should be parameterizeable, but I don't see anything about this in the documentation.

Can you tell me where I can specify the cost function as a Theano function?

[lmjohns3]

Yes, at the moment you need to do this by creating a new subclass for your model. For example, to create a regression model that uses mean absolute error:

class MaeRegressor(theanets.Regressor):
    @property
    def cost(self):
        err = self.outputs[-1] - self.targets
        return TT.mean(abs(err).sum(axis=1))

I've been thinking of making several mixin classes to simplify this sort of task, but mixins can be annoying at times. I haven't really hit upon the "right" way to do this, so for now creating subclasses is it.

Also, I'm still working on the docs for extending theanets; I'll go ahead and put this in as an example.

[mzargham]

Thanks!

On Jan 9, 2015, at 11:58 AM, Leif Johnson notifications@github.com wrote:

Yes, at the moment you need to do this by creating a new subclass for your model. For example, to create a regression model that uses mean absolute error:

class MaeRegressor(theanets.Regressor): @property def cost(self): err = self.outputs[-1] - self.targets return TT.mean(abs(err).sum(axis=1)) I've been thinking of making several mixin classes to simplify this sort of task, but mixins can be annoying at times. I haven't really hit upon the "right" way to do this, so for now creating subclasses is it.

Also, I'm still working on the docs for extending theanets; I'll go ahead and put this in as an example.

— Reply to this email directly or view it on GitHub.