[theano-users] Compiled cost function (with theano.function) ?

[Sym]

Hi,

I have a fairly complex cost function, composed of various pieces that are
looping over the data.

Should I compile the total cost as a theano function or not ?
Would the network still be able to compute gradients if I do this ?
Would it be faster / slower than compiling everything (network + cost
functions) in one theano function ?

I am not sure about what would be the benefits of doing this, if there are
any.

[Sym]

I would like to add that the cost function is looping over the input
tensors.

I need these loops to be differentiable, so I've implemented them with
scan/map. But in this answer
<https://groups.google.com/forum/#!msg/theano-users/uqYtNlpwiHw/k-3QCro0SbkJ>
I saw that it is possible to compile a theano function and call it with a
Python for loop.

So again I don't know what's best.. Should I compile a step of the loop in
a theano function, and then loop with Python's for op ?
Or should I compile a theano function with the whole cost (where the loops
are now done with scan) ?
Or should I just compile one theano function, with the network and the cost
?


Thank you for your help !

Post by Sym
Hi,
I have a fairly complex cost function, composed of various pieces that
are looping over the data.
Should I compile the total cost as a theano function or not ?
Would the network still be able to compute gradients if I do this ?
Would it be faster / slower than compiling everything (network + cost
functions) in one theano function ?
I am not sure about what would be the benefits of doing this, if there are
any.