theano and sympy

[draz...@gmail.com]

Hi,

I was experimenting with the following code:

import sympy

from sympy.abc import x

expr = sympy.exp(-x ** 2)

from sympy.printing.theanocode import theano_function

fn_theano  = theano_function([x], [expr], dims={x: 1}, dtypes={x: 'float64'})

print fn_theano([29.]),sympy.exp(-29.**2)

I got 

[ 0.] 5.73245586032578e-366

My question is why fn_theano does not  give the same value?

Thank you,

Costas

[Jason Moore]

Theano is doing floating point computations whereas SymPy does the computation using arbitrary precision.

Jason
moorepants.info
+01 530-601-9791

--
You received this message because you are subscribed to the Google Groups "sympy" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sympy+un...@googlegroups.com.
To post to this group, send email to sy...@googlegroups.com.
Visit this group at http://groups.google.com/group/sympy.
To view this discussion on the web visit https://groups.google.com/d/msgid/sympy/4662668c-68b0-482e-b467-92a6d2206d2e%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[draz...@gmail.com]

Thank you for your answer.

So, in what sense theano cooperates with sympy (or the inverse)?

Since, we can not have the advantage of sympy functions in theano.

Costas

[Frédéric Bastien]

The execution by Theano should be faster when you work big not too small inputs size. If you work only on scalar, Theano is probably not faster then sympy.

But if you work on vector/tensor big enough (this size depend of the function...) then Theano get faster.

Fred

--
You received this message because you are subscribed to the Google Groups "sympy" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sympy+un...@googlegroups.com.
To post to this group, send email to sy...@googlegroups.com.
Visit this group at http://groups.google.com/group/sympy.

To view this discussion on the web visit https://groups.google.com/d/msgid/sympy/3d09a5c6-eaea-4898-83cb-99b8a74f48a7%40googlegroups.com.

[draz...@gmail.com]

ok, I oversimplified my example. In fact my computations are in vectors. The problem seems to be that I cannot pass the results of 

[sympy.exp(x[i]) for i in range(dim)] to Theano (for x[i] small enough). Theano evaluate them as zero but here I need the real values of exp([x[i]]), where only

sympy.exp can give me.

Costas  

[Frédéric Bastien]

Theano use real number (float32 or float64), not bigfloat. If you need bigfloat precission, there isn't easy way to do it in Theano. We have stability optimization that help in some cases, but it isn't equivalent to bigfloat.

If you have just one op that need bigfloat, you can build a quick Theano op to wrap other python code that do the execution:

http://deeplearning.net/software/theano/tutorial/extending_theano.html#as-op

But your code:

[sympy.exp(x[i]) for i in range(dim)]

do not seem vectorised. In theano x would be a vector and you would just do theano.tensor.exp(x).

Fred

To view this discussion on the web visit https://groups.google.com/d/msgid/sympy/95980104-7afc-4b3c-b0e0-5d618f222d78%40googlegroups.com.