Problems using breze with GPU

[makarl]

when using breze with theano using the GPU the program crashes in breze.arch.util.Model._unify_exprs(self, exprs) with the error raise ValueError("length not known")

this happens because exprs is not of the type theano.tensor.basic.TensorVariable but of the type theano.sandbox.cuda.var.CudaNdarrayVariable.

_unify_exprs then tries to iterate trough exprs, which results in raising the raise ValueError("length not known") exception

im using cudamat from http://code.google.com/p/cudamat/
and gnumpy from http://www.cs.toronto.edu/~tijmen/gnumpy.html

[makarl]

additional to your fix i needed to add
if isinstance(expr, theano.sandbox.cuda.var.CudaNdarrayVariable):
return expr
inside breze.arch.util.cpu_expr_to_gpu to to catch the case when the expression is already a Gpu expression

its now working!
thank you all

[bayerj]

Can someone please add a summary of this as a comment on top of this code?
I can then merge it.

On 25 April 2014 12:57, makarl [email protected] wrote:

additional to your fix i needed to add
if isinstance(expr, theano.sandbox.cuda.var.CudaNdarrayVariable):
return expr
inside breze.arch.util.cpu_expr_to_gpu to to catch the case when the
expression is already a Gpu expression

its now working!
thank you all

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/40#issuecomment-41380802
.

Dipl. Inf. Justin Bayer
Lehrstuhl für Robotik und Echtzeitsysteme, Technische Universität München
http://www6.in.tum.de/Main/Bayerj

[makarl]

I just added a merge request, don't know how to connect it to this conversation

[makarl]

please do not merge it right now, I want to make some tests to be really sure everything works fine

[bayerj]

Is #42 fixing this?

[makarl]

It's now working for my network, but it's extremely slow, like ten times compared to the CPU version.
Additionally you will need fix #43, since lbfgs can't handle gnumpy scalars. I set numpy_result to True only for loss functions. Is this the correct way ?