Anyone here a Theano expert? A few questions after my first "successful" usage.

Hi all,

So I tried using Theano to see how well the gpu stuff worked today. I have to say, I was a bit disappointed: I only got maybe a 14% speedup over using just numpy.

I'm implementing gradient descent on a convex function (softmax regression). The goal is to find the input vector x which produces a given output vector.

Iteration computations are dominated by a few matrix multiplications (a roughly 1 x 75k matrix by a 75k x 300 matrix). Is this perhaps not big enough to justify the additional code complexities of Theano?

A couple things: - I'm only using shared variables, so there's no unnecessary bus transfer back and forth AFAIK - I compute my loss function on the gpu, then transfer that one scalar value back to the host to print it each iteration and test

Additionally, do you know if there's any major advantage to using theano's automatic grad differentiator if I've already computed the derivative by hand?

Is there any reason to use theano's scan function (which looks difficult to use, to say the least) over just keeping track of my iteration count in pure python code?

Thanks, slashcom