gistfile1.py

import theano
import theano.tensor as T
import gnumpy
import theano.misc.gnumpy_utils as gput

g = gnumpy.zeros((1, 2))
g += np.array([[2, 1]])

x = T.matrix()
expr = 2 * x

# Naive approach. This does not work when called with a CudaNdArray.
f = theano.function([x], 2 * x)

def function(inpt, inpt_templates, expr):
    # Create shared variables for each input template.
    shared = [theano.shared(np.empty(i.shape, dtype='float32')) for i in inpt_templates]
    
    # Let the function we will compile with theano point at the shared variables instead of the symbolic ones.
    for i, s in zip(inpt, shared):
        expr = theano.clone(expr, {i: s})
    
    # Compile a function with no arguments to call.
    f = theano.function([], expr)
    
    def inner(*args):
        for a, s in zip(args, shared):
            s.set_value(gput.garray_to_cudandarray(a))
        return f()
    
    return inner

# Does the job.
f2 = function([x], [g], expr)
f2(g)