We do this as in some case the test was failing during theano fct call.(i.e. under ubuntu)
As we use theano.tensor.scalar and not theano.scalar.scalar normally this is not a problem.