The dot product of int32 and float32 produces float64 in Theano

Question

When I compile the function

x = theano.tensor.imatrix('x')
y = theano.tensor.fmatrix('y')
z = x.dot(y)
f = theano.function([x, y], z)

the resulting output is a float64, even when x is of type int32 and y is of type float32. When I compute the same operation where x is a Theano fmatrix of type float32, the resulting matrix is a float32. Why is the smaller bit size not preserved in the former case? In other words, why is the dot product of an int32 and a float32 equal to a float64 instead of float32 in Theano?

I'm using Theano version 0.9.0

Answer 1

In general, it is assumed that you want to keep whatever precision you start out with. int32 has 31 significand bits, float32 has 24 significand bits and 8 exponent bits. Thus staying at either may cause reduced precision, and float64 is chosen, with its 53 significand bits.

You can also configure theano not to use float64 by default .