我应该如何在theano函数中分配一个numpy数组？

Question

Let's say I have a theano function:

def my_fun(x, y):
  # Create output array for example sake
  z = np.asarray(
    shape=(x.shape[0], y.shape[1]),
    dtype=theano.config.floatX
  )

  z = x + y

  # this is wrong, how should I convert this to a theano
  # tensor?
  return z

x = theano.tensor.dmatrix("x")
y = theano.tensor.dmatrix("y")

f = function(
  inputs=[x, y],
  outputs=[my_fun]
)

a = numpy.asarray([[1,2],[3,4]])
b = numpy.asarray([[1,2],[3,4]])

c = my_fun(a,b)

1. How should I allocate tensors/ arrays or memory within the actual theano to be optimized when compiled by theano.
2. How should I convert that allocated tensor/ array whatever to a theano like variable to be returned? I've tried converting it to a shared variable in the function but that didn't work.

Answer 1

I'm sorry but I don't understand your specific questions but can comment on the code sample you provided.

Firstly, your comment above return z is incorrect. If x and y are Theano variables then z will also be a Theano variable after z = x + y .

Secondly, there is no need to pre-allocate memory, using numpy, for return variables. So your my_fun can change to simply

def my_fun(x, y):
  z = x + y
  return z

Thirdly, the output(s) of Theano functions need to be Theano variables, not Python functions. And the output needs to be a function of the inputs. So your theano.function call needs to be changed to

f = function(
  inputs=[x, y],
  outputs=[my_fun(x, y)]
)

The most important point to grasp about Theano, which can be a little difficult to get one's head around when starting out, is the difference between the symbolic world and the executable world. Tied in to that is the difference between Python expressions and Theano expressions.

The modified my_fun above could be used like a symbolic function or like a normal executable Python function but it behaves differently for each. If you pass in normal Python inputs then the addition operation occurs immediately and the return value is the result of the computation. So my_fun(1,2) returns 3 . If instead you pass in symbolic Theano variables then the addition operation does not take place immediately. Instead the function returns a symbolic expression that after later being compiled and executed will return the result of adding two inputs. So the result of my_fun(theano.tensor.scalar(), theano.tensor.scalar()) is a Python object that represents a symbolic Theano computation graph. When that result is passed as the output to a theano.function it is compiled into something that is executable. Thean when the compiled function is executed, and given some concrete values for the inputs, you actually get the result you were looking for.