I know this isn’t necessarily part of the course, but thought a lot of people here might have a better idea on how Theano works internally. Say that I want to do a simple dot product between a matrix and a vector
It makes sense that we first need to create the variables as that’s how the computation graph is built
A = T.matrix("A")
v = T.vector("v")
then we define the dot product, which continues building the computation graph adding an operation between A and v and stores the result of the dot product in a new graph node x.
x = A.dot(v)
then it makes sense that we have to compile the graph into actual executable code, so I assume this is where the CUDA’s nvcc gets the job done and builds a native function:
f = theano.function(inputs=[A, v], outputs=w)
but what I don’t is why we need to specify both the inputs and the outputs? Why can’t theano simply infer what are the inputs based on the resulting node w? Is this because we might create a function for only a portion of the whole graph?
If so, can I take an arbitrary computation graph and “slice” a part of it as a function by simply saying which nodes are the starting points and which I want as the result, and build a Theano function out of that?
edit: Just to clarify, my question is why do I need to specify both the inputs and outputs when the computation dependencies are already defined by the outputs and Theano would have to verify that the graph is complete anyway, right?