advanced.md

Advanced Uses

Collect Variables

You can collect intermediate variables you are interested in with ' on that variable. This allows you to define and train a model, and only make changes to its structure (and not the layers of the model itself) if you need access to intermediate outputs for downstream tasks (e.g. style transfer using a pretrained CNN). For example:

julia> @functopo (x,y) => (a,b,c,d') => (w',r',y) => (m,n)' => z
# FuncTopo{"(x, y) => ((a, b, c, d') => ((w', r', y) => (((m, n))' => z)))"}
# function(model, x, y)
#     (a, b, c, d) = model[1](x, y)
#     %1 = d
#     (w, r, y) = model[2](a, b, c, d)
#     %2 = (w, r)
#     (m, n) = model[3](w, r, y)
#     %3 = (m, n)
#     z = model[4](m, n)
#     (z, (var"%1", var"%2", var"%3"))
# end

Interpolation

TopoChains.jl supports interpolation, so you can use a variable to hold a substructure or the unroll number. !!! note The interpolation variable should always be at the top level of the module since we can only get that value with eval. (To interpolate local variables, use @functopo_str "topo_pattern" instead)

@functopo_str

N = 3

topo = @functopo (y => (z1, z2) => t) => $N

# alternatively
# topo = @functopo_str "(y => (z1, z2) => t) => $N"

topo
# FuncTopo{"(y => ((z1, z2) => t)) => 3"}
# function(model, y)
#     (z1, z2) = model[1](y)
#     t = model[2](z1, z2)
#     (z1, z2) = model[3](t)
#     t = model[4](z1, z2)
#     (z1, z2) = model[5](t)
#     t = model[6](z1, z2)
#     t
# end