jondeuce / TopoChains.jl

A flexible data structure for multi-input multi-output models

TopoChains.jl

TopoChains.jl provides a new data structure, a TopoChain, which picks up where Flux.jl's Chain left off. Unlike a Chain, which is designed for a sequential single-input single-output use case, a TopoChain can handle multi-input multi-output compositions of layers and functions. We achieve this by seperately specifying the topology (that is, the structure) of a model from the actual layers of a model. Check the docs for more!

The TopoChain was originally designed as the Stack, as part of Transformers.jl by Peter Cheng, and is repackaged here into a standalone package for general purpose use.

Usage

Suppose you want to define the following model:

First, we define the structure of the model:

topo = @functopo x:x => a:x => b:(a, b) => c => o

Then we define the model itself:

model = TopoChain(topo,
            Dense(32, 64),
            Dense(32, 64),
            (x, y) -> x .* y, 
            Dropout(0.1))

And that's it, you're done! You can see that the model both contains your layers, as well as information on how to run the layers:

model
# TopoChain(Dense(32, 64), Dense(32, 64), #7, Dropout(0.1)) representing the following function composition: 
# function(x)
#     a = Dense(32, 64)(x)
#     b = Dense(32, 64)(x)
#     c = #7(a, b)
#     o = Dropout(0.1)(c)
#     o
# end

TopoChain is a drop-in replacement for a Chain, which means all the features of Chain, such as parameter collection, indexing, layer slicing, etc. will work as intended.

Installation

You can get started by installing TopoChains.jl with the following command:

] add TopoChains