FluxML/Flux.jl

Shape-propagating Chain

Open

#703 opened on Mar 26, 2019

View on GitHub
 (8 comments) (8 reactions) (0 assignees)Julia (4,725 stars) (619 forks)batch import
discussionenhancementhelp wanted

Description

It'd be nice to be able to write something like

model = @Chain(
  Input(28^2),
  Dense(32, relu),
  Dense(10),
  softmax)

It's a relatively minor convenience but it does avoid some redundancy when specifying chains, which is tedious to correct and easy to get wrong when trying different layer sizes.

Here's roughly how I imagine this working. The @Chain would expand to something like

shape = nothing
layer1, shape = fromshape(Input, shape, 10)
layer2, shape = fromshape(Dense, shape, 32, relu)
...
Chain(layer1, layer2, ...)

fromshape can then forward to an appropriate constructor or error for non-supported layers. Hopefully this strikes the right balance of simplicity/generality and we don't end up having to turn it into a full shape inference system.

Contributor guide

Tech stack
None
Domain
machine learning
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
Basic knowledge of Flux.jlUnderstanding of Julia macrosFamiliarity with neural network layer construction
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
Examine Flux.jl's existing @Chain macro implementation (likely in src/chains.jl) to understand current syntax and expansion. Define a fromshape function that takes a layer type, previous shape, and parameters to infer the output shape. Implement fromshape for common layers like Dense and Input. Ensure backward compatibility. Write tests for various chain configurations. Consult issue comments for additional use cases.