Data

Tools

Indexes

Applications

Experiments

Open Source Science

https://github.com/darsnack/fluxprune.jl

Pruning framework and methods for Flux

https://github.com/darsnack/fluxprune.jl

Science Score: 23.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • ○
    CITATION.cff file
  • ✓
    codemeta.json file
    Found codemeta.json file
  • ○
    .zenodo.json file
  • ○
    DOI references
  • ○
    Academic publication links
  • ✓
    Committers with academic emails
    1 of 1 committers (100.0%) from academic institutions
  • ○
    Institutional organization owner
  • ○
    JOSS paper metadata
  • ○
    Scientific vocabulary similarity
    Low similarity (6.4%) to scientific vocabulary
Last synced: 10 months ago · JSON representation

Repository

Pruning framework and methods for Flux

Basic Info
  • Host: GitHub
  • Owner: darsnack
  • License: mit
  • Language: Julia
  • Default Branch: main
  • Size: 31.3 KB
Statistics
  • Stars: 27
  • Watchers: 2
  • Forks: 1
  • Open Issues: 1
  • Releases: 1
Created about 5 years ago · Last pushed about 3 years ago
Metadata Files
Readme License

README.md

FluxPrune

Build Status

FluxPrune.jl provides iterative pruning algorithms for Flux models. Pruning strategies can be unstructured or structured. Unstructured strategies operate on arrays, while structured strategies operate on layers.

Examples

Unstructured edge pruning

```julia using Flux, FluxPrune using MLUtils: flatten

m = Chain(Conv((3, 3), 3 => 16), Conv((3, 3), 16 => 32), flatten, Dense(512, 10))

prune all weights to 70% sparsity

m̄ = prune(LevelPrune(0.7), m)

prune all weights with magnitude lower than 0.5

m̄ = prune(ThresholdPrune(0.5), m)

prune each layer in a Chain at a different rate

(just uses broadcasting then re-Chains)

m̄ = prune([LevelPrune(0.4), LevelPrune(0.6), identity, LevelPrune(0.7)], m) ```

Structured channel pruning

```julia using Flux, FluxPrune using MLUtils: flatten

m = Chain(Conv((3, 3), 3 => 16), Conv((3, 3), 16 => 32), flatten, Dense(512, 10))

prune all conv layer channels to 30% sparsity

m̄ = prune(ChannelPrune(0.3), m) ```

Mixed pruning

```julia using Flux, FluxPrune using MLUtils: flatten

m = Chain(Conv((3, 3), 3 => 16), Conv((3, 3), 16 => 32), flatten, Dense(512, 10))

apply channel and edge pruning

m̄ = prune([ChannelPrune(0.3), ChannelPrune(0.4), identity, LevelPrune(0.8)], m) ```

Iterative pruning

Target pruning levels step-by-step. The first argument to iterativeprune (or the function block after the do statement) will finetune the model and return true to indicate moving onto the next stage, or false to indicate that finetune must be called again. ```julia using Flux, FluxPrune using MLUtils: flatten using Statistics: mean

features = rand(Float32, 8, 8, 3, 100); labels = Flux.onehotbatch(rand(0:9, 100), 0:9); data = (features, labels); loss(m, x, y) = Flux.Losses.mse(m(x), y) accuracy(m, data) = mean(Flux.onecold(m(data[1]), 0:9) .== Flux.onecold(data[2], 0:9)) target_accuracy = 0.08 # random data, so this is a low target

m = Chain(Conv((3, 3), 3 => 16), Conv((3, 3), 16 => 32), flatten, Dense(512, 10), softmax) opt_state = Flux.setup(Momentum(), m);

stages = [ [ChannelPrune(0.1), ChannelPrune(0.1), identity, LevelPrune(0.4), identity], [ChannelPrune(0.2), ChannelPrune(0.3), identity, LevelPrune(0.7), identity], [ChannelPrune(0.3), ChannelPrune(0.5), identity, LevelPrune(0.9), identity] ] m̄ = iterativeprune(stages, m) do m̄ for epoch in 1:10 Flux.train!(loss, m̄, [data], optstate) end return accuracy(m̄, data) > targetaccuracy end ```

Owner

  • Name: Kyle Daruwalla
  • Login: darsnack
  • Kind: user
  • Location: Cold Spring Harbor Lab, NY

NeuroAI scholar at CSHL

GitHub Events

Total
Last Year

Committers

Last synced: over 3 years ago

All Time
  • Total Commits: 29
  • Total Committers: 1
  • Avg Commits per committer: 29.0
  • Development Distribution Score (DDS): 0.0
Top Committers
Name Email Commits
Kyle Daruwalla d****a@w****u 29
Committer Domains (Top 20 + Academic)
wisc.edu: 1

Issues and Pull Requests

Last synced: over 1 year ago

All Time
  • Total issues: 4
  • Total pull requests: 6
  • Average time to close issues: about 7 hours
  • Average time to close pull requests: 6 months
  • Total issue authors: 3
  • Total pull request authors: 3
  • Average comments per issue: 1.25
  • Average comments per pull request: 0.17
  • Merged pull requests: 2
  • Bot issues: 0
  • Bot pull requests: 4
Past Year
  • Issues: 2
  • Pull requests: 0
  • Average time to close issues: about 15 hours
  • Average time to close pull requests: N/A
  • Issue authors: 1
  • Pull request authors: 0
  • Average comments per issue: 1.5
  • Average comments per pull request: 0
  • Merged pull requests: 0
  • Bot issues: 0
  • Bot pull requests: 0
View more stats
Top Authors
Issue Authors
  • vonpetersenn (2)
  • natema (1)
  • JuliaTagBot (1)
Pull Request Authors
  • github-actions[bot] (4)
  • darsnack (1)
  • natema (1)
Top Labels
Issue Labels
Pull Request Labels

Packages

  • Total packages: 1
  • Total downloads: unknown
  • Total dependent packages: 0
  • Total dependent repositories: 0
  • Total versions: 1
juliahub.com: FluxPrune

Pruning framework and methods for Flux

  • Versions: 1
  • Dependent Packages: 0
  • Dependent Repositories: 0
Rankings
Dependent repos count: 9.9%
Stargazers count: 22.0%
Average: 31.1%
Dependent packages count: 38.9%
Forks count: 53.5%
Last synced: 11 months ago

Dependencies

.github/workflows/CI.yml actions
  • actions/cache v1 composite
  • actions/checkout v2 composite
  • julia-actions/julia-buildpkg v1 composite
  • julia-actions/julia-runtest v1 composite
  • julia-actions/setup-julia v1 composite
.github/workflows/CompatHelper.yml actions
.github/workflows/TagBot.yml actions
  • JuliaRegistries/TagBot v1 composite