Computer vision models for Flux
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
datasets Update README.md Feb 9, 2019
src Update utils.jl Feb 9, 2019
test Refactor tests a bit Sep 19, 2018
.codecov.yml Metalhead.jl generated files. Feb 15, 2018
.gitignore move to BSON Feb 27, 2018
.travis.yml Julia 1.0 updates Aug 26, 2018
LICENSE.md Metalhead.jl generated files. Feb 15, 2018
README.md Merge pull request #35 from Sklan/master Feb 11, 2019
REQUIRE bump julia version Sep 13, 2018
appveyor.yml Metalhead.jl generated files. Feb 15, 2018

README.md

Metalhead

Build Status Coverage

Pkg.add("Metalhead")

This package provides computer vision models that run on top of the Flux machine learning library.

IJulia Screenshot

Each model (like VGG19) is a Flux layer, so you can do anything you would normally do with a model; like moving it to the GPU, training or freezing components, and extending it to carry out other tasks (such as neural style transfer).

# Run with dummy image data
julia> x = rand(Float32, 224, 224, 3, 1)
224×224×3×1 Array{Float32,4}:
[:, :, 1, 1] =
 0.353337   0.252493    0.444695   0.7671930.107599   0.424298   0.218889    0.377959
 0.247294   0.039822    0.829367   0.832303       0.582103   0.359319   0.259342    0.12293
  ⋮

julia> vgg(x)
1000×1 Array{Float32,2}:
 0.000851723
 0.00079913# See the underlying model structure
julia> vgg.layers
Chain(Conv2D((3, 3), 3=>64, NNlib.relu), Conv2D((3, 3), 64=>64, NNlib.relu), Metalhead.#3, Conv2D((3, 3), 64=>128, NNlib.relu), Conv2D((3, 3), 128=>128, NNlib.relu), Metalhead.#4, Conv2D((3, 3), 128=>256, NNlib.relu), Conv2D((3, 3), 256=>256, NNlib.relu), Conv2D((3, 3), 256=>256, NNlib.relu), Conv2D((3, 3), 256=>256, NNlib.relu), Metalhead.#5, Conv2D((3, 3), 256=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Metalhead.#6, Conv2D((3, 3), 512=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Conv2D((3, 3), 512=>512, NNlib.relu), Metalhead.#7, Metalhead.#8, Dense(25088, 4096, NNlib.relu), Flux.Dropout{Float32}(0.5f0, false), Dense(4096, 4096, NNlib.relu), Flux.Dropout{Float32}(0.5f0, false), Dense(4096, 1000), NNlib.softmax)

# Run the model up to the last convolution/pooling layer
julia> vgg.layers[1:21](x)
7×7×512×1 Array{Float32,4}:
[:, :, 1, 1] =
 0.657502  0.598338  0.594517  0.594425  0.594522  0.597183  0.59534
 0.663341  0.600874  0.596379  0.596292  0.596385  0.598204  0.590494

Working with common datasets

Metalhead includes support for working with several common object recognition datasets. The datasets() function will attempt to auto-detect any common dataset placed in the datasets/. The Metalhead.download function can be used to download these datasets (where such automatic download is possible - for other data sets, see datasets/README.md), e.g.:

Metalhead.download(CIFAR10)

Once a dataset is loaded, it's training, validation, and test images are available using the trainimgs, valimgs, and testimgs functions. E.g.

julia> valimgs(dataset(ImageNet))[rand(1:50000, 10)]

will fetch 10 random validation images from the ImageNet data set.

Inline Images at the REPL

If you are using OS X, it is recommended that you use iTerm2 and install the TerminalExtensions.jl package. This will allow you to see inference results as well as the corresponding images directly in your terminal:

REPL Screenshot