ImageNetDataset.jl

Data loader for the ImageNet 2012 Classification Dataset (ILSVRC 2012-2017) in Julia.

Installation

The ImageNet dataset can be downloaded at image-net.org after signing up and accepting the terms of access. It is therefore required that you download this dataset manually.

Installation instructions can be found in the documentation.

Afterwards, add this package via

julia> ]add https://github.com/adrhill/ImageNetDataset.jl

Examples

By default, the ImageNet dataset will be loaded with the CenterCropNormalize transformation.

This uses JpegTurbo.jl to open the image and applies a center-cropping view to (224, 224) resolution to it. Afterwards, the image is normalized over color channels using normalization constants which are compatible with most pretrained models from Metalhead.jl. The output is in WHC[N] format (width, height, color channels, batchsize).

using ImageNetDataset

dataset = ImageNet(:val)            # load validation set
X, y = dataset[1:5]                 # load features and targets

convert2image(dataset, X)           # convert features back to images

class(dataset, y)                   # obtain class names

Preprocessing

The dataset can also be loaded in a custom size with custom normalization parameters by configuring the preprocessing transformations. ImageNetDataset.jl currently provides CenterCropNormalize and RandomCropNormalize:

output_size = (224, 224)
mean = (0.485f0, 0.456f0, 0.406f0)
std  = (0.229f0, 0.224f0, 0.225f0)

tfm = CenterCropNormalize(; output_size, mean, std)

dataset = ImageNet(:val; transform=tfm)

Custom transformations can be implemented by extending AbstractTransformation.

To apply a transformation outside of the ImageNet dataset, e.g. to preprocess a single image add a given path, run

transform(tfm, path)

DataAugmentation.jl compatibility

Alternatively, ImageNetDataset is compatible with transformations from DataAugmentation.jl:

using ImageNetDataset, DataAugmentation

tfm = CenterResizeCrop((224, 224)) |> ImageToTensor() |> Normalize(mean, std)

dataset = ImageNet(:val; transform=tfm)

Warning

Note that DataAugmentation.jl returns features in HWC[N] format instead of WHC[N]. Transformations from DataAugmentation.jl are also slightly less performant and not compatible with convert2image.

Related packages

• MLDatasets.jl: Utility package for accessing common Machine Learning datasets in Julia

Note

This repository is based on MLDatasets.jl PR #146 and mirrors the MLDatasets API.

Copyright (c) 2015 Hiroyuki Shindo and contributors.