krishvishal / ResNet.jl

ResNet implementation in Julia w/ pretrained Imagenet weights

ResNet

Note this package requires Flux#master in order to be able to disable bias in convolutional layers, as there is no release with this ability yet.

Implementation of ResNet in Julia language.

Pretrained on ImageNet weights were ported from PyTorch, tested and confirmed to give identical results with the PyTorch's version.

Model	Weights
ResNet18	Download
ResNet34	Download
ResNet50	Download
ResNet101	Download
ResNet152	Download

How to load

Because currently Flux.jl defines only bias β and scale γ as trainable parameters for BatchNorm, you have to redefine trainable function for BatchNorm as follows.

Flux.trainable(bn::Flux.BatchNorm) = (bn.β, bn.γ, bn.μ, bn.σ²)

After that you can load weights using BSON.jl.

model = resnet(18)
path = "./resnet18-pretrained.bson"
@load path parameters
loadweights!(model, parameters)

Inference

Given image one can perform inference simply by calling model

image = ...
y = model(image)

or if you want to extract features

features = image |> model.entry |> model.encoder

or extract list of features

entry = image |> model.entry
features = [model.encoder[1](entry)]
for encoder in model.encoder[2:end]
    push!(features, encoder(features[end]))
end