Reference

IterLayers(model::Model)

Pre-order (parents before children) iterator over all wrapped layers.

Example

> model = Model(Chain(Conv((3, 3), 3 => 16), BatchNorm(16)))
> collect(IterModels(model))

[Chain{...}, Conv{...}, BatchNorm{...}]

IterModels(model::Model)

Pre-order (parents before children) iterator over all Models. If you want to iterate over the wrapped layers instead, use IterLayers.

Example

> model = Model(Chain(Conv((3, 3), 3 => 16), BatchNorm(16)))
> collect(IterModels(model))

[Model{Chain{...}}, Model{Conv{...}}, Model{BatchNorm{...}}]

Model(layer)

Wrapper for Flux layer that recursively wraps children layers.

Use model.layer to access wrapped layer

Use children(model) to access children wrappers

Hook(layer, [forward, backward, state = Dict()])

Wraps a Flux layer, and holds a Dict with state (Hook.state)

After the forward pass of the wrapped layer calls Hook.forward(state, layeroutput) After the gradient calculation of the wrapped layer calls Hook.backward(state, layergradient)

For adding hooks to existing models, see addhook

addhook(model; [forward, backward, filter_fn])

Add a Hook with callbacks forward and backward to all layers of model where filter_fn(layer) == true.

See Hook

mapmodel(f, model::M, [constructor])

Applies f to every layer, resulting in a new model.

If your custom type does not have a constructor of that form, like e.g. Flux.Chain, you will have to extend this function with a signature like

`mapmodel(f, model::MyType, (fields) -> ...)`

where the last argument results in a MyType(fields...)

paramsdict(model::Model) -> Dict{Symbol, Array}

Return a Dict with mapping parameter name => parameter value.

Useful for inspecting the distribution of parameters.

gradientsdict(model::Model, gs::Grads) -> Dict{Symbol, Array}

Return a Dict with mapping parameter name => gradient value

Useful for inspecting the distribution of gradients.

printmodel(model)

Print model as a tree

Initialization(layertype, field, initializer)

Describes an initialization for field field of all layers of type layertype to be initialized with initializer.

Use initmodel! to apply to a model.

Example

Initialization(Conv, :bias, init_zeros) -> initialize the bias of all Conv layers with zeros.

init_normal_uniform(T, dims...; gain=sqrt(2))

Kaiming normal initializer

init_kaiming_uniform(T, dims...; gain=sqrt(2))

Kaiming uniform initializer

init!(a, initializer)

Initialize an array a with initializer.

Example

layer = Conv((3, 3), 3 => 16)
init!(layer.weight, init_kaiming_normal)

initmodel!(model, inits::Vector{Initialization})

Applies inits to model.

See Initialization.

Example

model = Chain(Dense(10, 10), Dense(10, 10))
initmodel!(model, [
    Initialization(Dense, :W, init_kaiming_normal),
    Initialization(Dense, :bias, init_zeros)
])