remove learnable parameters when bias=false

[CarloLucibello]

On current master, bias=false initialized the weights to zero, but the bias remains a learnable parameter. This PR fixes this problem and makes the interface consistent with Flux's initialization.

This PR also imposes stricter parametrization on layers' fields

This a breaking PR

[CarloLucibello]

this should be ready for review and merge

[yuehhua]

I am curious about assigning the element types of layer parameters. Shouldn't we provide a way for specifying the element types, such as T=Float32?

[CarloLucibello]

I am curious about assigning the element types of layer parameters. Shouldn't we provide a way for specifying the element types, such as T=Float32?

We discussed this at length in Flux. In the end the approach we took is the following:

• the default init method (glorot_uniform), returns Float32 arrays
• if one wants to use a different eltype, he can do so by passing e.g. init = (dims...) -> randn(Float16, dims...)
• the bias has the same eltype of the weights
• layers can also be initialized with specific weights: Dense(w, b)
• one can cast a whole model to the desired floating point type: model |> fp32