[Discussion]: doctests, docstrings, documentation manual, and unclear internal API (for newcomers)

[Saransh-cpp]

Hello! I am opening this issue for further discussions on docstrings, doctests, making internal API visible (by prepending _), and updating the manual. Before I open a PR, I think it would be best to discuss the documentation here. I hope this is okay as I couldn't find a discussions tab in Flux's repository.

onhot.jl

onehot.jl misses doctests for the structs returned by onehotbatch and onehot. I am not sure if these structs are supposed to be user-facing? The docstring of the struct Embedding uses OneHotMatrix directly, but other than that, I couldn't find any other use cases. Should I add some doctests to these structs? I can also add them to the manual if they are supposed to be user-facing.

Additionally, right now if a user tries to access the documentation of OneHotVector and OneHotMatrix, they are presented with the documentation (with the signature type) of OneHotArray. After reading Flux's code, the reason becomes obvious, but I don't think every user will want to dive into Flux's code. Should these structs have their independent docstrings?

utils.jl

utils.jl has a few functions with no doctests, but it isn't clear if they are user-facing or not. From my previous PRs I have learned that the internal API doesn't need doctests, so should I prepend a _ to these functions, or should I add doctests to them (if they are user-facing)? The functions -

Not present in the manual -

• rng_from_array
• ones32
• zeros32
• create_bias

Present in the manual -

• rand32
• randn32
• throttle

functor.jl

The exact same case as that of utils.jl -

Present in the manual -

• testmode!
• trainmode!
• f32
• f64

Additionally, I think cpu and gpu are user-facing, but their APIs aren't documented in the manual.

optimisers.jl

Currently, none of the optimisers have doctests, and I am not sure how to add them without excessive repetition. For example, I came up with the following doctest for the Descent optimiser -

julia> f(x) = 4x + 2;

julia> x = hcat(0:5...); y = f.(x);

julia> model = Dense(1 => 1);

julia> opt = Descent(0.01);

julia> loss(x, y) = Flux.Losses.mse(model(x), y);

julia> ps = Flux.params(model)  # initialised parameters
Params([Float32[-1.1358223;;], Float32[0.0]])

julia> loss(x, y)  # loss with initial parameters
297.14438f0

julia> gs = gradient(ps) do
                  loss(x, y)
              end;

julia> Flux.Optimise.update!(opt, ps, gs) # the parameters will be updated, and the loss should ideally go down

julia> ps
Params([Float32[-0.09425485;;], Float32[0.29679117]])

julia> loss(x, y)
191.4279f0

This does look a bit extra, but I am not sure how else I can document it properly. My main concern is that this would have to be repeated for all the optimisers, if it is added. A workaround can be initializing the same variables only once -

"""
    Descent(η = 0.1)
...
\```jldoctest optimisers
julia> f(x) = 4x + 2;

julia> x = hcat(0:5...); y = f.(x);

julia> model = Dense(1 => 1);

julia> opt = Descent(0.01);

julia> loss(x, y) = Flux.Losses.mse(model(x), y);

julia> ps = Flux.params(model)  # initialised parameters
Params([Float32[-1.1358223;;], Float32[0.0]])

julia> loss(x, y)  # loss with initial parameters
297.14438f0

julia> gs = gradient(ps) do
                  loss(x, y)
              end;

julia> Flux.Optimise.update!(opt, ps, gs) # the parameters will be updated, and the loss should ideally go down

julia> ps
Params([Float32[-0.09425485;;], Float32[0.29679117]])

julia> loss(x, y)
191.4279f0
\```
"""
mutable struct Descent <: AbstractOptimiser
   ...
"""
    Momentum(η = 0.01, ρ = 0.9)
...
\```jldoctest optimisers
julia> opt = Momentum(0.03, 0.99);

julia> ps = Flux.params(model)
Params([Float32[-0.09425485;;], Float32[0.29679117]])

julia> loss(x, y)
191.4279f0

julia> gs = gradient(ps) do
                  loss(x, y)
       end;

julia> Flux.Optimise.update!(opt, ps, gs) # the parameters will be updated

julia> ps  # updated parameters
Params([Float32[2.4130664;;], Float32[1.013122]])

julia> loss(x, y)
31.88943f0
\```
"""
mutable struct Momentum <: AbstractOptimiser
   ...

I am not sure if I should proceed ahead with this.

Suggestions from everyone are welcome :)

[mcabbott]

• optimisers.jl we should kill, https://github.com/FluxML/Flux.jl/issues/1986 . And I don't think adding a complete worked example for every rule is a good idea.
• onhot.jl we should kill, https://github.com/FluxML/Flux.jl/issues/1544 . The package exists but isn't registered yet https://github.com/JuliaRegistries/General/pull/58049
• utils.jl : ones32, rand32 etc. have docstrings, I'm not sure it matters too much whether these have examples, they should probably be in the manual though. rng_from_array and create_bias are internal.
• functors.jl : cpu & gpu should surely be in the manual.

[Saransh-cpp]

Thank you for the suggestion! I will create a PR for updating the docstrings. I will leave Optimise.jl and onehot.jl as it is for now. I'll definitely pick up some documentation work in the new packages!