See Documentation for more details.
Clone repository locally, install Julia v1.7.x. In the ActionRNN folder.
Start the julia repl with julia
julia> ]add Revise, Plots
julia> ]activate .
julia> ]instantiatejulia> using Revise; includet("experiment/ringworld_er.jl")
julia> ret = RingWorldRNNExperiment.main_experiment(; progress=true)This should run a Ring World experiment with the action RNN. This return a dictionary containing predictions and errors. You should easily be able to analyze the error and see relatively good performance here.
To analyze the data:
julia> using Statistics
julia> rmse = sqrt.(mean(ret["err"].^2; dims=2))
julia> mean(rmse) # You should get 0.006968971...To plot:
julia> using Plots
julia> gr()
julia> plot(rmse) # this will plot all the data points, and will be noisy
julia> plot(mean(reshape(rmse, 1000, 300); dims=1)') #this will plot a windowed average of pointsAdding a consistency test requires you know which experiment you are targeting, what cell, and the various arguments you want. For example. There is some magic macros in the ActionRNNsTests to make this simpler. Bellow is the easiest way to find the test for an experiment with an example in ringworld_er. This workflow is still WIP, but this should get you all you need to add new cells to the files in tests/consistency.
push!(LOAD_PATH, "../ActionRNNs.jl/test/")
using ReTest
import ActionRNNsTests: ActionRNNsTests, Consistency.@run_experiment
begin
ret = @run_experiment RingWorldERExperiment "FacMARNN" Consistency.RINGWORLD_ER_BASE_CONFIG factors=3
sum(ret["err"])
endWe use ReTest to run all our tests. This allows for testing certain experiments or cells across all experiments.
push!(LOAD_PATH, "../ActionRNNs.jl/test/")
using ReTest; import ActionRNNsTests
retest() # Run all tests
retest("AARNN") # Run all AARNN tests
retest("Ringworld") # Run all Ring World experiments
retest("Fac") # run all factored tests (including FacMA* and FacTucMA*)