Real World Examples

Easy local directories

I setup all my science projects using DrWatson's suggested setup, using initialize_project. Then, every file in every project has a start that looks like this:

using DrWatson
quickactivate(@__DIR__, "MagneticBilliardsLyapunovs")
using DynamicalBilliards, PyPlot, LinearAlgebra


In all projects I save data/plots using datadir/plotdir:

@tagsave(datadir("mushrooms", "Λ_N=$N.bson"), (@dict Λ Λσ ws hs description))

The advantage of this approach is that it will always work regardless of if I move the specific file to a different subfolder (which is very often necessary) or whether I move the entire project folder somewhere else! Please be sure you have understood the caveat of using quickactivate!

Here is an example from another project. You will notice that another advantage is that I can use identical syntax to access the data or source folders even though I have different projects!

using DrWatson
quickactivate "EmbeddingResearch"
using Parameters
using TimeseriesPrediction, LinearAlgebra, Statistics

include(srcdir("systems", "barkley.jl"))

# stuff...

save(datadir("sim", "barkley", "astonishing_results.bson"), data)

Making your project a usable module

For some projects, it is often the case that some packages and files from the source folder are loaded at the beginning of every file of the project. For example, I have a project that I know that for any script I will write, the first five lines will be:

using DrWatson
@quickactivate "AlbedoProperties"
using Dates, Statistics, NCDatasets

It would be quite convenient to group all of these commands into one file and instead load that file, for example do include(srcdir("everything.jl")) and all commands are in there.

We can do even better though! Because of the way Julia handles project and module paths, it is in fact possible to transform the currently active project into a usable module. If one defines inside the src folder a file AlbedoProperties.jl and in that file define a module AlbedoProperties (notice that these names must match exactly the project name), then upon doing using AlbedoProperties Julia will in fact just bring this module into scope.

So what I end up doing (for some projects where this makes sense) is creating the aforementioned file and putting inside things like

module AlbedoProperties

using Reexport
@reexport using Dates, Statistics
using NCDatasets: NCDataset, dimnames, NCDatasets
export NCDataset, dimnames
include("core.jl") # this file now also has export statements


and then the header of all my files is transformed to

using DrWatson
@quickactivate :AlbedoProperties

which takes advantage of @quickactivate's feature to essentially combine the commands @quickactivate "AlbedoProperties" and using AlbedoProperties into one.

If you intend to share your project with a non-DrWatson user, you should consider the verbose syntax instead, as the above syntax is not really clear for someone that doesn't know what @quickactivate does.

savename and tagging

The combination of using savename and tagsave makes it easy and fast to save output in a way that is consistent, robust and reproducible. Here is an example from a project:

using DrWatson
quickactivate(@__DIR__, "EmbeddingResearch")
using TimeseriesPrediction, LinearAlgebra, Statistics
include(srcdir("systems", "barkley.jl"))

ΔTs = [1.0, 0.5, 0.1] # resolution of the saved data
Ns = [50, 150] # spatial extent
for N ∈ Ns, ΔT ∈ ΔTs
    T = 10050 # we can offset up to 1000 units
    every = round(Int, ΔT/barkley_Δt)
    seed = 1111

    simulation = @ntuple T N ΔT seed
    U, V = barkley(T, N, every; seed = seed)

        datadir("sim", "bk", savename(simulation, "bson")),
        @dict U V simulation

This saves files that look like:


and each file is a dictionary that has my data fields: :U, :V, :simulation, but also :gitcommit, :script. When I read this file I know exactly what was the source code that produced it (provided that I am not sloppy and commit code changes regularly :P).

Customizing savename

Here is a simple example for customizing savename. We are using a common struct Experiment across different experiments with cats and mice. In this example we are also using Parameters.jl for a convenient default constructor.

We first define the relevant types.

using DrWatson, Dates
using Base: @kwdef

# Define a type hierarchy we use at experiments
abstract type Species end
struct Mouse <: Species end
struct Cat <: Species end

# @with_kw comes from Parameters.jl
@kwdef struct Experiment{S<:Species}
    n::Int = 50
    c::Float64 = 10.0
    x::Float64 = 0.2
    date::Date = Date(
    species::S = Mouse()
    scientist::String = "George"

e1 = Experiment()
e2 = Experiment(species = Cat())
Main.ex-customizing.Experiment{Main.ex-customizing.Cat}(50, 10.0, 0.2, Date("2020-11-11"), Main.ex-customizing.Cat(), "George")

For analyzing our experiments we need information about the species used, and to use multiple dispatch latter on we decided to make this information associated with a Type. This is why we defined Species.

Now, we want to customize savename. We start by extending DrWatson.default_prefix:

DrWatson.default_prefix(e::Experiment) = "Experiment_"*string(


However this is not good enough for us, as the information about the species is not contained in savename and also the date information is duplicated. We have to extend DrWatson.default_allowed to specify which data types should be extended in savename:

DrWatson.default_allowed(::Experiment) = (Real, String, Species)


To make printing of Species better we can extend Base.string, which is what DrWatson uses internally in savename to display values.

Base.string(::Mouse) = "mouse"
Base.string(::Cat) = "cat"

Lastly, let's say that the information of which scientist performed the experiment is not really relevant for savename. We can extend the last method, DrWatson.allaccess:

DrWatson.allaccess(::Experiment) = (:n, :c, :x, :species)

so that only those four fields will be used (notice that the date field is already used in default_prefix). We finally have:

println( savename(e1) )
println( savename(e2) )

savename and nested containers

In the case of user-defined structs and projects of significant complexity, it is often necessary that your "main" container has other containers as subfields. savename can adapt to these situations as well. Consider the following example, where I need a core struct that represents a spatio temporal system, and its simulation:

struct SpatioTemporalSystem
    model::String # system codeword
    N        # Integer or Tuple of integers: spatial extent
    Δt::Real # sampling time in real time units
    p        # parameters. nothing or Dict{Symbol}
const STS = SpatioTemporalSystem

struct SpatioTemporalTimeseries
    T::Int       # total frame amount
    ic           # initial condition (matrix, string, seed)
    fields::Dict # resulting timeseries, dictionary of string to vector
const STT = SpatioTemporalTimeseries

For my use case, p can be nothing or it can be a dictionary itself, containing the possible parameters the spatiotemporal systems can have. To adapt savename to situations like this, we use the functionality surrounding DrWatson.default_expand.

Expanding the necessary methods allows me to do:

DrWatson.allaccess(c::STS) = (:N, :Δt, :p)
DrWatson.default_prefix(c::STS) = c.model
DrWatson.default_allowed(c::STS) = (Real, Tuple, Dict, String)
DrWatson.default_expand(c::STS) = ["p"]

bk = STS("barkley", 60, 0.1, nothing)

and when I do want to use different parameters than the default:

a = 0.3; b = 0.5
bk = STS("barkley", 60, 0.1, @dict a b)

Expanding to the second struct is also fine:

DrWatson.default_prefix(c::STT) = savename(c.sts)
stt = STT(bk, 1000, nothing, Dict("U"=>rand(100), "V"=>rand(100)))

Stopping "Did I run this?"

It can become very tedious to have a piece of code that you may or may not have run and may or may not have saved the produced data. You then constantly ask yourself "Did I run this?". Typically one uses isfile and an if clause to either load a file or run some code. Especially in the cases where the code takes only a couple of minutes to finish you are left in a dilemma "Is it even worth it to save?".

This is the dilemma that produce_or_load resolves. You can wrap your code in a function and then produce_or_load will take care of the rest for you! I found it especially useful in scripts that generate figures for a publication.

Here is an example; originally I had this piece of code:

HTEST = 0.1:0.1:2.0
WS = [0.5, 1.0, 1.5]
N = 10000; T = 10000.0

toypar_h = [[] for l in HS]
for (wi, w) in enumerate(WS)
    println("w = $w")
    for h in HTEST
        toyp = toyparameters(h, w, N, T)
        push!(toypar_h[wi], toyp)

that was taking some minutes to run. To use the function produce_or_load I first have to wrap this code in a high level function like so:

function g(d)
    HTEST = 0.1:0.1:2.0
    WS = [0.5, 1.0, 1.5]
    @unpack N, T = d
    toypar_h = [[] for l in HS]

    for (wi, w) in enumerate(WS)
        println("w = $w")
        for h in HTEST
            toyp = toyparameters(h, w, N, T)
            push!(toypar_h[wi], toyp)
    return @dict toypar_h

N = 2000; T = 2000.0
file = produce_or_load(
    datadir("mushrooms", "toy"), # path
    @dict(N, T), # container
    g, # function
    prefix = "fig5_toyparams" # prefix for savename
@unpack toypar_h = file

Now, every time I run this code block the function tests automatically whether the file exists. Only if it does not, then the code is run while the new result is saved to ensure I won't have to run it again.

The extra step is that I have to extract the useful data I need from the container file. Thankfully the @unpack macro from Parameters.jl makes this super easy.

Preparing & running jobs

Preparing the dictionaries

Here is a shortened script from a project that uses dict_list:

using DrWatson

general_args = Dict(
    "model" => ["barkley", "kuramoto"],
    "noise" => 0.075,
    "noisy_training" => [true, false],
    "N" => [100],
    "embedding" => [ #(γ, τ, r, c)
    (4, 5, 1, 0.34), (4, 6, 1, 0.28)]
Dict{String,Any} with 5 entries:
  "embedding"      => [(4, 5, 1, 0.34), (4, 6, 1, 0.28)]
  "model"          => ["barkley", "kuramoto"]
  "N"              => [100]
  "noise"          => 0.075
  "noisy_training" => Bool[1, 0]
dicts = dict_list(general_args)
println("Total dictionaries made: ", length(dicts))
Dict{String,Any} with 5 entries:
  "embedding"      => (4, 5, 1, 0.34)
  "model"          => "barkley"
  "N"              => 100
  "noise"          => 0.075
  "noisy_training" => true

Now, how you use these dictionaries is up to you. Typically each dictionary is given to a main-like Julia function which extracts the necessary data and calls the necessary functions.

Let's say I have written a function that takes in one of these dictionaries and saves the file somewhere locally:

function cross_estimation(data)
    γ, τ, r, c = data["embedding"]
    N = data["N"]
    # add fake results:
    data["x"] = rand()
    data["error"] = rand(10)
    # Save data:
    prefix = datadir("results", data["model"])
    get(data, "noisy_training", false) && (prefix *= "_noisy")
    get(data, "symmetric_training", false) && (prefix *= "_symmetric")
    sname = savename((@dict γ τ r c N), "bson")
    mkpath(datadir("results", data["model"]))
    save(datadir("results", data["model"], sname), data)
    return true
cross_estimation (generic function with 1 method)

Using map and pmap

One way to run many simulations is with map (identical process for using pmap). To run all my simulations I just do:

dicts = dict_list(general_args)
map(cross_estimation, dicts) # or pmap

# load one of the files to be sure everything is ok:
filename = readdir(datadir("results", "barkley"))[1]
file = load(datadir("results", "barkley", filename))
Dict{String,Any} with 7 entries:
  "embedding"      => (4, 6, 1, 0.28)
  "model"          => "barkley"
  "N"              => 100
  "x"              => 0.617931
  "error"          => [0.393519, 0.160122, 0.938565, 0.585471, 0.532048, 0.7819…
  "noise"          => 0.075
  "noisy_training" => false

Using a Serial Cluster

In case that I can't store the results of dict_list in memory, I have to change my approach and load them from disk. This is easy with the function tmpsave.

Instead of using Julia to run all jobs from one process with map/pmap one can use Julia to submit many jobs to a cluster que. For our example above, the Julia program that does this would look like this:

dicts = dict_list(general_args)
res = tmpsave(dicts)
for r in res
    submit = `qsub -q queuename julia runjob.jl $r`

Now the file runjob.jl would have contents that look like:

f = ARGS[1]
dict = load(projectdir("_research", "tmp", f))

i.e. it just loads the dict and straightforwardly uses the "main" function cross_estimation. Remember to routinely clear the tmp directory! You could do that by e.g. adding a line rm(projectdir("_research", "tmp", f) at the end of the runjob.jl script.

Listing completed runs

Continuing from the Preparing & running jobs section, we now want to collect the results of all these simulations into a single DataFrame. We will do that with the function collect_results!.

It is quite simple actually! But because we don't want to include the error, we have to black-list it:

using DataFrames # this is necessary to access collect_results!
bl = ["error"]
res = collect_results!(datadir("results"); black_list = bl, subfolders = true)

4 rows × 7 columns

1(4, 6, 1, 0.28)barkley1000.6179310.0750/home/travis/build/JuliaDynamics/DrWatson.jl/docs/data/results/barkley/N=100_c=0.28_r=1_γ=4_τ=6.bson
2(4, 5, 1, 0.34)barkley1000.7104160.0750/home/travis/build/JuliaDynamics/DrWatson.jl/docs/data/results/barkley/N=100_c=0.34_r=1_γ=4_τ=5.bson
3(4, 6, 1, 0.28)kuramoto1000.9530360.0750/home/travis/build/JuliaDynamics/DrWatson.jl/docs/data/results/kuramoto/N=100_c=0.28_r=1_γ=4_τ=6.bson
4(4, 5, 1, 0.34)kuramoto1000.292160.0750/home/travis/build/JuliaDynamics/DrWatson.jl/docs/data/results/kuramoto/N=100_c=0.34_r=1_γ=4_τ=5.bson

We can take also advantage of the basic processing functionality of collect_results! to use the excluded "error" column, replacing it with its average value:

using Statistics: mean
special_list = [:avrg_error => data -> mean(data["error"])]
res = collect_results(
      black_list = bl,
      special_list = special_list,
      subfolders = true

select!(res, Not(:path)) # don't show path this time

4 rows × 7 columns

1(4, 6, 1, 0.28)barkley1000.6179310.07500.469971
2(4, 5, 1, 0.34)barkley1000.7104160.07500.49904
3(4, 6, 1, 0.28)kuramoto1000.9530360.07500.481486
4(4, 5, 1, 0.34)kuramoto1000.292160.07500.382188

As you see here we used collect_results instead of the in-place version, since there already exists a DataFrame with all results processed (and thus everything would be skipped).

Adapting to new data/parameters

We once again continue from the above example. But we now need to run some new simulations with some new parameters that do not exist in the old simulations... Well, DrWatson says "no problem!" :)

Let's save these new parameters in a different subfolder, to have a neatly organized project:

general_args_new = Dict(
    "model" => ["bocf"],
    "symmetry" => "radial",
    "symmetric_training" => [true, false],
    "N" => [100],
    "embedding" => [ #(γ, τ, r, c)
    (4, 5, 1, 0.34), (4, 6, 1, 0.28)]
Dict{String,Any} with 5 entries:
  "symmetry"           => "radial"
  "model"              => ["bocf"]
  "symmetric_training" => Bool[1, 0]
  "N"                  => [100]
  "embedding"          => [(4, 5, 1, 0.34), (4, 6, 1, 0.28)]

As you can see, there here there are two parameters not existing in previous simulations, namely "symmetry", "symmetric_training". In addition, the parameters "noise", "noisy_training" that existed in the previous simulations do not exist in the current one.

No problem though, let's run the new simulations:

dicts = dict_list(general_args_new)
map(cross_estimation, dicts)

# load one of the files to be sure everything is ok:
filename = readdir(datadir("results", "bocf"))[1]
file = load(datadir("results", "bocf", filename))
Dict{String,Any} with 7 entries:
  "symmetric_training" => false
  "model"              => "bocf"
  "N"                  => 100
  "embedding"          => (4, 6, 1, 0.28)
  "symmetry"           => "radial"
  "x"                  => 0.444977
  "error"              => [0.648368, 0.990382, 0.060906, 0.925573, 0.830422, 0.…

Alright, now we want to add these new runs to our existing dataframe that has collected all previous results. This is straight-forward:

res = collect_results!(datadir("results"); black_list = bl, subfolders = true)

select!(res, Not(:path)) # don't show path this time

6 rows × 8 columns

1(4, 6, 1, 0.28)barkley1000.6179310.0750missingmissing
2(4, 5, 1, 0.34)barkley1000.7104160.0750missingmissing
3(4, 6, 1, 0.28)kuramoto1000.9530360.0750missingmissing
4(4, 5, 1, 0.34)kuramoto1000.292160.0750missingmissing
5(4, 6, 1, 0.28)bocf1000.444977missingmissing0radial
6(4, 5, 1, 0.34)bocf1000.943005missingmissing0radial

All missing entries were adjusted automatically :)

Defining parameter sets with restrictions

As already demonstrated in the examples above, for functions where the set of input parameters is the same for each simulation run, a basic dictionary can be used to define these parameters. However, often some of the parameters or values should only be considered if another parameter is also included in the set or has a specific value. The macro @onlyif allows to place such restrictions on values and parameters. The following dictionary defines values and parameters for a genetic algorithm:

ga_parameters = Dict(
    :population_size => [20,50,100],
    :selection => ["roulette-selection", "SUS", "tournament-selection", "linear ranking"],
    :fitness_scaling => @onlyif(:selection in ("SUS", "roulette-selection"), collect(1.0:20.0)),
    :tournamet_size => @onlyif(:selection == "tournament-selection", collect(2:10)),
    :chromosome => [:A, @onlyif(begin
        size_constr = (:population_size <= 50)
        select_constr = (:selection != "SUS")
        size_constr && select_constr
    end, :B)])
Dict{Symbol,Array{T,1} where T} with 5 entries:
  :selection       => ["roulette-selection", "SUS", "tournament-selection", "li…
  :population_size => [20, 50, 100]
  :chromosome      => Any[:A, DependentParameter{Symbol}(:B, #7)]
  :fitness_scaling => DrWatson.DependentParameter{Float64}[DependentParameter{F…
  :tournamet_size  => DrWatson.DependentParameter{Int64}[DependentParameter{Int…
dicts = dict_list(ga_parameters)
Dict{Symbol,Any} with 4 entries:
  :selection       => "roulette-selection"
  :population_size => 20
  :chromosome      => :A
  :fitness_scaling => 1.0

The parameter restriction for the chromosome type shows that one can use arbitrary Julia expressions that return true or false. In this case, first the conditions for the population size and for the selection method are evaluated and stored. The expression then only returns true, if both conditions are met, thus restricting the usage of chromosome type :B.

Advanced Usage of collect_results

At some point in your work you may want to run a single function that returns multiple fields that you want to include in your results DataFrame. Depending on the problem you are trying to solve it may just make more sense to use a single function that extracts most or all of the meta-data. For this case DrWatson has another syntax available. Let us, for the sake of simplicity, assume that your data files contain a very long array of numbers called "manynumbers" and the information that you care about are the three largest values.

One way to implement this would be to write

special_list = [
    :first  => data -> sort(data["manynumbers"])[1],
    :second => data -> sort(data["manynumbers"])[2],
    :third  => data -> sort(data["manynumbers"])[3],

which makes very obvious that there should be a better way to do this. There is no point in sorting the very long vector three times. A better thing to do is the following

function largestthree(data)
    sorted = sort(data["manynumbers"])
    return [:first  => sorted[1],
            :second => sorted[2],
            :third  => sorted[3]]

special_list = [largestthree,]