A companion tool for the AnimalAI.
Specify task templates and procedurally generate all possible variations.
pip install git+[https://github.com/wschella/procgen-companion.git](https://github.com/Kinds-of-Intelligence-CFI/procgen-companion.git)
or if you already installed it, maybe safer to like this:
pip install --upgrade --force-reinstall git+https://github.com/wschella/procgen-companion.git
If you use this software, please cite it as below.
@software{Schellaert_ProcGen_2023,
author = {Schellaert, Wout},
title = {{AnimalAI Procedural Generation Companion}},
url = {https://github.com/wschella/procgen-companion},
version = {0.1.0},
month = {5},
year = {2023}
}
$ procgen --help
usage: procgen-companion [-h] {sample,gen,sample-bulk,gen-bulk,count-bulk} ...
Procedurally generate variations of an AnimalAI based on a task template.
positional arguments:
{sample,gen,sample-bulk,gen-bulk,count-bulk}
sample Sample variations of a task definition.
gen Generate variations of a task definition.
sample-bulk Sample variations of multiple task definitions.
gen-bulk Generate variations of multiple task definitions.
count-bulk Count the number of variations of multiple task definitions.
A procgen template is structured like an AnimalAI .yaml
file, but one can make use of certain special tags like !ProcList
, !ProcColor
, !ProcIf
that define the allowed variations of a field's value.
Run procgen gen template.yaml
...
... and the procgen-companion tool will generate all possible combinations of these fields, and output a .yaml
file for each variation in path/to/template_variations/
, which are directly usable by the AnimalAI environment.
The example below is not a valid AnimalAI config, but that doesn't matter right now.
# example.template.yaml
field1: !ProcList [0, 120]
field2: !ProcList [90, 180]
If we run procgen example.template.yaml
, four new files will be created:
# example_variations/example_1.yaml
field1: 0
field2: 90
# example_variations/example_2.yaml
field1: 0
field2: 180
# example_variations/example_3.yaml
field1: 120
field2: 90
# example_variations/example_4.yaml
field1: 120
field2: 180
The template used will also be copied example_variations/template.yaml
.
Every valid AnimalAI .yaml file is also a valid procgen-companion template. All the fields you don't want to procedurally generate will be ignored and copied.
It is also possible to take a random sample instead of generating all possible combinations, or to generate in bulk. See procgen --help
or ``procgen [subcommand] --help` for more info.
All tags used in procgen-companion start with !Proc
.
With !ProcList
, you define all the possible options a value can take. The items of the list can be any kind of valid yaml, i.e. lists, tags, mappings (see below).
But !ProcList
does not support nested !Proc
-tags, i.e. the items of the list can not be !Proc
-tags.
# Define three possible options for a scalar field value.
!ProcList [0, 90, 180]
# Define three possible options for a mapping
!ProcList [
!Vector3 { x: 10, y: 0, z: 1 },
!Vector3 { x: 20, y: 1, z: 5 },
!Vector3 { x: 20, y: 1, z: 10 },
]
# Exactly the same as above, just with different notation.
!ProcList
- !Vector3 { x: 10, y: 0, z: 1 }
- !Vector3 { x: 20, y: 1, z: 5 }
- !Vector3 { x: 20, y: 1, z: 10 }
You can also perfectly do this instead:
!Vector3 { x: !ProcList [2, 4, 6], y: 1, z: !ProcList [5, 10, 15] }
This defines 3x3=9 possible variations: !Vector3 {x: 2, y: 1, z: 5}
, !Vector3 {x: 2, y: 1, z: 10}
, !Vector3 {x: 2, y: 1, z: 15}
, and the same for x: 4
and x: 6
.
This will generate a new color each time, until 10 different ones have been picked. The colors will be picked from a fixed list of colors. If the amount specified here is larger than the list of colors we use, we will throw an error so you can fix it or ask us to create more colors.
!ProcColor 5
A common use case is varying all dimensions of a Vec3 at the same time, e.g. scaling the vector with a single scalar number.
The base !Vector3
will not be included in the possible options. If you want that, just include 1
as a scale factor.
!ProcVector3Scaled
base: !Vector3 { x: 2, y: 1, z: 1 }
scales: [1, 2, 3, 4]
The base
is an optional argument. If it is not present, the unit vector will be assumed, i.e. !Vector3 { x: 1, y: 1, z: 1}
. The following is thus also valid:
!ProcVector3Scaled
scales: [1, 2, 3, 4]
You can use !ProcRepeatChoice
when for example three walls need to vary their color, but it must always be the same color, i.e. they must vary at the same time. This tag will produce a list of values, each of them identical.
The value
field can contain any valid yaml including (nested) !Proc
-tags, as that is what is designed for. Any choice made in the !Proc
tags will be copied.
It is allowed for the value
field to not contain any !Proc
tags. Everything works just the same.
colors: !ProcRepeatChoice
amount: 3 # The amount of repetitions, i.e. the number of elements in the resulting list.
value: !ProcColor 5 # The amount of different values a list item can take.
# One possible realization of the possible 5 would like like this:
colors:
- !RGB {r: 128, g: 128, b: 128 }
- !RGB {r: 128, g: 128, b: 128 }
- !RGB {r: 128, g: 128, b: 128 }
You can use !ProcRestrictCombinations
when you know some part of the yaml file is responsible for many different possibilities, but you want to limit the total, without restricting any of the individual choices.
An example is when seven walls need to vary their color, you don't want it to be the same color, and you only want 25 different combinations in total.
This works by uniformly random sampling all the !Proc
-tags nested inside.
The item
field does not need to be a list as is the case right now. It can be anything.
!ProcRestrictCombinations
amount: 25 # Number of possible values this tag will generate.
item:
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
- !ProcColor 10
!ProcIf
value: [agent.positions.0.x, agents.positions.0.z]
cases: [[1, 5], [1, 7], [3, !R [10, 20]]]
then: [90, 120, 150]
default: 0
Depending on the the values the variables take, different values are generated by the !ProcIf
tag. A !ProcIf
tag is reactive. It does not increase the number of combinations, only sets the values of some fields in an existing combination.
Conceptually, if the variables in the value
field are equal to the numbers defined in the cases
field (or within a range !R
) then the corresponding value in the then
field will be selected. If none of the cases match the value of the default
field is taken. If your cases overlap by accident, the first matching case will be taken.
Here (first case): if x=1 and z=5, then we take 90.
Instead of a single number, you can also specify a range with !R
.
Here (last case): if x=3 and z between 10 and 20 (inclusive), then we take 150.
- The
value
field can be a string, or a list of strings. Multiple strings indicate there multiple variables we care about. Each string is a reference to a specific field in the generated yaml file (see below). - The
cases
field is always list. Ifvalue
is a list,cases
must be a list of lists, and an inner list needs to match the length ofvalue
. The outer list can be any length, each element represent a "case". It can be of length 1 (representing a single "if", withdefault
acting as the "else".) - The
then
field results is always a list. Its length equals the length of the outer list of thecases
field. - The
default
is always a scalar, i.e. a number, a!Vector3
, etc. It is an optional argument. If it is not provided, the tool will throw an error and stop execution when any variables take on a value that is not in the list of cases.
TODO: On references. Generated yaml file. First part of the dot separated must refer to an AnimalAI mapping (!ArenaConfig
, !Arena
, !Item
, !Vector3
, ...) with the corresponding id
field.
TODO: !ProcIf's can refer to other values decided by !ProcIf's (not circular of course). Add example for this.
# With only 1 variable
!ProcIf
value: agent.positions.0.z,
cases: [5, 7, 10]
then: [90, 120, 150]
default: 0
# With only 1 case and 1 variable
!ProcIf
value: agent.positions.0.z,
cases: [5]
then: [180]
default: 0
# With only 1 case, but two variables (one of them in a range)
!ProcIf
value: [agent.positions.0.x, agents.positions.0.z]
cases: [[1, !R [5, 10]]]
then: [180]
default: 0
# With three cases, and no default.
# The tool will throw an error and stop if agent.positions.0.x takes values other than those specified, e.g. 15.
!ProcIf
value: agent.positions.0.x
cases: [1, !R [2, 9], 10]
then: [0, 90, 180]
If you want to change the filename depending on some of the generated values, you can use can add a labels
field to !ProcIf
and!ProcVector3Scaled
, or you can use the new !ProcListLabelled
.
Make sure that your labels are filename friendly (e.g. no colons).
!ProcIf
value: [agent.positions.0.x, agents.positions.0.z]
cases: [[1, 5], [1, 7], [3, !R [10, 20]]]
then: [90, 120, 150]
labels: [dist_close, dist_med, dist_far]
default: 0
default_label: dist_default
# Currently no support for a default label yet.
!ProcVector3Scaled
base: !Vector3 { x: 2, y: 1, z: 1 }
scales: [1, 2, 3]
labels: [size_small, size_med, size_large]
!ProcListLabelled
- label: dist_far
value: !Vector3 { x: 20, y: 0, z: 1 }
- label: dist_medium
value: !Vector3 { x: 20, y: 1, z: 5 }
- label: dist_close
value: !Vector3 { x: 20, y: 1, z: 10 }
There is also a new special section proc_meta
in the beginning of the yaml file, which will not be present in the generate AnimalAI config, but allows you to specify independent labels. You have to use !ProcIfLabels
, which works exactly like !ProcIf
, but it will have the corresponding then
value included in the filename.
!ArenaConfig
proc_meta:
proc_labels:
- !ProcIfLabels
value: [agent.positions.0.x, agents.positions.0.z]
cases: [[1, 5], [1, 7], [3, !R [10, 20]]]
labels: [dist_far, dist_close, dist_medium]
default: dist_default # Optional
arenas:
0: !Arena
items: # ...