Multimon

Description

Multimon is a dataset specifically made for Multitask Learning (MTL) based on the popular Pokémon videogame series. It contains sprites of every pokemon in the series and has labels for multiple tasks.

Dataset Structure

Note that due to copyright restrictions, the dataset is not available for direct download. Other than the main script, the files described below are not present but are constructed by running the script main.py.

The Main Script

This script contains everything you need to construct the dataset. It will download the sprites and labels from the PokeAPI, perform postprocessing (standardising the sprites and labels) and construct partitions for train/test and train/test/validation splits.

The Sprites

Sprites are stored in two directories, sprites/processed for the actual sprites to use and sprites/raw for raw images (these are provided in case you want to do custom postprocessing, though it's unlikely you'll ever need to)

The Data

This is the main file containing the labels. It contains the following columns.

Name	Description	Type
index	The filename of the sprite	string
type1, type21	The first and second type of the pokemon	int
gen	The generation the pokemon was introduced in	integer
hp_raw, hp_norm, hp_std2	The HP stat of the pokemon	float
atk_raw, atc_norm, atc_std2	The Attack stat of the pokemon	float
def_raw, def_norm, def_std2	The Defense stat of the pokemon	float
spatk_raw, spatk_norm, spatk_std2	The Special Attack stat of the pokemon	float
spdef_raw, spdef_norm, spdef_std2	The Special Defense stat of the pokemon	float
spd_raw, spd_norm, spd_std2	The Speed stat of the pokemon	float
height_raw, height_norm, height_std2	The height of the pokemon	float
weight_raw, weight_norm, weight_std2	The weight of the pokemon	float

Partitions

Partitions are an important part of a dataset to ensure consistency between runs. Since this repo only contains the script for constructing the dataset, the partitions are also not available for direct download. To ensure consistency the partitions are created by hashing the name of the pokemon and using that value to assign partitions. This method is consistent between python versions, download instances and also in the event of more Pokémon being released.

The partitions have a 70/15/15 split for train/test/validation splits.

Footnotes

1. Pokémon have one or two of 17 types, for consistency we create an 18th "null" type. type1 and type2 should be interpreted as a two-hot vector together. ↩

2. hp and all the columns below are provided as the raw value, as well as the normalised and standardised values (suffixed as "_raw", "_norm" and "_std" respectively) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸