background_task.py

import random
import pandas as pd
from datetime import datetime
from huggingface_hub import HfApi


env = "Soccer"


class Model:
    """
    Class containing the info of a model.

    :param name: Name of the model
    :param elo: Elo rating of the model
    :param games_played: Number of games played by the model (useful if we implement sigma uncertainty)
    """
    def __init__(self, author, name, elo=1200, games_played=0):
        self.author = author
        self.name = name
        self.elo = elo
        self.games_played = games_played


class Matchmaking:
    """
    Class managing the matchmaking between the models.

    :param models: List of models
    :param queue: Temporary list of models used for the matching process
    :param k: Dev coefficient
    :param max_diff: Maximum difference considered between two models' elo
    :param matches: Dictionary containing the match history (to later upload as CSV)
    """
    def __init__(self, models):
        self.models = models
        self.queue = self.models.copy()
        self.k = 20
        self.max_diff = 500
        self.matches = {
            "model1": [],
            "model2": [],
            "result": [],
            "datetime": [],
            "env": []
        }

    def run(self):
        """
        Run the matchmaking process.
        Add models to the queue, shuffle it, and match the models one by one to models with close ratings.
        Compute the new elo for each model after each match and add the match to the match history.
        """
        self.queue = self.models.copy()
        random.shuffle(self.queue)
        while len(self.queue) > 1:
            model1 = self.queue.pop(0)
            model2 = self.queue.pop(self.find_n_closest_indexes(model1, 10))
            result = match(model1, model2)
            self.compute_elo(model1, model2, result)
            self.matches["model1"].append(model1.name)
            self.matches["model2"].append(model2.name)
            self.matches["result"].append(result)
            self.matches["datetime"].append(datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f"))
            self.matches["env"].append(env)

    def compute_elo(self, model1, model2, result):
        """ Compute the new elo for each model based on a match result. """
        delta = model1.elo - model2.elo
        win_probability = 1 / (1 + 10 ** (-delta / 500))
        model1.elo += self.k * (result - win_probability)
        model2.elo -= self.k * (result - win_probability)

    def find_n_closest_indexes(self, model, n) -> int:
        """
        Get a model index with a fairly close rating. If no model is found, return the last model in the queue.
        We don't always pick the closest rating to add variety to the matchups.

        :param model: Model to compare
        :param n: Number of close models from which to pick a candidate
        :return: id of the chosen candidate
        """
        indexes = []
        closest_diffs = [9999999] * n
        for i, m in enumerate(self.queue):
            if m.name == model.name:
                continue
            diff = abs(m.elo - model.elo)
            if diff < max(closest_diffs):
                closest_diffs.append(diff)
                closest_diffs.sort()
                closest_diffs.pop()
                indexes.append(i)
        random.shuffle(indexes)
        return indexes[0]

    def to_csv(self):
        """ Save the match history as a CSV file to the hub. """
        data_dict = {"author": [], "name": [], "elo": [], "games_played": []}
        for model in self.models:
            data_dict["author"].append(model.author)
            data_dict["name"].append(model.name)
            data_dict["elo"].append(model.elo)
            data_dict["games_played"].append(model.games_played)

        df = pd.DataFrame(data_dict)
        df.to_csv('env_elos/elo.csv', index=False)
        df_matches = pd.DataFrame(self.matches)
        date = datetime.now()
        df_matches.to_csv(f"matches/{date.strftime('%Y-%m-%d_%H-%M-%S_%f')}.csv", index=False)


def match(model1, model2) -> float:
    """
    !!! Current code is placeholder !!!
    TODO: Launch a Unity process with the 2 models and get the result of the match

    :param model1: First Model object
    :param model2: Second Model object
    :return: match result (0: model1 lost, 0.5: draw, 1: model1 won)
    """
    result = random.randint(0, 2) / 2

    model1.games_played += 1
    model2.games_played += 1
    return result


def get_models_list() -> list:
    """
    !!! Current code is placeholder !!!
    TODO: Create a list of Model objects from the models found on the hub

    :return: list of Model objects
    """
    models = []
    models_names = []
    data = pd.read_csv("env_elos/elo.csv")
    # models_on_hub = api.list_models(filter=["reinforcement-learning", env, "stable-baselines3"])
    models_on_hub = []
    for i, row in data.iterrows():
        models.append(Model(row["author"], row["name"], row["elo"], row["games_played"]))
        models_names.append(row["name"])
    for model in models_on_hub:
        if model.modelId not in models_names:
            models.append(Model(model.author, model.modelId))
    return models


def init_matchmaking():
    models = get_models_list()
    matchmaking = Matchmaking(models)
    matchmaking.run()
    matchmaking.to_csv()


if __name__ == "__main__":
    print("It's running!")
    api = HfApi()
    init_matchmaking()