import os import time import random import asyncio import subprocess import pandas as pd from datetime import datetime from huggingface_hub import HfApi, Repository DATASET_REPO_URL = "https://huggingface.co/datasets/CarlCochet/BotFightData" ELO_FILENAME = "soccer_elo.csv" ELO_DIR = "soccer_elo" HF_TOKEN = os.environ.get("HF_TOKEN") repo = Repository( local_dir=ELO_DIR, clone_from=DATASET_REPO_URL, use_auth_token=HF_TOKEN ) 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)) match(model1, model2) self.load_results() def load_results(self): """ Load the match history from the hub. """ repo.git_pull() results = pd.read_csv( "https://huggingface.co/datasets/huggingface-projects/temp-match-results/raw/main/results.csv" ) # while len(results) < len(self.matches["model1"]): # time.sleep(60) # results = pd.read_csv( # "https://huggingface.co/datasets/huggingface-projects/temp-match-results/raw/main/results.csv" # ) for i, row in results.iterrows(): model1 = row["model1"].split("/") model2 = row["model2"].split("/") model1 = self.find_model(model1[0], model1[1]) model2 = self.find_model(model2[0], model2[1]) result = row["result"] self.compute_elo(row["model1"], row["model2"], row["result"]) self.matches["model1"].append(model1.name) self.matches["model2"].append(model2.name) self.matches["result"].append(result) self.matches["timestamp"].append(row["timestamp"]) def find_model(self, author, name): """ Find a model in the models list. """ for model in self.models: if model.author == author and model.name == name: return model return None 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 = {"rank": [], "author": [], "model": [], "elo": [], "games_played": []} sorted_models = sorted(self.models, key=lambda x: x.elo, reverse=True) for i, model in enumerate(sorted_models): data_dict["rank"].append(i + 1) data_dict["author"].append(model.author) data_dict["model"].append(model.name) data_dict["elo"].append(model.elo) data_dict["games_played"].append(model.games_played) df = pd.DataFrame(data_dict) print(df.head()) repo.git_pull() df.to_csv(os.path.join(ELO_DIR, ELO_FILENAME), index=False) repo.push_to_hub(commit_message="Update ELO") def match(model1, model2): """ :param model1: First Model object :param model2: Second Model object :return: match result (0: model1 lost, 0.5: draw, 1: model1 won) """ model1_id = model1.author + "/" + model1.name model2_id = model2.author + "/" + model2.name subprocess.run(["UnityEnvironment.exe", "-model1", model1_id, "-model2", model2_id]) model1.games_played += 1 model2.games_played += 1 def get_models_list() -> list: """ :return: list of Model objects """ models = [] models_names = [] data = pd.read_csv(os.path.join(DATASET_REPO_URL, "resolve", "main", ELO_FILENAME)) models_on_hub = api.list_models(filter=["reinforcement-learning", "ml-agents", "ML-Agents-SoccerTwos"]) for i, row in data.iterrows(): models.append(Model(row["author"], row["model"], row["elo"], row["games_played"])) models_names.append(row["model"]) for model in models_on_hub: print("New model found: ", model.author, "-", model.modelId) 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() print("Matchmaking done ---", datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")) async def run_background_loop(): while True: print("It's running!", datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")) await asyncio.sleep(60) if __name__ == "__main__": print("It's running!") api = HfApi() init_matchmaking()