File size: 7,119 Bytes
b62b9e0
e155868
513af34
e155868
 
513af34
 
f980695
513af34
 
b62b9e0
eb1e3b7
3ee7ba6
b62b9e0
513af34
f980695
3ee7ba6
f980695
 
513af34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e155868
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
513af34
 
 
e155868
 
 
 
 
 
 
 
513af34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a5a13b
 
 
 
513af34
4b1b8a4
513af34
 
 
72e56e4
485e890
3ee7ba6
f980695
513af34
 
e155868
513af34
 
 
 
 
e155868
 
 
513af34
 
 
 
 
 
 
 
 
 
eb1e3b7
e155868
513af34
4b1b8a4
44c646a
513af34
88faa07
513af34
 
 
 
 
 
 
7a5e1aa
 
 
3a5a13b
513af34
 
e155868
 
 
 
 
 
513af34
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
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()