app.py

import os
import json
import requests

import gradio as gr
import pandas as pd
from huggingface_hub import HfApi, hf_hub_download, snapshot_download
from huggingface_hub.repocard import metadata_load
from apscheduler.schedulers.background import BackgroundScheduler

from utils import *

DATASET_REPO_URL = "https://huggingface.co/datasets/huggingface-projects/drlc-leaderboard-data"
DATASET_REPO_ID = "huggingface-projects/drlc-leaderboard-data"
HF_TOKEN = os.environ.get("HF_TOKEN")

block = gr.Blocks()
api = HfApi(token=HF_TOKEN)

# Containing the data
rl_envs = [
{
"rl_env_beautiful": "LunarLander-v2 🚀",
"rl_env": "LunarLander-v2",
"video_link": "",
"global": None
},    
{
"rl_env_beautiful": "CartPole-v1",
"rl_env": "CartPole-v1",
"video_link": "https://huggingface.co/sb3/ppo-CartPole-v1/resolve/main/replay.mp4",
"global": None
},
{
"rl_env_beautiful": "FrozenLake-v1-4x4-no_slippery ❄️",
"rl_env": "FrozenLake-v1-4x4-no_slippery",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "FrozenLake-v1-8x8-no_slippery ❄️",
"rl_env": "FrozenLake-v1-8x8-no_slippery",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "FrozenLake-v1-4x4 ❄️",
"rl_env": "FrozenLake-v1-4x4",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "FrozenLake-v1-8x8 ❄️",
"rl_env": "FrozenLake-v1-8x8",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "Taxi-v3 🚖",
"rl_env": "Taxi-v3",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "CarRacing-v0 🏎️",
"rl_env": "CarRacing-v0",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "MountainCar-v0 ⛰️",
"rl_env": "MountainCar-v0",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "SpaceInvadersNoFrameskip-v4 👾",
"rl_env": "SpaceInvadersNoFrameskip-v4",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "PongNoFrameskip-v4 🎾",
"rl_env": "PongNoFrameskip-v4",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "BreakoutNoFrameskip-v4 🧱",
"rl_env": "BreakoutNoFrameskip-v4",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "QbertNoFrameskip-v4 🐦",
"rl_env": "QbertNoFrameskip-v4",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "BipedalWalker-v3",
"rl_env": "BipedalWalker-v3",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "Walker2DBulletEnv-v0",
"rl_env": "Walker2DBulletEnv-v0",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "AntBulletEnv-v0",
"rl_env": "AntBulletEnv-v0",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "HalfCheetahBulletEnv-v0",
"rl_env": "HalfCheetahBulletEnv-v0",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "PandaReachDense-v2",
"rl_env": "PandaReachDense-v2",
"video_link": "",
"global": None
},
{
"rl_env_beautiful": "Pixelcopter-PLE-v0",
"rl_env": "Pixelcopter-PLE-v0",
"video_link": "",
"global": None
}
]

def get_metadata(model_id):
    try:
        readme_path = hf_hub_download(model_id, filename="README.md")
        return metadata_load(readme_path)
    except requests.exceptions.HTTPError:
        # 404 README.md not found
        return None
        
def parse_metrics_accuracy(meta):
    if "model-index" not in meta:
        return None
    result = meta["model-index"][0]["results"]
    metrics = result[0]["metrics"]
    accuracy = metrics[0]["value"]
    return accuracy

# We keep the worst case episode
def parse_rewards(accuracy):
    default_std = -1000
    default_reward=-1000
    if accuracy !=  None:
        accuracy = str(accuracy)
        parsed =  accuracy.split(' +/- ')
        if len(parsed)>1:
            mean_reward = float(parsed[0])
            std_reward =  float(parsed[1])
        elif len(parsed)==1: #only mean reward   
            mean_reward = float(parsed[0])
            std_reward =  float(0)
             
        else: 
            mean_reward = float(default_std)
            std_reward = float(default_reward)

    else:
        mean_reward = float(default_std)
        std_reward = float(default_reward)
    return mean_reward, std_reward


def get_model_ids(rl_env):
    api = HfApi()
    models = api.list_models(filter=rl_env)
    model_ids = [x.modelId for x in models]
    return model_ids

def update_leaderboard_dataset(rl_env):
    # Get model ids associated with rl_env
    model_ids = get_model_ids(rl_env)
    data = []
    for model_id in model_ids:
        """
        readme_path = hf_hub_download(model_id, filename="README.md")
        meta = metadata_load(readme_path)
        """
        meta = get_metadata(model_id)
        #LOADED_MODEL_METADATA[model_id] = meta if meta is not None else ''
        if meta is None:
            continue
        user_id = model_id.split('/')[0]
        row = {}
        row["User"] = user_id
        row["Model"] = model_id
        accuracy = parse_metrics_accuracy(meta)
        mean_reward, std_reward = parse_rewards(accuracy)
        mean_reward = mean_reward if not pd.isna(mean_reward) else 0
        std_reward = std_reward if not pd.isna(std_reward) else 0
        row["Results"] = mean_reward - std_reward
        row["Mean Reward"] = mean_reward
        row["Std Reward"] = std_reward
        data.append(row)

    ranked_dataframe = rank_dataframe(pd.DataFrame.from_records(data))
    new_history = ranked_dataframe
    filename = rl_env + ".csv"
    new_history.to_csv(filename, index=False)
    csv_path = os.path.abspath(filename)

    api.upload_file(
    path_or_fileobj= csv_path,
    path_in_repo= filename,
    repo_id="huggingface-projects/drlc-leaderboard-data",
    repo_type="dataset",
    commit_message="Update dataset")

    return ranked_dataframe

def download_leaderboard_dataset():
    path = snapshot_download(repo_id=DATASET_REPO_ID, repo_type="dataset")
    return path

def get_data(rl_env, path) -> pd.DataFrame:
    """
    Get data from rl_env
    :return: data as a pandas DataFrame
    """
    csv_path = path + "/" + rl_env + ".csv"
    data = pd.read_csv(csv_path)

    for index, row in data.iterrows():
        user_id = row["User"]
        data.loc[index, "User"] = make_clickable_user(user_id)
        model_id = row["Model"]
        data.loc[index, "Model"] = make_clickable_model(model_id)
        
    return data

def get_data_no_html(rl_env, path) -> pd.DataFrame:
    """
    Get data from rl_env
    :return: data as a pandas DataFrame
    """
    csv_path = path + "/" + rl_env + ".csv"
    data = pd.read_csv(csv_path)

    return data
    
def rank_dataframe(dataframe):
    dataframe = dataframe.sort_values(by=['Results'], ascending=False)
    if not 'Ranking' in dataframe.columns:
        dataframe.insert(0, 'Ranking', [i for i in range(1,len(dataframe)+1)])
    else:
        dataframe['Ranking'] =   [i for i in range(1,len(dataframe)+1)]
    return dataframe


def run_update_dataset():
    for i in range(0, len(rl_envs)):
        rl_env = rl_envs[i]
        update_leaderboard_dataset(rl_env["rl_env"])


def filter_data(rl_env, path, user_id):
    print("RL ENV", rl_env)
    print("PATH", path)
    print("USER ID", user_id)
    data_df = get_data_no_html(rl_env, path)
    print(data_df)
    models = []
    models = data_df[data_df["User"] == user_id]

    for index, row in models.iterrows():
        user_id = row["User"]
        models.loc[index, "User"] = make_clickable_user(user_id)
        model_id = row["Model"]
        models.loc[index, "Model"] = make_clickable_model(model_id)
        

    print(models)
    return models
    
with block:
    gr.Markdown(f"""
    # 🏆 The Deep Reinforcement Learning Course Leaderboard 🏆 
    
    This is the leaderboard of trained agents during the <a href="https://huggingface.co/learn/deep-rl-course/unit0/introduction?fw=pt">Deep Reinforcement Learning Course</a>. A free course from beginner to expert.
    
    ### We only display the best 100 models
    If you want to **find yours, type your user id and click on Search my models.**
    You **can click on the model's name** to be redirected to its model card, including documentation.
    
    ### How are the results calculated?
    We use **lower bound result to sort the models: mean_reward - std_reward.**

    ### I can't find my model 😭
    The leaderboard is **updated every hour** if you can't find your models, just wait for the next update.
    
    ### The Deep RL Course
    🤖 You want to try to train your agents? <a href="https://huggingface.co/deep-rl-course/unit0/introduction?fw=pt" target="_blank"> Check the Hugging Face free Deep Reinforcement Learning Course 🤗 </a>.
        
    🔧 There is an **environment missing?** Please open an issue.
    """)
    path_ = download_leaderboard_dataset()

    for i in range(0, len(rl_envs)):
        rl_env = rl_envs[i]
        with gr.TabItem(rl_env["rl_env_beautiful"]) as rl_tab:
            with gr.Row():
                markdown = """
                    # {name_leaderboard}
                    
                    """.format(name_leaderboard = rl_env["rl_env_beautiful"], video_link = rl_env["video_link"])
                gr.Markdown(markdown)
                #gr.Textbox(value=filter_data(rl_env["rl_env"], path_))
            
            with gr.Row():
                gr.Markdown("""
                    ## Search your models
                    Simply type your user id to find your models
                    """)
                
            with gr.Row():
                user_id = gr.Textbox(label= "Your user id")
                search_btn = gr.Button("Search my models 🔎")
                reset_btn = gr.Button("Clear my search")
                env = gr.Variable(rl_env["rl_env"])
                grpath = gr.Variable(path_)
            with gr.Row():
                gr_dataframe = gr.components.Dataframe(value=get_data(rl_env["rl_env"], path_), headers=["Ranking 🏆", "User 🤗", "Model id 🤖", "Results", "Mean Reward", "Std Reward"], datatype=["number", "markdown", "markdown", "number", "number", "number"],  row_count=(100, 'fixed'))
             
            with gr.Row():
                #gr_search_dataframe = gr.components.Dataframe(headers=["Ranking 🏆", "User 🤗", "Model id 🤖", "Results", "Mean Reward", "Std Reward"], datatype=["number", "markdown", "markdown", "number", "number", "number"], visible=False)
                search_btn.click(fn=filter_data, inputs=[env, grpath, user_id], outputs=gr_dataframe, api_name="filter_data")

            with gr.Row():
                search_btn.click(fn=filter_data, inputs=[env, grpath, user_id], outputs=gr_dataframe, api_name="filter_data")
                reset_btn.click(fn=get_data, inputs=[env, grpath], outputs=gr_dataframe, api_name="get_data")
                       
block.launch()

def refresh_leaderboard():
    """
    Here we refresh the leaderboard:
    we update the rl_env["global"] for each rl_envs in rl_env
    """
    for i in range(0, len(rl_envs)):
        rl_env = rl_envs[i]
        temp = get_model_dataframe(rl_env)
        rl_env["global"] = temp
    print("The leaderboard has been updated")
                       
block.launch()

scheduler = BackgroundScheduler()
# Refresh every hour
scheduler.add_job(func=run_update_dataset, trigger="interval", seconds=3600)
scheduler.start()