chore: init

.gitignore

@@ -0,0 +1 @@
+__pycache__/*

app.py

@@ -0,0 +1,385 @@
+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 tqdm.contrib.concurrent import thread_map
+
+from utils import *
+
+DATASET_REPO_URL = "https://huggingface.co/datasets/mshamrai/rlc-leaderboard-data"
+DATASET_REPO_ID = "mshamrai/rlc-leaderboard-data"
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+STUDENTS_SET = {"mshamrai", "Kolosok", "grinvolod", "ostap-khm", "elusivephantasm", "letaldir"}
+
+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": "CarRacing-v2 🏎️",
+"rl_env": "CarRacing-v2",
+"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": "PandaReachDense-v3",
+"rl_env": "PandaReachDense-v3",
+"video_link": "",
+"global": None
+},  
+{
+"rl_env_beautiful": "Pixelcopter-PLE-v0",
+"rl_env": "Pixelcopter-PLE-v0",
+"video_link": "",
+"global": None
+}
+]
+
+def restart():
+    print("RESTART")
+    api.restart_space(repo_id="huggingface-projects/Deep-Reinforcement-Learning-Leaderboard")
+
+def get_metadata(model_id):
+    try:
+        readme_path = hf_hub_download(model_id, filename="README.md", etag_timeout=180)
+        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].strip())
+            std_reward =  float(parsed[1].strip())
+        elif len(parsed)==1: #only mean reward
+            mean_reward = float(parsed[0].strip())
+            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 filter_students(model_ids):
+    filtered = []
+    for model_id in model_ids:
+        user_id = model_id.split('/')[0]
+        if user_id in STUDENTS_SET:
+            filtered.append(model_id)
+    return filtered
+
+# Parralelized version
+def update_leaderboard_dataset_parallel(rl_env, path):
+    # Get model ids associated with rl_env
+    model_ids = get_model_ids(rl_env)
+    model_ids = filter_students(model_ids)
+
+    def process_model(model_id):
+        meta = get_metadata(model_id)
+        #LOADED_MODEL_METADATA[model_id] = meta if meta is not None else ''
+        if meta is None:
+            return None
+        try:
+            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
+            return row
+        except:
+            return None
+
+    data = list(thread_map(process_model, model_ids, desc="Processing models"))
+
+    # Filter out None results (models with no metadata)
+    data = [row for row in data if row is not None]
+
+    ranked_dataframe = rank_dataframe(pd.DataFrame.from_records(data))
+    new_history = ranked_dataframe
+    file_path = path + "/" + rl_env + ".csv"
+    new_history.to_csv(file_path, index=False)
+
+    return ranked_dataframe
+
+
+def update_leaderboard_dataset(rl_env, path):
+    # 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
+    file_path = path + "/" + rl_env + ".csv"
+    new_history.to_csv(file_path, index=False)
+
+    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):
+    if dataframe.empty:
+        return pd.DataFrame(columns=['User', 'Model', 'Results', 'Mean Reward', 'Std Reward'])
+    dataframe = dataframe.sort_values(by=['Results', 'User', 'Model'], 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():
+    path_ = download_leaderboard_dataset()
+    for i in range(0, len(rl_envs)):
+        rl_env = rl_envs[i]
+        update_leaderboard_dataset_parallel(rl_env["rl_env"], path_)
+
+    api.upload_folder(
+    folder_path=path_,
+    repo_id="mshamrai/rlc-leaderboard-data",
+    repo_type="dataset",
+    commit_message="Update dataset")
+
+run_update_dataset()
+
+with block:
+    gr.Markdown(f"""
+    # 🏆 Reinforcement Learning Course Leaderboard 🏆 
+
+    This leaderboard is for Kyiv Academic University students to see their results during the Hugging Face <a href="https://huggingface.co/learn/deep-rl-course/unit0/introduction?fw=pt">Deep Reinforcement Learning Course</a>.
+    
+    ### 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 two hours** if you can't find your models, just wait for the next update.
+    """)
+    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)
+                
+            
+            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=(15, 'dynamic'))
+    """
+    block.load(
+        download_leaderboard_dataset,
+        inputs=[],
+        outputs=[
+            grpath
+        ],
+    )
+    """
+
+
+scheduler = BackgroundScheduler()
+# Refresh every hour
+#scheduler.add_job(func=run_update_dataset, trigger="interval", seconds=3600)
+#scheduler.add_job(download_leaderboard_dataset, 'interval', seconds=3600)
+#scheduler.add_job(run_update_dataset, 'interval', seconds=3600)
+scheduler.add_job(restart, 'interval', seconds=10800)
+scheduler.start()
+
+block.launch()

requirements.txt

@@ -0,0 +1,4 @@
+APScheduler==3.10.1
+gradio==4.44.1
+httpx>=0.24.1
+tqdm

utils.py

@@ -0,0 +1,14 @@
+# Based on Omar Sanseviero work
+# Make model clickable link
+def make_clickable_model(model_name):
+    # remove user from model name
+    model_name_show = ' '.join(model_name.split('/')[1:])
+
+    link = "https://huggingface.co/" + model_name
+    return f'<a target="_blank" href="{link}">{model_name_show}</a>'
+
+# Make user clickable link
+def make_clickable_user(user_id):
+    link = "https://huggingface.co/" + user_id
+    return f'<a  target="_blank" href="{link}">{user_id}</a>'
+