Big Leaderboard Update

README.md

@@ -4,7 +4,7 @@ emoji: 🚀
 colorFrom: green
 colorTo: indigo
 sdk: gradio
-sdk_version: 3.0.20
+sdk_version: 3.11.0
 app_file: app.py
 pinned: false
 ---

app.css

@@ -1,37 +0,0 @@
-
-.infoPoint h1 {
-    font-size: 30px;
-    text-decoration: bold;
-    
-    }
-    
-a {
-    text-decoration: underline; 
-    color: #1f3b54 ;
-}
-
-table {
-  
-    margin: 25px 0;
-    font-size: 0.9em;
-    font-family: sans-serif;
-    min-width: 400px;
-    box-shadow: 0 0 20px rgba(0, 0, 0, 0.15);
-}
-
-table th,
-table td {
-    padding: 12px 15px;
-}
-
-tr {
-text-align: left;
-}
-thead tr {
-text-align: left;
-}
-
-.flex
-{
-    overflow:auto;
-}

app.py

@@ -1,238 +1,225 @@
+import json
+
 import requests
-import pandas as pd
-from tqdm.auto import tqdm
-from utils import *
-import gradio as gr
 
+from datasets import load_dataset
+import gradio as gr
 from huggingface_hub import HfApi, hf_hub_download
 from huggingface_hub.repocard import metadata_load
+import pandas as pd
+from utils import *
 
-class DeepRL_Leaderboard:
-    def __init__(self) -> None:
-        self.leaderboard= {} 
 
-    def add_leaderboard(self,id=None, title=None):
-        if id is not None and title is not None:
-            id = id.strip()
-            title = title.strip()
-            self.leaderboard.update({id:{'title':title,'data':get_data_per_env(id)}})
-            
-    def get_data(self):
-        return self.leaderboard
+block = gr.Blocks()
+
+# Containing the data
+rl_envs = [{
+"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": "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
+}      
+]
+
+
+
+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)
 
-    def get_ids(self):
-        return list(self.leaderboard.keys())
+    else:
+        mean_reward = float(default_std)
+        std_reward = float(default_reward)
+    return mean_reward, std_reward
 
-# CSS file for the
-with open('app.css','r') as f:
-    BLOCK_CSS = f.read() 
 
-LOADED_MODEL_IDS = {}
-LOADED_MODEL_METADATA = {}
+def get_model_ids(rl_env):
+    api = HfApi()
+    models = api.list_models(filter=rl_env)
+    model_ids = [x.modelId for x in models]
+    print(model_ids)
+    return model_ids
 
-def get_data(rl_env):
-    global LOADED_MODEL_IDS ,LOADED_MODEL_METADATA
-    data = []
+def get_model_dataframe(rl_env):
+    # Get model ids associated with rl_env
     model_ids = get_model_ids(rl_env)
-    LOADED_MODEL_IDS[rl_env]=model_ids
-
-    for model_id in tqdm(model_ids):
-        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)
-    return pd.DataFrame.from_records(data)
-
-def get_data_per_env(rl_env):
-    dataframe = get_data(rl_env)
-    dataframe = dataframe.fillna("")
-
-    if not dataframe.empty:
-        # turn the model ids into clickable links
-        dataframe["User"] = dataframe["User"].apply(make_clickable_user)
-        dataframe["Model"] = dataframe["Model"].apply(make_clickable_model)
-        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)]   
-        table_html = dataframe.to_html(escape=False, index=False,justify = 'left')
-        return table_html,dataframe,dataframe.empty
-    else: 
-        html = """<div style="color: green">
-                <p> ⌛ Please wait. Results will be out soon... </p>
-                </div>
-               """
-        return html,dataframe,dataframe.empty   
-
-
-
-rl_leaderboard = DeepRL_Leaderboard()
-rl_leaderboard.add_leaderboard('CartPole-v1','The Cartpole-v1 Leaderboard')
-rl_leaderboard.add_leaderboard('LunarLander-v2',"The Lunar Lander 🌕 Leaderboard")
-rl_leaderboard.add_leaderboard('FrozenLake-v1-4x4-no_slippery','The FrozenLake-v1-4x4-no_slippery Leaderboard')
-rl_leaderboard.add_leaderboard('FrozenLake-v1-8x8-no_slippery','The FrozenLake-v1-8x8-no_slippery Leaderboard')
-rl_leaderboard.add_leaderboard('FrozenLake-v1-4x4','The FrozenLake-v1-4x4 Leaderboard')
-rl_leaderboard.add_leaderboard('FrozenLake-v1-8x8','The FrozenLake-v1-8x8 Leaderboard')
-rl_leaderboard.add_leaderboard('Taxi-v3','The Taxi-v3🚖 Leaderboard')
-rl_leaderboard.add_leaderboard('CarRacing-v0'," The Car Racing 🏎️ Leaderboard")
-rl_leaderboard.add_leaderboard('MountainCar-v0',"The Mountain Car ⛰️ 🚗 Leaderboard")
-rl_leaderboard.add_leaderboard('BipedalWalker-v3',"The BipedalWalker Leaderboard")
-rl_leaderboard.add_leaderboard('SpaceInvadersNoFrameskip-v4','The SpaceInvadersNoFrameskip-v4 Leaderboard')
-rl_leaderboard.add_leaderboard('Pixelcopter-PLE-v0','The Pixelcopter-PLE-v0 🚁 Leaderboard')
-rl_leaderboard.add_leaderboard('Pong-PLE-v0','The Pong-PLE-v0 🎾 Leaderboard')
-rl_leaderboard.add_leaderboard('Walker2DBulletEnv-v0','The Walker2DBulletEnv-v0 🤖 Leaderboard')
-rl_leaderboard.add_leaderboard('AntBulletEnv-v0','The AntBulletEnv-v0 🕸️ Leaderboard')
-rl_leaderboard.add_leaderboard('HalfCheetahBulletEnv-v0','The HalfCheetahBulletEnv-v0 🤖 Leaderboard')
-RL_ENVS = rl_leaderboard.get_ids()
-RL_DETAILS = rl_leaderboard.get_data()
-
-
-def update_data(rl_env):
-    global LOADED_MODEL_IDS,LOADED_MODEL_METADATA
     data = []
-    model_ids = [x for x in get_model_ids(rl_env)] #if x not in LOADED_MODEL_IDS[rl_env]] # For now let's update all
-    
-    LOADED_MODEL_IDS[rl_env]+=model_ids
-
-    for model_id in tqdm(model_ids):
+    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 ''
+        #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
+        row["User"] = make_clickable_user(user_id)
+        row["Model"] = make_clickable_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)
-    return pd.DataFrame.from_records(data)
-
-
-def update_data_per_env(rl_env):
-    global RL_DETAILS
-
-    _,old_dataframe,_ = RL_DETAILS[rl_env]['data']
-    new_dataframe = update_data(rl_env)
-
-    new_dataframe = new_dataframe.fillna("")
-    if not new_dataframe.empty:
-        new_dataframe["User"] = new_dataframe["User"].apply(make_clickable_user)
-        new_dataframe["Model"] = new_dataframe["Model"].apply(make_clickable_model)
-
-    dataframe = pd.concat([old_dataframe,new_dataframe])
-
-    if not dataframe.empty:
-       
-        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)]   
-        table_html = dataframe.to_html(escape=False, index=False,justify = 'left')
-        return table_html,dataframe,dataframe.empty
-    else: 
-        html = """<div style="color: green">
-                <p> ⌛ Please wait. Results will be out soon... </p>
-                </div>
-               """
-        return html,dataframe,dataframe.empty   
-
-
-def get_info_display(dataframe,env_name,name_leaderboard,is_empty):
-    if not is_empty:
-        markdown = """
-        <div class='infoPoint'>
-        <h1> {name_leaderboard} </h1>
-        <br>
-        <p> This is a leaderboard of <b>{len_dataframe}</b> agents, from <b>{num_unique_users}</b> unique users, playing {env_name} 👩‍🚀. </p>
-        <br>
-        <p> We use <b>lower bound result to sort the models: mean_reward - std_reward.</b> </p>
-        <br>    
-        <p> You can click on the model's name to be redirected to its model card which includes documentation. </p>
-        <br>
-        <p> You want to try to train your agents? <a href="http://eepurl.com/h1pElX" target="_blank">Sign up to the Hugging Face free Deep Reinforcement Learning Class 🤗 </a>.
-        </p>
-        <br>
-        <p> You want to compare two agents? <a href="https://huggingface.co/spaces/ThomasSimonini/Compare-Reinforcement-Learning-Agents" target="_blank">It's possible using this Spaces demo 👀 </a>.
-        </p>
-        </div>
-        """.format(len_dataframe = len(dataframe),env_name = env_name,name_leaderboard = name_leaderboard,num_unique_users = len(set(dataframe['User'].values)))
-
-    else:
-        markdown = """
-        <div class='infoPoint'>
-        <h1> {name_leaderboard} </h1>
-        <br>
-        </div>                  
-        """.format(name_leaderboard =  name_leaderboard)
-    return markdown    
-
-def reload_all_data():
-
-    global RL_DETAILS,RL_ENVS
-
-    for rl_env in RL_ENVS:
-        RL_DETAILS[rl_env]['data'] = update_data_per_env(rl_env)
-
-    html = """<div style="color: green">
-                <p> ✅ Leaderboard updated! </p>
-                </div>
-               """    
-    return html            
-
-
-def reload_leaderboard(rl_env):
-    global RL_DETAILS
- 
-    data_html,data_dataframe,is_empty = RL_DETAILS[rl_env]['data'] 
-
-    markdown = get_info_display(data_dataframe,rl_env,RL_DETAILS[rl_env]['title'],is_empty)            
+    ranked_dataframe = rank_dataframe(pd.DataFrame.from_records(data))
+    print("RANKED", ranked_dataframe)
+    return ranked_dataframe
     
-    return markdown,data_html     
-            
-
-           
-
+    
+def rank_dataframe(dataframe):
+    print("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
 
 
-block = gr.Blocks(css=BLOCK_CSS)
 with block:
-    notification = gr.HTML("""<div style="color: green">
-                <p> ⌛ Updating leaderboard... </p>
-                </div>
-               """)
-    block.load(reload_all_data,[],[notification])
+    gr.Markdown(f"""
+    # 🏆 The Deep Reinforcement Learning Course Leaderboard 🏆 
     
-    with gr.Tabs():
-        for rl_env in RL_ENVS:
-            with gr.TabItem(rl_env) as rl_tab:
-                data_html,data_dataframe,is_empty = RL_DETAILS[rl_env]['data'] 
-                markdown = get_info_display(data_dataframe,rl_env,RL_DETAILS[rl_env]['title'],is_empty)            
-                env_state =gr.Variable(value=f'\"{rl_env}\"')  
-                output_markdown = gr.HTML(markdown)
-
-                output_html = gr.HTML(data_html)
+    This is the leaderboard of trained agents during the Deep Reinforcement Learning Course. A free course from beginner to expert.
+    
+    Just choose which environment you trained your agent on and with Ctrl+F find your rank 🏆
+    
+    We use **lower bound result to sort the models: mean_reward - std_reward.**
 
-                rl_tab.select(reload_leaderboard,inputs=[env_state],outputs=[output_markdown,output_html])
+    You **can click on the model's name** to be redirected to its model card which includes documentation.
+    
+    🤖 You want to try to train your agents? <a href="http://eepurl.com/ic5ZUD" target="_blank">Sign up to the Hugging Face free Deep Reinforcement Learning Course 🤗 </a>.
+    
+    You want to compare two agents? <a href="https://huggingface.co/spaces/ThomasSimonini/Compare-Reinforcement-Learning-Agents" target="_blank">It's possible using this Spaces demo 👀 </a>.
+    
+    🔧 There is an **environment missing?** Please open an issue.
+    """)
+    
+    #for rl_env in RL_ENVS:
+    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():
+                rl_env["global"] = gr.components.Dataframe(value= get_model_dataframe(rl_env["rl_env"]), headers=["Ranking 🏆", "User 🤗", "Model id 🤖", "Results", "Mean Reward", "Std Reward"], datatype=["number", "markdown", "markdown", "number", "number", "number"])
+            with gr.Row():
+                data_run = gr.Button("Refresh")
+                print("rl_env", rl_env["rl_env"])
+                val = gr.Variable(value=[rl_env["rl_env"]])
+                data_run.click(get_model_dataframe, inputs=[val], outputs =rl_env["global"])
+            
+                       
+block.launch()
 
-block.launch()    

utils.py

@@ -1,10 +1,3 @@
-import pandas as pd
-import requests
-from tqdm.auto import tqdm
-from huggingface_hub import HfApi, hf_hub_download
-from huggingface_hub.repocard import metadata_load
-
-
 # Based on Omar Sanseviero work
 # Make model clickable link
 def make_clickable_model(model_name):
@@ -18,51 +11,4 @@ def make_clickable_model(model_name):
 def make_clickable_user(user_id):
     link = "https://huggingface.co/" + user_id
     return f'<a  target="_blank" href="{link}">{user_id}</a>'
-    
-
-
-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 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
-
+