ThomasSimonini's picture
Test video
48f6ca1
import requests
import pandas as pd
from tqdm.auto import tqdm
from utils import *
import gradio as gr
from huggingface_hub import HfApi, hf_hub_download
from huggingface_hub.repocard import metadata_load
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
def get_ids(self):
return list(self.leaderboard.keys())
# CSS file for the
with open('app.css','r') as f:
BLOCK_CSS = f.read()
LOADED_MODEL_IDS = {}
LOADED_MODEL_METADATA = {}
def get_data(rl_env):
global LOADED_MODEL_IDS ,LOADED_MODEL_METADATA
data = []
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> βŒ› 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 πŸŒ• 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-v3πŸš– Leaderboard')
rl_leaderboard.add_leaderboard('CarRacing-v0'," The Car Racing 🏎️ Leaderboard")
rl_leaderboard.add_leaderboard('MountainCar-v0',"The Mountain Car ⛰️ πŸš— 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 πŸ€– Leaderboard')
rl_leaderboard.add_leaderboard('AntBulletEnv-v0','The AntBulletEnv-v0 πŸ•ΈοΈ Leaderboard')
rl_leaderboard.add_leaderboard('HalfCheetahBulletEnv-v0','The HalfCheetahBulletEnv-v0 πŸ€– 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):
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 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> βŒ› 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 style="text-align:center"> {name_leaderboard} </h1>
<br>
<video controls width="250">
<source src="https://huggingface.co/sb3/ppo-CartPole-v1/resolve/main/replay.mp4"
>
</video>
</video>
<p> This is a leaderboard of <b>{len_dataframe}</b> agents, from <b>{num_unique_users}</b> unique users, playing {env_name} πŸ‘©β€πŸš€. </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 πŸ€— </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 πŸ‘€ </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> βœ… 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)
return markdown,data_html
block = gr.Blocks(css=BLOCK_CSS)
with block:
notification = gr.HTML("""<div style="color: green">
<p> βŒ› Updating leaderboard... </p>
</div>
""")
block.load(reload_all_data,[],[notification])
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)
rl_tab.select(reload_leaderboard,inputs=[env_state],outputs=[output_markdown,output_html])
block.launch()