import json
import requests
from datasets import load_dataset
import gradio as gr
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import HfApi, hf_hub_download
from huggingface_hub.repocard import metadata_load
import pandas as pd
from utils import *
block = gr.Blocks()
# 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": "PitfallNoFrameskip-v4 🏃",
"rl_env": "PitfallNoFrameskip-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)
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]
#print(model_ids)
return model_ids
def get_model_dataframe(rl_env):
# Get model ids associated with rl_env
model_ids = get_model_ids(rl_env)
#print(model_ids)
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"] = 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)
ranked_dataframe = rank_dataframe(pd.DataFrame.from_records(data))
#print("RANKED", ranked_dataframe)
return ranked_dataframe
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
with block:
gr.Markdown(f"""
# 🏆 The Deep Reinforcement Learning Course Leaderboard 🏆
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 🏆
**The leaderboard is updated every hour. If you don't find your model, go to the bottom of the page and click on the refresh button**
We use **lower bound result to sort the models: mean_reward - std_reward.**
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? Check the Hugging Face free Deep Reinforcement Learning Course 🤗 .
You want to compare two agents? It's possible using this Spaces demo 👀 .
🔧 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()
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")
scheduler = BackgroundScheduler()
# Refresh every hour
scheduler.add_job(func=refresh_leaderboard, trigger="interval", seconds=3600)
scheduler.start()