app.py

import gradio as gr
import os
from moviepy.editor import *

def replay(option):
	path = ""
	# Get the correct model
	if (option == "LunarLander-v2 🚀👩‍🚀"):
  		path = "./LunarLander-v2.mp4"
	elif(option == "CartPole-v1 🕹️"):
  		path = "./CartPole-v1.mp4"
	elif(option == "Atari Space Invaders 👾"):
  		path = "./SpaceInvadersNoFrameskip-v4.mp4"
	
	
  	# The only turnaround I found (Base64 video pb)
	videoclip = VideoFileClip(path)
	videoclip.write_videofile("new_filename.mp4")
	return 'new_filename.mp4'

iface = gr.Interface(
    replay,
    [
        gr.inputs.Dropdown(["Atari Space Invaders 👾", "CartPole-v1 🕹️", "LunarLander-v2 🚀👩‍🚀"]),
    ],
    "video",
     title = 'Stable Baselines 3 with 🤗',
            description = '',
             article = 
                        '''<div>
                        	<p style="text-align: center">This version of the RL library allows you to load models directly from the Hugging Face Hub</p>
                            <p style="text-align: center"> Select the trained agent you want to watch perform.
                            These models are from <a href="https://github.com/araffin/rl-baselines-zoo">Stable Baseline Zoo</a></p>
                            <p>
                            There are currently 3 models:
                            <ul>
                            	<li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-SpaceInvadersNoFrameskip-v4">PPO SpaceInvadersNoFrameskip-v4</a></li>
                            	<li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-LunarLander-v2">PPO LunarLander-v2</a></li>
                            	<li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-CartPole-v1">PPO CartPole-v1</a></li>
                            </ul>
                        </div>'''
            )
   

iface.launch()

"""
TODO: Next version with live video generation
import gradio as gr
import os

from Recorder import Recorder

from stable_baselines3 import PPO


#The Agent plays and we generate the video
def replay(option):
  video_path = ""
  # Get the correct model
  if (option == "LunarLander-v2 🚀👩‍🚀"):
    env_name = "Lunar Lander v2"
    agent_name = "PPO"
    print("TEST")
    hf_model_filename = "LunarLander-v2"
    hf_model_id = "ThomasSimonini/stable-baselines3-ppo-LunarLander-v2"
    video_path = replay_gym(hf_model_filename, hf_model_id)
  elif(option == "CartPole-v1 🕹️"):
      hf_model_filename = "CartPole-v1"
      hf_model_id = "ThomasSimonini/stable-baselines3-ppo-CartPole-v1"
      video_path = replay_gym(hf_model_filename, hf_model_id)
  elif(option == "Atari Space Invaders 👾"):
    hf_model_filename = "SpaceInvadersNoFrameskip-v4"
    hf_model_id = "ThomasSimonini/stable-baselines3-ppo-SpaceInvadersNoFrameskip-v4"
    video_path = replay_atari(hf_model_filename, hf_model_id)
      #video_path = "./SpaceInvadersNoFrameskip-v4.mp4"

  return video_path


def replay_gym(hf_model_filename, hf_model_id):
  import gym
  from stable_baselines3.common.evaluation import evaluate_policy


  model = PPO.load_from_huggingface(hf_model_id,hf_model_filename)

  eval_env = gym.make(hf_model_filename)

  directory = './video'
  env = Recorder(eval_env, directory)

  obs = env.reset()
  done = False
  while not done:
      action, _state = model.predict(obs)
      obs, reward, done, info = env.step(action)
  clip = env.play()
  return clip


def replay_atari(hf_model_filename, hf_model_id):
  os.system("python -m atari_py.import_roms \"content/atari_roms\"")
  import gym
  from stable_baselines3.common.env_util import make_atari_env
  from stable_baselines3.common.vec_env import VecFrameStack

  from stable_baselines3.common.evaluation import evaluate_policy

  model = PPO.load_from_huggingface(hf_model_id, hf_model_filename)


  eval_env = make_atari_env(hf_model_filename, n_envs=1, seed=0)
  eval_env = VecFrameStack(eval_env, n_stack=4)

  model = PPO.load_from_huggingface(hf_model_id, hf_model_filename)

  import gym
  directory = './video'
  env = Recorder(eval_env, directory)

  obs = env.reset()
  done = False
  while not done:
      action, _state = model.predict(obs)
      obs, reward, done, info = env.step(action)
  clip = env.play()
  return clip



iface = gr.Interface(
    replay,
    [
        gr.inputs.Dropdown(["Atari Space Invaders 👾", "CartPole-v1 🕹️", "LunarLander-v2 🚀👩‍🚀"]),
    ],
    "video",
     title = 'Stable Baselines 3 with 🤗',
            description = '',
             article = 
                        '''<div>
                          <p style="text-align: center">This version of the RL library allows you to load models directly from the Hugging Face Hub</p>
                            <p style="text-align: center"> Select the trained agent you want to watch perform. We record your agent playing.
                            <p style="text-align: center"> Don't forget to <b>click on clear between each record.</b> </p>
                            These models are from <a href="https://github.com/araffin/rl-baselines-zoo">Stable Baseline Zoo</a></p>
                            <p>
                            There are currently 3 models:
                            <ul>
                              <li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-SpaceInvadersNoFrameskip-v4">PPO SpaceInvadersNoFrameskip-v4</a></li>
                              <li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-LunarLander-v2">PPO LunarLander-v2</a></li>
                              <li><a href="https://huggingface.co/ThomasSimonini/stable-baselines3-ppo-CartPole-v1">PPO CartPole-v1</a></li>
                            </ul>
                        </div>'''
            )
   

iface.launch()
"""