app.py

import cv2
import gradio as gr
import time

from huggingface_sb3 import load_from_hub

from stable_baselines3 import PPO
from stable_baselines3.common.env_util import make_atari_env
from stable_baselines3.common.vec_env import VecFrameStack

from stable_baselines3.common.env_util import make_atari_env

max_steps = 5000 # Let's try with 5000 steps.

# Loading functions were taken from Edward Beeching code
def load_env(env_name):
    env = make_atari_env(env_name, n_envs=1)
    env = VecFrameStack(env, n_stack=4)
    return env

def load_model(env_name):
    custom_objects = {
        "learning_rate": 0.0,
        "lr_schedule": lambda _: 0.0,
        "clip_range": lambda _: 0.0,
    }

    checkpoint = load_from_hub(
        f"ThomasSimonini/ppo-{env_name}",
        f"ppo-{env_name}.zip",
    )

    model = PPO.load(checkpoint, custom_objects=custom_objects)

    return model
    
def replay(env_name, time_sleep):
  max_steps = 500
  env = load_env(env_name)
  model = load_model(env_name)
  #for i in range(num_episodes):
  obs = env.reset()
  done = False
  i = 0
  while not done:
    i+= 1
    if i < max_steps:
      frame = env.render(mode="rgb_array")
      action, _states = model.predict(obs)
      obs, reward, done, info = env.step([action])
      time.sleep(time_sleep)
      yield frame
    else:
      break

demo = gr.Interface(
    replay,
    [gr.Dropdown(["SpaceInvadersNoFrameskip-v4",
        "PongNoFrameskip-v4",
        "SeaquestNoFrameskip-v4",
        "QbertNoFrameskip-v4",
        ]),
     #gr.Slider(100, 10000, value=500),
     gr.Slider(0.01, 1, value=0.05),
     #gr.Slider(1, 20, value=5)
     ],
    gr.Image(shape=(300, 150)),
    title="Watch Agents playing Atari games 🤖",
    description="Select an environment to watch a Hugging Face's trained deep reinforcement learning agent.",
    article = "time_sleep is the time delay between each frame (0.05 by default)."
).launch().queue(max_concurrency=20, max_size=20)