app.py

import cv2
import streamlit as st
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

st.subheader("Atari 2600 Deep RL Environments Live AI")

# @st.cache This is not cachable :(
def load_env(env_name):
    env = make_atari_env(env_name, n_envs=1)
    env = VecFrameStack(env, n_stack=4)
    return env

# @st.cache This is not cachable :(
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
    
st.write("In game theory and optimization Nash Equilibrium loss minimization starts playing randomly but then by understanding ratios of action success to action-reward with an action (observe, decide/predict, act and then observe outcome the Deep RL agents go from 50% efficiency to 98-99% efficiency based on quality of decision without making mistakes. A good reference to environments is here https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/benchmark.md")
#st.write("Deep RL models: https://huggingface.co/sb3")

env_name = st.selectbox(
    "Select environment",
    (
        "SeaquestNoFrameskip-v4",
        "QbertNoFrameskip-v4",
        "SpaceInvadersNoFrameskip-v4",
        "PongNoFrameskip-v4",
        #"AsteroidsNoFrameskip-v4",
        #"BeamRiderNoFrameskip-v4",
        #"BreakoutNoFrameskip-v4	",
        #"EnduroNoFrameskip-v4",
        #"MsPacmanNoFrameskip-v4",
        #"RoadRunnerNoFrameskip-v4",
        #"Swimmer-v3",
        #"Walker2d-v3",
    ),
)

num_episodes = st.slider("Number of Episodes", 1, 20, 5)
env = load_env(env_name)
model = load_model(env_name)

obs = env.reset()

with st.empty():
    for i in range(num_episodes):
        obs = env.reset()
        done = False
        while not done:
            frame = env.render(mode="rgb_array")
            im = st.image(frame, width=400)
            action, _states = model.predict(obs)
            obs, reward, done, info = env.step([action])

            time.sleep(0.1)