{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "view-in-github"
},
"source": [
"![\"Open](\"https://colab.research.google.com/assets/colab-badge.svg\")
"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "-cf5-oDPjwf8"
},
"source": [
"# Unit 8: Proximal Policy Gradient (PPO) with PyTorch 🤖\n",
"\n",
"In this unit, you'll learn to **code your PPO agent from scratch with PyTorch**.\n",
"\n",
"To test its robustness, we're going to train it in 2 different classical environments:\n",
"\n",
"- [Cartpole-v1](https://www.gymlibrary.dev/environments/classic_control/cart_pole/?highlight=cartpole)\n",
"- [LunarLander-v2 🚀](https://www.gymlibrary.dev/environments/box2d/lunar_lander/)\n",
"\n",
"We finish this foundations part of the course by **understanding how PPO works deeply.** In the first Unit, you learned to train a PPO agent on LunarLander-v2. And now, with Unit 8, you can code it from scratch. **How incredible is that 🤩.**\n",
"\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "2Fl6Rxt0lc0O"
},
"source": [
"⬇️ Here is an example of what you will achieve. ⬇️"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "PYv9Br6slgR-"
},
"outputs": [],
"source": [
"%%html\n",
""
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "DbKfCj5ilgqT"
},
"outputs": [],
"source": [
"%%html\n",
""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "YcOFdWpnlxNf"
},
"source": [
"💡 We advise you to use Google Colab since some environments work only with Ubuntu. The free version of Google Colab is perfect for this tutorial. Let's get started 🚀"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "gkLf2KjRl8cm"
},
"source": [
"## This notebook is from Deep Reinforcement Learning Class\n",
"\n",
""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "kRZVF-uOl8_n"
},
"source": [
"In this free course, you will:\n",
"\n",
"- 📖 Study Deep Reinforcement Learning in **theory and practice**.\n",
"- 🧑💻 Learn to **use famous Deep RL libraries** such as Stable Baselines3, RL Baselines3 Zoo, and RLlib.\n",
"- 🤖 Train **agents in unique environments** \n",
"\n",
"And more check 📚 the syllabus 👉 https://github.com/huggingface/deep-rl-class\n",
"\n",
"The best way to keep in touch is to join our discord server to exchange with the community and with us 👉🏻 https://discord.gg/aYka4Yhff9"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "AMuTi9xgl_tS"
},
"source": [
"## Prerequisites 🏗️\n",
"Before diving into the notebook, you need to:\n",
"\n",
"🔲 📚 [Read the Unit 8 Readme](https://github.com/huggingface/deep-rl-class/blob/main/unit8/README.md) contains all the information.\n",
"\n",
"🔲 📚 **Study Proximal Policy Optimization (PPO)** by reading the chapter 👉 https://huggingface.co/blog/deep-rl-ppo"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "dQKcncVkmFtE"
},
"source": [
"### Step 0: Set the GPU 💪\n",
"- To **faster the agent's training, we'll use a GPU** to do that go to `Runtime > Change Runtime type`\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "7Ov87REzmHkU"
},
"source": [
"- `Hardware Accelerator > GPU`"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "tjcQVcC-mK3p"
},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "ncIgfNf3mOtc"
},
"source": [
"### Step 1: Install dependencies 🔽 and virtual screen 💻\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "ixxlRGmhM6gY"
},
"outputs": [],
"source": [
"!pip install imageio-ffmpeg\n",
"!pip install huggingface_hub\n",
"!pip install gym[box2d]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "sQ2_tjvxM6mP"
},
"outputs": [],
"source": [
"!apt install python-opengl\n",
"!apt install ffmpeg\n",
"!apt install xvfb\n",
"!pip3 install pyvirtualdisplay\n",
"\n",
"# Virtual display\n",
"from pyvirtualdisplay import Display\n",
"\n",
"virtual_display = Display(visible=0, size=(500, 500))\n",
"virtual_display.start()"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "oDkUufewmq6v"
},
"source": [
"### Step 2: Let's code PPO from scratch with Costa Huang tutorial\n",
"- For the core implementation of PPO we're going to use the excellent [Costa Huang](https://costa.sh/) tutorial.\n",
"- In addition to the tutorial, to go deeper you can read the 13 core implementation details: https://iclr-blog-track.github.io/2022/03/25/ppo-implementation-details/\n",
"\n",
"👉 The video tutorial: https://youtu.be/MEt6rrxH8W4"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "aNgEL1_uvhaq"
},
"outputs": [],
"source": [
"from IPython.display import HTML\n",
"\n",
"HTML('')"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "f34ILn7AvTbt"
},
"source": [
"- The best is to code first on the cell below, this way, if you kill the machine **you don't loose the implementation**."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "_bE708C6mhE7"
},
"outputs": [],
"source": [
"### Your code here:"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "mk-a9CmNuS2W"
},
"source": [
"### Step 3: Add the Hugging Face Integration 🤗\n",
"- In order to push our model to the Hub, we need to define a function `package_to_hub`"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "TPi1Nme-oGWd"
},
"source": [
"- Add dependencies we need to push our model to the Hub"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Sj8bz-AmoNVj"
},
"outputs": [],
"source": [
"from huggingface_hub import HfApi, upload_folder\n",
"from huggingface_hub.repocard import metadata_eval_result, metadata_save\n",
"\n",
"from pathlib import Path\n",
"import datetime\n",
"import tempfile\n",
"import json\n",
"import shutil\n",
"import imageio\n",
"\n",
"from wasabi import Printer\n",
"msg = Printer()"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "5rDr8-lWn0zi"
},
"source": [
"- Add new argument in `parse_args()` function to define the repo-id where we want to push the model."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "iHQiqQEFn0QH"
},
"outputs": [],
"source": [
"# Adding HuggingFace argument\n",
"parser.add_argument(\"--repo-id\", type=str, default=\"ThomasSimonini/ppo-CartPole-v1\", help=\"id of the model repository from the Hugging Face Hub {username/repo_name}\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "blLZMiBAoUVT"
},
"source": [
"- Next, we add the methods needed to push the model to the Hub\n",
"\n",
"- These methods will:\n",
" - `_evalutate_agent()`: evaluate the agent.\n",
" - `_generate_model_card()`: generate the model card of your agent.\n",
" - `_record_video()`: record a video of your agent."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "WlLcz4L9odXs"
},
"outputs": [],
"source": [
"def package_to_hub(repo_id, \n",
" model,\n",
" hyperparameters,\n",
" eval_env,\n",
" video_fps=30,\n",
" commit_message=\"Push Reinforce agent to the Hub\",\n",
" token= None,\n",
" logs=None\n",
" ):\n",
" \"\"\"\n",
" Evaluate, Generate a video and Upload a model to Hugging Face Hub.\n",
" This method does the complete pipeline:\n",
" - It evaluates the model\n",
" - It generates the model card\n",
" - It generates a replay video of the agent\n",
" - It pushes everything to the hub\n",
" :param repo_id: id of the model repository from the Hugging Face Hub\n",
" :param model: trained model\n",
" :param eval_env: environment used to evaluate the agent\n",
" :param fps: number of fps for rendering the video\n",
" :param commit_message: commit message\n",
" :param logs: directory on local machine of tensorboard logs you'd like to upload\n",
" \"\"\"\n",
" msg.info(\n",
" \"This function will save, evaluate, generate a video of your agent, \"\n",
" \"create a model card and push everything to the hub. \"\n",
" \"It might take up to 1min. \\n \"\n",
" \"This is a work in progress: if you encounter a bug, please open an issue.\"\n",
" )\n",
" # Step 1: Clone or create the repo\n",
" repo_url = HfApi().create_repo(\n",
" repo_id=repo_id,\n",
" token=token,\n",
" private=False,\n",
" exist_ok=True,\n",
" )\n",
" \n",
" with tempfile.TemporaryDirectory() as tmpdirname:\n",
" tmpdirname = Path(tmpdirname)\n",
"\n",
" # Step 2: Save the model\n",
" torch.save(model.state_dict(), tmpdirname / \"model.pt\")\n",
" \n",
" # Step 3: Evaluate the model and build JSON\n",
" mean_reward, std_reward = _evaluate_agent(eval_env, \n",
" 10, \n",
" model)\n",
"\n",
" # First get datetime\n",
" eval_datetime = datetime.datetime.now()\n",
" eval_form_datetime = eval_datetime.isoformat()\n",
"\n",
" evaluate_data = {\n",
" \"env_id\": hyperparameters.env_id, \n",
" \"mean_reward\": mean_reward,\n",
" \"std_reward\": std_reward,\n",
" \"n_evaluation_episodes\": 10,\n",
" \"eval_datetime\": eval_form_datetime,\n",
" }\n",
" \n",
" # Write a JSON file\n",
" with open(tmpdirname / \"results.json\", \"w\") as outfile:\n",
" json.dump(evaluate_data, outfile)\n",
"\n",
" # Step 4: Generate a video\n",
" video_path = tmpdirname / \"replay.mp4\"\n",
" record_video(eval_env, model, video_path, video_fps)\n",
" \n",
" # Step 5: Generate the model card\n",
" generated_model_card, metadata = _generate_model_card(\"PPO\", hyperparameters.env_id, mean_reward, std_reward, hyperparameters)\n",
" _save_model_card(tmpdirname, generated_model_card, metadata)\n",
"\n",
" # Step 6: Add logs if needed\n",
" if logs:\n",
" _add_logdir(tmpdirname, Path(logs))\n",
" \n",
" msg.info(f\"Pushing repo {repo_id} to the Hugging Face Hub\")\n",
" \n",
" repo_url = upload_folder(\n",
" repo_id=repo_id,\n",
" folder_path=tmpdirname,\n",
" path_in_repo=\"\",\n",
" commit_message=commit_message,\n",
" token=token,\n",
" )\n",
"\n",
" msg.info(f\"Your model is pushed to the Hub. You can view your model here: {repo_url}\")\n",
" return repo_url\n",
"\n",
"def _evaluate_agent(env, n_eval_episodes, policy):\n",
" \"\"\"\n",
" Evaluate the agent for ``n_eval_episodes`` episodes and returns average reward and std of reward.\n",
" :param env: The evaluation environment\n",
" :param n_eval_episodes: Number of episode to evaluate the agent\n",
" :param policy: The Reinforce agent\n",
" \"\"\"\n",
" episode_rewards = []\n",
" for episode in range(n_eval_episodes):\n",
" state = env.reset()\n",
" step = 0\n",
" done = False\n",
" total_rewards_ep = 0\n",
" \n",
" while done is False:\n",
" state = torch.Tensor(state)\n",
" action, _, _, _ = policy.get_action_and_value(state)\n",
" new_state, reward, done, info = env.step(action.numpy())\n",
" total_rewards_ep += reward \n",
" if done:\n",
" break\n",
" state = new_state\n",
" episode_rewards.append(total_rewards_ep)\n",
" mean_reward = np.mean(episode_rewards)\n",
" std_reward = np.std(episode_rewards)\n",
"\n",
" return mean_reward, std_reward\n",
"\n",
"def record_video(env, policy, out_directory, fps=30):\n",
" images = [] \n",
" done = False\n",
" state = env.reset()\n",
" img = env.render(mode='rgb_array')\n",
" images.append(img)\n",
" while not done:\n",
" state = torch.Tensor(state)\n",
" # Take the action (index) that have the maximum expected future reward given that state\n",
" action, _, _, _ = policy.get_action_and_value(state)\n",
" state, reward, done, info = env.step(action.numpy()) # We directly put next_state = state for recording logic\n",
" img = env.render(mode='rgb_array')\n",
" images.append(img)\n",
" imageio.mimsave(out_directory, [np.array(img) for i, img in enumerate(images)], fps=fps)\n",
"\n",
"def _generate_model_card(model_name, env_id, mean_reward, std_reward, hyperparameters):\n",
" \"\"\"\n",
" Generate the model card for the Hub\n",
" :param model_name: name of the model\n",
" :env_id: name of the environment\n",
" :mean_reward: mean reward of the agent\n",
" :std_reward: standard deviation of the mean reward of the agent\n",
" :hyperparameters: training arguments\n",
" \"\"\"\n",
" # Step 1: Select the tags\n",
" metadata = generate_metadata(model_name, env_id, mean_reward, std_reward)\n",
"\n",
" # Transform the hyperparams namespace to string\n",
" converted_dict = vars(hyperparameters)\n",
" converted_str = str(converted_dict)\n",
" converted_str = converted_str.split(\", \")\n",
" converted_str = '\\n'.join(converted_str)\n",
" \n",
" # Step 2: Generate the model card\n",
" model_card = f\"\"\"\n",
" # PPO Agent Playing {env_id}\n",
"\n",
" This is a trained model of a PPO agent playing {env_id}.\n",
" To learn to code your own PPO agent and train it Unit 8 of the Deep Reinforcement Learning Class: https://github.com/huggingface/deep-rl-class/tree/main/unit8\n",
" \n",
" # Hyperparameters\n",
" ```python\n",
" {converted_str}\n",
" ```\n",
" \"\"\"\n",
" return model_card, metadata\n",
"\n",
"def generate_metadata(model_name, env_id, mean_reward, std_reward):\n",
" \"\"\"\n",
" Define the tags for the model card\n",
" :param model_name: name of the model\n",
" :param env_id: name of the environment\n",
" :mean_reward: mean reward of the agent\n",
" :std_reward: standard deviation of the mean reward of the agent\n",
" \"\"\"\n",
" metadata = {}\n",
" metadata[\"tags\"] = [\n",
" env_id,\n",
" \"ppo\",\n",
" \"deep-reinforcement-learning\",\n",
" \"reinforcement-learning\",\n",
" \"custom-implementation\",\n",
" \"deep-rl-class\"\n",
" ]\n",
"\n",
" # Add metrics\n",
" eval = metadata_eval_result(\n",
" model_pretty_name=model_name,\n",
" task_pretty_name=\"reinforcement-learning\",\n",
" task_id=\"reinforcement-learning\",\n",
" metrics_pretty_name=\"mean_reward\",\n",
" metrics_id=\"mean_reward\",\n",
" metrics_value=f\"{mean_reward:.2f} +/- {std_reward:.2f}\",\n",
" dataset_pretty_name=env_id,\n",
" dataset_id=env_id,\n",
" )\n",
"\n",
" # Merges both dictionaries\n",
" metadata = {**metadata, **eval}\n",
"\n",
" return metadata\n",
"\n",
"def _save_model_card(local_path, generated_model_card, metadata):\n",
" \"\"\"Saves a model card for the repository.\n",
" :param local_path: repository directory\n",
" :param generated_model_card: model card generated by _generate_model_card()\n",
" :param metadata: metadata\n",
" \"\"\"\n",
" readme_path = local_path / \"README.md\"\n",
" readme = \"\"\n",
" if readme_path.exists():\n",
" with readme_path.open(\"r\", encoding=\"utf8\") as f:\n",
" readme = f.read()\n",
" else:\n",
" readme = generated_model_card\n",
"\n",
" with readme_path.open(\"w\", encoding=\"utf-8\") as f:\n",
" f.write(readme)\n",
"\n",
" # Save our metrics to Readme metadata\n",
" metadata_save(readme_path, metadata)\n",
"\n",
"def _add_logdir(local_path: Path, logdir: Path):\n",
" \"\"\"Adds a logdir to the repository.\n",
" :param local_path: repository directory\n",
" :param logdir: logdir directory\n",
" \"\"\"\n",
" if logdir.exists() and logdir.is_dir():\n",
" # Add the logdir to the repository under new dir called logs\n",
" repo_logdir = local_path / \"logs\"\n",
" \n",
" # Delete current logs if they exist\n",
" if repo_logdir.exists():\n",
" shutil.rmtree(repo_logdir)\n",
"\n",
" # Copy logdir into repo logdir\n",
" shutil.copytree(logdir, repo_logdir)"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "TqX8z8_rooD6"
},
"source": [
"- Finally, we call this function at the end of the PPO training"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "I8V1vNiTo2hL"
},
"outputs": [],
"source": [
"# Create the evaluation environment\n",
"eval_env = gym.make(args.env_id)\n",
"\n",
"package_to_hub(repo_id = args.repo_id,\n",
" model = agent, # The model we want to save\n",
" hyperparameters = args,\n",
" eval_env = gym.make(args.env_id),\n",
" logs= f\"runs/{run_name}\",\n",
" )"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "muCCzed4o5TC"
},
"source": [
"- Here's what look the ppo.py final file"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "LviRdtXgo7kF"
},
"outputs": [],
"source": [
"# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/ppo/#ppopy\n",
"\n",
"import argparse\n",
"import os\n",
"import random\n",
"import time\n",
"from distutils.util import strtobool\n",
"\n",
"import gym\n",
"import numpy as np\n",
"import torch\n",
"import torch.nn as nn\n",
"import torch.optim as optim\n",
"from torch.distributions.categorical import Categorical\n",
"from torch.utils.tensorboard import SummaryWriter\n",
"\n",
"from huggingface_hub import HfApi, upload_folder\n",
"from huggingface_hub.repocard import metadata_eval_result, metadata_save\n",
"\n",
"from pathlib import Path\n",
"import datetime\n",
"import tempfile\n",
"import json\n",
"import shutil\n",
"import imageio\n",
"\n",
"from wasabi import Printer\n",
"msg = Printer()\n",
"\n",
"def parse_args():\n",
" # fmt: off\n",
" parser = argparse.ArgumentParser()\n",
" parser.add_argument(\"--exp-name\", type=str, default=os.path.basename(__file__).rstrip(\".py\"),\n",
" help=\"the name of this experiment\")\n",
" parser.add_argument(\"--seed\", type=int, default=1,\n",
" help=\"seed of the experiment\")\n",
" parser.add_argument(\"--torch-deterministic\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"if toggled, `torch.backends.cudnn.deterministic=False`\")\n",
" parser.add_argument(\"--cuda\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"if toggled, cuda will be enabled by default\")\n",
" parser.add_argument(\"--track\", type=lambda x: bool(strtobool(x)), default=False, nargs=\"?\", const=True,\n",
" help=\"if toggled, this experiment will be tracked with Weights and Biases\")\n",
" parser.add_argument(\"--wandb-project-name\", type=str, default=\"cleanRL\",\n",
" help=\"the wandb's project name\")\n",
" parser.add_argument(\"--wandb-entity\", type=str, default=None,\n",
" help=\"the entity (team) of wandb's project\")\n",
" parser.add_argument(\"--capture-video\", type=lambda x: bool(strtobool(x)), default=False, nargs=\"?\", const=True,\n",
" help=\"weather to capture videos of the agent performances (check out `videos` folder)\")\n",
"\n",
" # Algorithm specific arguments\n",
" parser.add_argument(\"--env-id\", type=str, default=\"CartPole-v1\",\n",
" help=\"the id of the environment\")\n",
" parser.add_argument(\"--total-timesteps\", type=int, default=50000,\n",
" help=\"total timesteps of the experiments\")\n",
" parser.add_argument(\"--learning-rate\", type=float, default=2.5e-4,\n",
" help=\"the learning rate of the optimizer\")\n",
" parser.add_argument(\"--num-envs\", type=int, default=4,\n",
" help=\"the number of parallel game environments\")\n",
" parser.add_argument(\"--num-steps\", type=int, default=128,\n",
" help=\"the number of steps to run in each environment per policy rollout\")\n",
" parser.add_argument(\"--anneal-lr\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"Toggle learning rate annealing for policy and value networks\")\n",
" parser.add_argument(\"--gae\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"Use GAE for advantage computation\")\n",
" parser.add_argument(\"--gamma\", type=float, default=0.99,\n",
" help=\"the discount factor gamma\")\n",
" parser.add_argument(\"--gae-lambda\", type=float, default=0.95,\n",
" help=\"the lambda for the general advantage estimation\")\n",
" parser.add_argument(\"--num-minibatches\", type=int, default=4,\n",
" help=\"the number of mini-batches\")\n",
" parser.add_argument(\"--update-epochs\", type=int, default=4,\n",
" help=\"the K epochs to update the policy\")\n",
" parser.add_argument(\"--norm-adv\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"Toggles advantages normalization\")\n",
" parser.add_argument(\"--clip-coef\", type=float, default=0.2,\n",
" help=\"the surrogate clipping coefficient\")\n",
" parser.add_argument(\"--clip-vloss\", type=lambda x: bool(strtobool(x)), default=True, nargs=\"?\", const=True,\n",
" help=\"Toggles whether or not to use a clipped loss for the value function, as per the paper.\")\n",
" parser.add_argument(\"--ent-coef\", type=float, default=0.01,\n",
" help=\"coefficient of the entropy\")\n",
" parser.add_argument(\"--vf-coef\", type=float, default=0.5,\n",
" help=\"coefficient of the value function\")\n",
" parser.add_argument(\"--max-grad-norm\", type=float, default=0.5,\n",
" help=\"the maximum norm for the gradient clipping\")\n",
" parser.add_argument(\"--target-kl\", type=float, default=None,\n",
" help=\"the target KL divergence threshold\")\n",
" \n",
" # Adding HuggingFace argument\n",
" parser.add_argument(\"--repo-id\", type=str, default=\"ThomasSimonini/ppo-CartPole-v1\", help=\"id of the model repository from the Hugging Face Hub {username/repo_name}\")\n",
"\n",
" args = parser.parse_args()\n",
" args.batch_size = int(args.num_envs * args.num_steps)\n",
" args.minibatch_size = int(args.batch_size // args.num_minibatches)\n",
" # fmt: on\n",
" return args\n",
"\n",
"def package_to_hub(repo_id, \n",
" model,\n",
" hyperparameters,\n",
" eval_env,\n",
" video_fps=30,\n",
" commit_message=\"Push Reinforce agent to the Hub\",\n",
" token= None,\n",
" logs=None\n",
" ):\n",
" \"\"\"\n",
" Evaluate, Generate a video and Upload a model to Hugging Face Hub.\n",
" This method does the complete pipeline:\n",
" - It evaluates the model\n",
" - It generates the model card\n",
" - It generates a replay video of the agent\n",
" - It pushes everything to the hub\n",
" :param repo_id: id of the model repository from the Hugging Face Hub\n",
" :param model: trained model\n",
" :param eval_env: environment used to evaluate the agent\n",
" :param fps: number of fps for rendering the video\n",
" :param commit_message: commit message\n",
" :param logs: directory on local machine of tensorboard logs you'd like to upload\n",
" \"\"\"\n",
" msg.info(\n",
" \"This function will save, evaluate, generate a video of your agent, \"\n",
" \"create a model card and push everything to the hub. \"\n",
" \"It might take up to 1min. \\n \"\n",
" \"This is a work in progress: if you encounter a bug, please open an issue.\"\n",
" )\n",
" # Step 1: Clone or create the repo\n",
" repo_url = HfApi().create_repo(\n",
" repo_id=repo_id,\n",
" token=token,\n",
" private=False,\n",
" exist_ok=True,\n",
" )\n",
" \n",
" with tempfile.TemporaryDirectory() as tmpdirname:\n",
" tmpdirname = Path(tmpdirname)\n",
"\n",
" # Step 2: Save the model\n",
" torch.save(model.state_dict(), tmpdirname / \"model.pt\")\n",
" \n",
" # Step 3: Evaluate the model and build JSON\n",
" mean_reward, std_reward = _evaluate_agent(eval_env, \n",
" 10, \n",
" model)\n",
"\n",
" # First get datetime\n",
" eval_datetime = datetime.datetime.now()\n",
" eval_form_datetime = eval_datetime.isoformat()\n",
"\n",
" evaluate_data = {\n",
" \"env_id\": hyperparameters.env_id, \n",
" \"mean_reward\": mean_reward,\n",
" \"std_reward\": std_reward,\n",
" \"n_evaluation_episodes\": 10,\n",
" \"eval_datetime\": eval_form_datetime,\n",
" }\n",
" \n",
" # Write a JSON file\n",
" with open(tmpdirname / \"results.json\", \"w\") as outfile:\n",
" json.dump(evaluate_data, outfile)\n",
"\n",
" # Step 4: Generate a video\n",
" video_path = tmpdirname / \"replay.mp4\"\n",
" record_video(eval_env, model, video_path, video_fps)\n",
" \n",
" # Step 5: Generate the model card\n",
" generated_model_card, metadata = _generate_model_card(\"PPO\", hyperparameters.env_id, mean_reward, std_reward, hyperparameters)\n",
" _save_model_card(tmpdirname, generated_model_card, metadata)\n",
"\n",
" # Step 6: Add logs if needed\n",
" if logs:\n",
" _add_logdir(tmpdirname, Path(logs))\n",
" \n",
" msg.info(f\"Pushing repo {repo_id} to the Hugging Face Hub\")\n",
" \n",
" repo_url = upload_folder(\n",
" repo_id=repo_id,\n",
" folder_path=tmpdirname,\n",
" path_in_repo=\"\",\n",
" commit_message=commit_message,\n",
" token=token,\n",
" )\n",
"\n",
" msg.info(f\"Your model is pushed to the Hub. You can view your model here: {repo_url}\")\n",
" return repo_url\n",
"\n",
"def _evaluate_agent(env, n_eval_episodes, policy):\n",
" \"\"\"\n",
" Evaluate the agent for ``n_eval_episodes`` episodes and returns average reward and std of reward.\n",
" :param env: The evaluation environment\n",
" :param n_eval_episodes: Number of episode to evaluate the agent\n",
" :param policy: The Reinforce agent\n",
" \"\"\"\n",
" episode_rewards = []\n",
" for episode in range(n_eval_episodes):\n",
" state = env.reset()\n",
" step = 0\n",
" done = False\n",
" total_rewards_ep = 0\n",
" \n",
" while done is False:\n",
" state = torch.Tensor(state)\n",
" action, _, _, _ = policy.get_action_and_value(state)\n",
" new_state, reward, done, info = env.step(action.numpy())\n",
" total_rewards_ep += reward \n",
" if done:\n",
" break\n",
" state = new_state\n",
" episode_rewards.append(total_rewards_ep)\n",
" mean_reward = np.mean(episode_rewards)\n",
" std_reward = np.std(episode_rewards)\n",
"\n",
" return mean_reward, std_reward\n",
"\n",
"def record_video(env, policy, out_directory, fps=30):\n",
" images = [] \n",
" done = False\n",
" state = env.reset()\n",
" img = env.render(mode='rgb_array')\n",
" images.append(img)\n",
" while not done:\n",
" state = torch.Tensor(state)\n",
" # Take the action (index) that have the maximum expected future reward given that state\n",
" action, _, _, _ = policy.get_action_and_value(state)\n",
" state, reward, done, info = env.step(action.numpy()) # We directly put next_state = state for recording logic\n",
" img = env.render(mode='rgb_array')\n",
" images.append(img)\n",
" imageio.mimsave(out_directory, [np.array(img) for i, img in enumerate(images)], fps=fps)\n",
"\n",
"def _generate_model_card(model_name, env_id, mean_reward, std_reward, hyperparameters):\n",
" \"\"\"\n",
" Generate the model card for the Hub\n",
" :param model_name: name of the model\n",
" :env_id: name of the environment\n",
" :mean_reward: mean reward of the agent\n",
" :std_reward: standard deviation of the mean reward of the agent\n",
" :hyperparameters: training arguments\n",
" \"\"\"\n",
" # Step 1: Select the tags\n",
" metadata = generate_metadata(model_name, env_id, mean_reward, std_reward)\n",
"\n",
" # Transform the hyperparams namespace to string\n",
" converted_dict = vars(hyperparameters)\n",
" converted_str = str(converted_dict)\n",
" converted_str = converted_str.split(\", \")\n",
" converted_str = '\\n'.join(converted_str)\n",
" \n",
" # Step 2: Generate the model card\n",
" model_card = f\"\"\"\n",
" # PPO Agent Playing {env_id}\n",
"\n",
" This is a trained model of a PPO agent playing {env_id}.\n",
" To learn to code your own PPO agent and train it Unit 8 of the Deep Reinforcement Learning Class: https://github.com/huggingface/deep-rl-class/tree/main/unit8\n",
" \n",
" # Hyperparameters\n",
" ```python\n",
" {converted_str}\n",
" ```\n",
" \"\"\"\n",
" return model_card, metadata\n",
"\n",
"def generate_metadata(model_name, env_id, mean_reward, std_reward):\n",
" \"\"\"\n",
" Define the tags for the model card\n",
" :param model_name: name of the model\n",
" :param env_id: name of the environment\n",
" :mean_reward: mean reward of the agent\n",
" :std_reward: standard deviation of the mean reward of the agent\n",
" \"\"\"\n",
" metadata = {}\n",
" metadata[\"tags\"] = [\n",
" env_id,\n",
" \"ppo\",\n",
" \"deep-reinforcement-learning\",\n",
" \"reinforcement-learning\",\n",
" \"custom-implementation\",\n",
" \"deep-rl-class\"\n",
" ]\n",
"\n",
" # Add metrics\n",
" eval = metadata_eval_result(\n",
" model_pretty_name=model_name,\n",
" task_pretty_name=\"reinforcement-learning\",\n",
" task_id=\"reinforcement-learning\",\n",
" metrics_pretty_name=\"mean_reward\",\n",
" metrics_id=\"mean_reward\",\n",
" metrics_value=f\"{mean_reward:.2f} +/- {std_reward:.2f}\",\n",
" dataset_pretty_name=env_id,\n",
" dataset_id=env_id,\n",
" )\n",
"\n",
" # Merges both dictionaries\n",
" metadata = {**metadata, **eval}\n",
"\n",
" return metadata\n",
"\n",
"def _save_model_card(local_path, generated_model_card, metadata):\n",
" \"\"\"Saves a model card for the repository.\n",
" :param local_path: repository directory\n",
" :param generated_model_card: model card generated by _generate_model_card()\n",
" :param metadata: metadata\n",
" \"\"\"\n",
" readme_path = local_path / \"README.md\"\n",
" readme = \"\"\n",
" if readme_path.exists():\n",
" with readme_path.open(\"r\", encoding=\"utf8\") as f:\n",
" readme = f.read()\n",
" else:\n",
" readme = generated_model_card\n",
"\n",
" with readme_path.open(\"w\", encoding=\"utf-8\") as f:\n",
" f.write(readme)\n",
"\n",
" # Save our metrics to Readme metadata\n",
" metadata_save(readme_path, metadata)\n",
"\n",
"def _add_logdir(local_path: Path, logdir: Path):\n",
" \"\"\"Adds a logdir to the repository.\n",
" :param local_path: repository directory\n",
" :param logdir: logdir directory\n",
" \"\"\"\n",
" if logdir.exists() and logdir.is_dir():\n",
" # Add the logdir to the repository under new dir called logs\n",
" repo_logdir = local_path / \"logs\"\n",
" \n",
" # Delete current logs if they exist\n",
" if repo_logdir.exists():\n",
" shutil.rmtree(repo_logdir)\n",
"\n",
" # Copy logdir into repo logdir\n",
" shutil.copytree(logdir, repo_logdir)\n",
"\n",
"def make_env(env_id, seed, idx, capture_video, run_name):\n",
" def thunk():\n",
" env = gym.make(env_id)\n",
" env = gym.wrappers.RecordEpisodeStatistics(env)\n",
" if capture_video:\n",
" if idx == 0:\n",
" env = gym.wrappers.RecordVideo(env, f\"videos/{run_name}\")\n",
" env.seed(seed)\n",
" env.action_space.seed(seed)\n",
" env.observation_space.seed(seed)\n",
" return env\n",
"\n",
" return thunk\n",
"\n",
"\n",
"def layer_init(layer, std=np.sqrt(2), bias_const=0.0):\n",
" torch.nn.init.orthogonal_(layer.weight, std)\n",
" torch.nn.init.constant_(layer.bias, bias_const)\n",
" return layer\n",
"\n",
"\n",
"class Agent(nn.Module):\n",
" def __init__(self, envs):\n",
" super().__init__()\n",
" self.critic = nn.Sequential(\n",
" layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),\n",
" nn.Tanh(),\n",
" layer_init(nn.Linear(64, 64)),\n",
" nn.Tanh(),\n",
" layer_init(nn.Linear(64, 1), std=1.0),\n",
" )\n",
" self.actor = nn.Sequential(\n",
" layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),\n",
" nn.Tanh(),\n",
" layer_init(nn.Linear(64, 64)),\n",
" nn.Tanh(),\n",
" layer_init(nn.Linear(64, envs.single_action_space.n), std=0.01),\n",
" )\n",
"\n",
" def get_value(self, x):\n",
" return self.critic(x)\n",
"\n",
" def get_action_and_value(self, x, action=None):\n",
" logits = self.actor(x)\n",
" probs = Categorical(logits=logits)\n",
" if action is None:\n",
" action = probs.sample()\n",
" return action, probs.log_prob(action), probs.entropy(), self.critic(x)\n",
"\n",
"\n",
"if __name__ == \"__main__\":\n",
" args = parse_args()\n",
" run_name = f\"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}\"\n",
" if args.track:\n",
" import wandb\n",
"\n",
" wandb.init(\n",
" project=args.wandb_project_name,\n",
" entity=args.wandb_entity,\n",
" sync_tensorboard=True,\n",
" config=vars(args),\n",
" name=run_name,\n",
" monitor_gym=True,\n",
" save_code=True,\n",
" )\n",
" writer = SummaryWriter(f\"runs/{run_name}\")\n",
" writer.add_text(\n",
" \"hyperparameters\",\n",
" \"|param|value|\\n|-|-|\\n%s\" % (\"\\n\".join([f\"|{key}|{value}|\" for key, value in vars(args).items()])),\n",
" )\n",
"\n",
" # TRY NOT TO MODIFY: seeding\n",
" random.seed(args.seed)\n",
" np.random.seed(args.seed)\n",
" torch.manual_seed(args.seed)\n",
" torch.backends.cudnn.deterministic = args.torch_deterministic\n",
"\n",
" device = torch.device(\"cuda\" if torch.cuda.is_available() and args.cuda else \"cpu\")\n",
"\n",
" # env setup\n",
" envs = gym.vector.SyncVectorEnv(\n",
" [make_env(args.env_id, args.seed + i, i, args.capture_video, run_name) for i in range(args.num_envs)]\n",
" )\n",
" assert isinstance(envs.single_action_space, gym.spaces.Discrete), \"only discrete action space is supported\"\n",
"\n",
" agent = Agent(envs).to(device)\n",
" optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)\n",
"\n",
" # ALGO Logic: Storage setup\n",
" obs = torch.zeros((args.num_steps, args.num_envs) + envs.single_observation_space.shape).to(device)\n",
" actions = torch.zeros((args.num_steps, args.num_envs) + envs.single_action_space.shape).to(device)\n",
" logprobs = torch.zeros((args.num_steps, args.num_envs)).to(device)\n",
" rewards = torch.zeros((args.num_steps, args.num_envs)).to(device)\n",
" dones = torch.zeros((args.num_steps, args.num_envs)).to(device)\n",
" values = torch.zeros((args.num_steps, args.num_envs)).to(device)\n",
"\n",
" # TRY NOT TO MODIFY: start the game\n",
" global_step = 0\n",
" start_time = time.time()\n",
" next_obs = torch.Tensor(envs.reset()).to(device)\n",
" next_done = torch.zeros(args.num_envs).to(device)\n",
" num_updates = args.total_timesteps // args.batch_size\n",
"\n",
" for update in range(1, num_updates + 1):\n",
" # Annealing the rate if instructed to do so.\n",
" if args.anneal_lr:\n",
" frac = 1.0 - (update - 1.0) / num_updates\n",
" lrnow = frac * args.learning_rate\n",
" optimizer.param_groups[0][\"lr\"] = lrnow\n",
"\n",
" for step in range(0, args.num_steps):\n",
" global_step += 1 * args.num_envs\n",
" obs[step] = next_obs\n",
" dones[step] = next_done\n",
"\n",
" # ALGO LOGIC: action logic\n",
" with torch.no_grad():\n",
" action, logprob, _, value = agent.get_action_and_value(next_obs)\n",
" values[step] = value.flatten()\n",
" actions[step] = action\n",
" logprobs[step] = logprob\n",
"\n",
" # TRY NOT TO MODIFY: execute the game and log data.\n",
" next_obs, reward, done, info = envs.step(action.cpu().numpy())\n",
" rewards[step] = torch.tensor(reward).to(device).view(-1)\n",
" next_obs, next_done = torch.Tensor(next_obs).to(device), torch.Tensor(done).to(device)\n",
"\n",
" for item in info:\n",
" if \"episode\" in item.keys():\n",
" print(f\"global_step={global_step}, episodic_return={item['episode']['r']}\")\n",
" writer.add_scalar(\"charts/episodic_return\", item[\"episode\"][\"r\"], global_step)\n",
" writer.add_scalar(\"charts/episodic_length\", item[\"episode\"][\"l\"], global_step)\n",
" break\n",
"\n",
" # bootstrap value if not done\n",
" with torch.no_grad():\n",
" next_value = agent.get_value(next_obs).reshape(1, -1)\n",
" if args.gae:\n",
" advantages = torch.zeros_like(rewards).to(device)\n",
" lastgaelam = 0\n",
" for t in reversed(range(args.num_steps)):\n",
" if t == args.num_steps - 1:\n",
" nextnonterminal = 1.0 - next_done\n",
" nextvalues = next_value\n",
" else:\n",
" nextnonterminal = 1.0 - dones[t + 1]\n",
" nextvalues = values[t + 1]\n",
" delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t]\n",
" advantages[t] = lastgaelam = delta + args.gamma * args.gae_lambda * nextnonterminal * lastgaelam\n",
" returns = advantages + values\n",
" else:\n",
" returns = torch.zeros_like(rewards).to(device)\n",
" for t in reversed(range(args.num_steps)):\n",
" if t == args.num_steps - 1:\n",
" nextnonterminal = 1.0 - next_done\n",
" next_return = next_value\n",
" else:\n",
" nextnonterminal = 1.0 - dones[t + 1]\n",
" next_return = returns[t + 1]\n",
" returns[t] = rewards[t] + args.gamma * nextnonterminal * next_return\n",
" advantages = returns - values\n",
"\n",
" # flatten the batch\n",
" b_obs = obs.reshape((-1,) + envs.single_observation_space.shape)\n",
" b_logprobs = logprobs.reshape(-1)\n",
" b_actions = actions.reshape((-1,) + envs.single_action_space.shape)\n",
" b_advantages = advantages.reshape(-1)\n",
" b_returns = returns.reshape(-1)\n",
" b_values = values.reshape(-1)\n",
"\n",
" # Optimizing the policy and value network\n",
" b_inds = np.arange(args.batch_size)\n",
" clipfracs = []\n",
" for epoch in range(args.update_epochs):\n",
" np.random.shuffle(b_inds)\n",
" for start in range(0, args.batch_size, args.minibatch_size):\n",
" end = start + args.minibatch_size\n",
" mb_inds = b_inds[start:end]\n",
"\n",
" _, newlogprob, entropy, newvalue = agent.get_action_and_value(b_obs[mb_inds], b_actions.long()[mb_inds])\n",
" logratio = newlogprob - b_logprobs[mb_inds]\n",
" ratio = logratio.exp()\n",
"\n",
" with torch.no_grad():\n",
" # calculate approx_kl http://joschu.net/blog/kl-approx.html\n",
" old_approx_kl = (-logratio).mean()\n",
" approx_kl = ((ratio - 1) - logratio).mean()\n",
" clipfracs += [((ratio - 1.0).abs() > args.clip_coef).float().mean().item()]\n",
"\n",
" mb_advantages = b_advantages[mb_inds]\n",
" if args.norm_adv:\n",
" mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)\n",
"\n",
" # Policy loss\n",
" pg_loss1 = -mb_advantages * ratio\n",
" pg_loss2 = -mb_advantages * torch.clamp(ratio, 1 - args.clip_coef, 1 + args.clip_coef)\n",
" pg_loss = torch.max(pg_loss1, pg_loss2).mean()\n",
"\n",
" # Value loss\n",
" newvalue = newvalue.view(-1)\n",
" if args.clip_vloss:\n",
" v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2\n",
" v_clipped = b_values[mb_inds] + torch.clamp(\n",
" newvalue - b_values[mb_inds],\n",
" -args.clip_coef,\n",
" args.clip_coef,\n",
" )\n",
" v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2\n",
" v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped)\n",
" v_loss = 0.5 * v_loss_max.mean()\n",
" else:\n",
" v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean()\n",
"\n",
" entropy_loss = entropy.mean()\n",
" loss = pg_loss - args.ent_coef * entropy_loss + v_loss * args.vf_coef\n",
"\n",
" optimizer.zero_grad()\n",
" loss.backward()\n",
" nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm)\n",
" optimizer.step()\n",
"\n",
" if args.target_kl is not None:\n",
" if approx_kl > args.target_kl:\n",
" break\n",
"\n",
" y_pred, y_true = b_values.cpu().numpy(), b_returns.cpu().numpy()\n",
" var_y = np.var(y_true)\n",
" explained_var = np.nan if var_y == 0 else 1 - np.var(y_true - y_pred) / var_y\n",
"\n",
" # TRY NOT TO MODIFY: record rewards for plotting purposes\n",
" writer.add_scalar(\"charts/learning_rate\", optimizer.param_groups[0][\"lr\"], global_step)\n",
" writer.add_scalar(\"losses/value_loss\", v_loss.item(), global_step)\n",
" writer.add_scalar(\"losses/policy_loss\", pg_loss.item(), global_step)\n",
" writer.add_scalar(\"losses/entropy\", entropy_loss.item(), global_step)\n",
" writer.add_scalar(\"losses/old_approx_kl\", old_approx_kl.item(), global_step)\n",
" writer.add_scalar(\"losses/approx_kl\", approx_kl.item(), global_step)\n",
" writer.add_scalar(\"losses/clipfrac\", np.mean(clipfracs), global_step)\n",
" writer.add_scalar(\"losses/explained_variance\", explained_var, global_step)\n",
" print(\"SPS:\", int(global_step / (time.time() - start_time)))\n",
" writer.add_scalar(\"charts/SPS\", int(global_step / (time.time() - start_time)), global_step)\n",
"\n",
" envs.close()\n",
" writer.close()\n",
"\n",
" # Create the evaluation environment\n",
" eval_env = gym.make(args.env_id)\n",
"\n",
" package_to_hub(repo_id = args.repo_id,\n",
" model = agent, # The model we want to save\n",
" hyperparameters = args,\n",
" eval_env = gym.make(args.env_id),\n",
" logs= f\"runs/{run_name}\",\n",
" )\n",
" "
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "lteS7h5znWO3"
},
"source": [
"To be able to share your model with the community there are two more steps to follow:\n",
"\n",
"1️⃣ (If it's not already done) create an account to HF ➡ https://huggingface.co/join\n",
"\n",
"2️⃣ Sign in and then, you need to store your authentication token from the Hugging Face website.\n",
"- Create a new token (https://huggingface.co/settings/tokens) **with write role**"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "8sz1AtyenSCS"
},
"source": [
""
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Fms38KXMncKF"
},
"outputs": [],
"source": [
"from huggingface_hub import notebook_login\n",
"notebook_login()"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "jRqkGvk7pFQ6"
},
"source": [
"### Step 4: Let's start the training 🔥\n",
"- Now that you've coded from scratch PPO and added the Hugging Face Integration, we're ready to start the training 🔥"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "0tmEArP8ug2l"
},
"source": [
"- First, you need to copy all your code to a file you create called `ppo.py`"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "tjRIs94ku_K8"
},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "AvWTiKi9vPbJ"
},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "VrS80GmMu_j5"
},
"source": [
"- Now we just need to run this python script using `python .py` with the additional parameters we defined with `argparse`"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Sa0W5uv0vd-5"
},
"outputs": [],
"source": [
"!python ppo.py --env-id=\"LunarLander-v2\" --repo-id=\"YOUR_REPO_ID\" --total-timesteps=50000"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "eVsVJ5AdqLE7"
},
"source": [
"## Some additional challenges 🏆\n",
"In the [Leaderboard](https://huggingface.co/spaces/chrisjay/Deep-Reinforcement-Learning-Leaderboard) you will find your agents. Can you get to the top?\n",
"\n",
"Here are some ideas to achieve so:\n",
"* Train more steps\n",
"* Try different hyperparameters by looking at what your classmates have done.\n",
"* **Push your new trained model** on the Hub 🔥\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "nYdl758GqLXT"
},
"source": [
"See you on Next Time! 🔥\n",
"## Keep learning, stay awesome 🤗"
]
}
],
"metadata": {
"colab": {
"authorship_tag": "ABX9TyN2BU0ytdzum3AwXr4nlJMv",
"collapsed_sections": [],
"include_colab_link": true,
"name": "unit8.ipynb",
"private_outputs": true,
"provenance": []
},
"gpuClass": "standard",
"kernelspec": {
"display_name": "Python 3",
"name": "python3"
},
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 0
}