Converting RL loop (PPO, DQN, etc) into a selfplay setting by changing 2 lines

slimevolley_ppo_selfplay.py

"""
Imagine we have a training loop for an agent. E.g. PPO, or DQN, or whatever.

What is the easiest way to convert this into a selfplay? 

To make this happen we want to run 2 identical loop: 1 loop for each agent.
But. There are 2 sync points: `env.reset()`, and each `env.step()`.
Is there a way to avoid duplicating code for setting/loading agent model, creating/updating
buffers, running training loops etc... Yes, it is. By utilizing generators protocol: `yield` and `send`.

The idea is simple: turn loop into a generator by `yield`-ing on `reset` and on each `step`, feed back
the result of interacting with the environment using `.send`. This makes it possible for us to have a single
env, and orchestrate 2 loops independently. Easier to show than to explain. See "__main__" with the
implementation.

Code for PPO is taken from CleanRL: https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/ppo.py
SlimeVolley Gym environment: https://github.com/hardmaru/slimevolleygym
"""

import argparse
import os
import random
import time
from tqdm import tqdm
from distutils.util import strtobool

import gym
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim

from torch.distributions.categorical import Categorical
from torch.distributions.bernoulli import Bernoulli

from torch.utils.tensorboard import SummaryWriter

import slimevolleygym

def parse_args():
    # fmt: off
    parser = argparse.ArgumentParser()
    parser.add_argument("--exp-name", type=str, default=os.path.basename(__file__).rstrip(".py"),
        help="the name of this experiment")
    parser.add_argument("--gym-id", type=str, default="CartPole-v1",
        help="the id of the gym environment")
    parser.add_argument("--learning-rate", type=float, default=2.5e-4,
        help="the learning rate of the optimizer")
    parser.add_argument("--seed", type=int, default=1,
        help="seed of the experiment")
    parser.add_argument("--total-timesteps", type=int, default=25000,
        help="total timesteps of the experiments")
    parser.add_argument("--torch-deterministic", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="if toggled, `torch.backends.cudnn.deterministic=False`")
    parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="if toggled, cuda will be enabled by default")
    parser.add_argument("--track", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="if toggled, this experiment will be tracked with Weights and Biases")
    parser.add_argument("--wandb-project-name", type=str, default="cleanRL",
        help="the wandb's project name")
    parser.add_argument("--wandb-entity", type=str, default=None,
        help="the entity (team) of wandb's project")
    parser.add_argument("--capture-video", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="weather to capture videos of the agent performances (check out `videos` folder)")

    # Algorithm specific arguments
    parser.add_argument("--num-envs", type=int, default=4,
        help="the number of parallel game environments")
    parser.add_argument("--num-steps", type=int, default=128,
        help="the number of steps to run in each environment per policy rollout")
    parser.add_argument("--anneal-lr", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Toggle learning rate annealing for policy and value networks")
    parser.add_argument("--gae", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Use GAE for advantage computation")
    parser.add_argument("--gamma", type=float, default=0.99,
        help="the discount factor gamma")
    parser.add_argument("--gae-lambda", type=float, default=0.95,
        help="the lambda for the general advantage estimation")
    parser.add_argument("--num-minibatches", type=int, default=4,
        help="the number of mini-batches")
    parser.add_argument("--update-epochs", type=int, default=4,
        help="the K epochs to update the policy")
    parser.add_argument("--norm-adv", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Toggles advantages normalization")
    parser.add_argument("--clip-coef", type=float, default=0.2,
        help="the surrogate clipping coefficient")
    parser.add_argument("--clip-vloss", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Toggles wheter or not to use a clipped loss for the value function, as per the paper.")
    parser.add_argument("--ent-coef", type=float, default=0.01,
        help="coefficient of the entropy")
    parser.add_argument("--vf-coef", type=float, default=0.5,
        help="coefficient of the value function")
    parser.add_argument("--max-grad-norm", type=float, default=0.5,
        help="the maximum norm for the gradient clipping")
    parser.add_argument("--target-kl", type=float, default=None,
        help="the target KL divergence threshold")
    args = parser.parse_args()
    args.batch_size = int(args.num_envs * args.num_steps)
    args.minibatch_size = int(args.batch_size // args.num_minibatches)
    # fmt: on
    return args


def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
    torch.nn.init.orthogonal_(layer.weight, std)
    torch.nn.init.constant_(layer.bias, bias_const)
    return layer


class Agent(nn.Module):
    def __init__(self, envs):
        super(Agent, self).__init__()
        self.critic = nn.Sequential(
            layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
            nn.Tanh(),
            layer_init(nn.Linear(64, 64)),
            nn.Tanh(),
            layer_init(nn.Linear(64, 1), std=1.0),
        )
        self.actor = nn.Sequential(
            layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
            nn.Tanh(),
            layer_init(nn.Linear(64, 64)),
            nn.Tanh(),
            layer_init(nn.Linear(64, envs.single_action_space.n), std=0.01),
        )

    def get_value(self, x):
        return self.critic(x)

    def get_action_and_value(self, x, action=None):
        logits = self.actor(x)
        # probs = Categorical(logits=logits)
        probs = Bernoulli(logits=logits)
        if action is None:
            action = probs.sample()
        return action, probs.log_prob(action.float()).sum(dim=1), probs.entropy().sum(dim=1), self.critic(x)


def train_loop(args, envs, run_name, agent_id):
    run_name = f"{run_name}__{agent_id}"

    if args.track:
        import wandb

        wandb.init(
            project=args.wandb_project_name,
            entity=args.wandb_entity,
            sync_tensorboard=True,
            config=vars(args),
            name=run_name,
            monitor_gym=True,
            save_code=True,
        )
    writer = SummaryWriter(f"runs/{run_name}")
    writer.add_text(
        "hyperparameters",
        "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
    )

    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")

    # EDITED HERE
    # assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"

    agent = Agent(envs).to(device)
    optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)

    # ALGO Logic: Storage setup
    obs = torch.zeros((args.num_steps, args.num_envs) + envs.single_observation_space.shape).to(device)
    actions = torch.zeros((args.num_steps, args.num_envs) + envs.single_action_space.shape).to(device)
    logprobs = torch.zeros((args.num_steps, args.num_envs)).to(device)
    rewards = torch.zeros((args.num_steps, args.num_envs)).to(device)
    dones = torch.zeros((args.num_steps, args.num_envs)).to(device)
    values = torch.zeros((args.num_steps, args.num_envs)).to(device)

    # TRY NOT TO MODIFY: start the game
    global_step = 0

    # EDITED HERE
    # xxx(okachaiev): ideally i need 2 pbars for running env and for epoches
    pbar = tqdm(total=args.total_timesteps)
    pbar.set_description("Episodic return: n/a")

    start_time = time.time()

    # EDITED HERE
    reset_obs = yield

    next_obs = torch.Tensor(reset_obs).to(device)
    next_done = torch.zeros(args.num_envs).to(device)
    num_updates = args.total_timesteps // args.batch_size

    for update in range(1, num_updates + 1):
        # Annealing the rate if instructed to do so.
        if args.anneal_lr:
            frac = 1.0 - (update - 1.0) / num_updates
            lrnow = frac * args.learning_rate
            optimizer.param_groups[0]["lr"] = lrnow

        for step in range(0, args.num_steps):
            global_step += 1 * args.num_envs
            obs[step] = next_obs
            dones[step] = next_done

            # ALGO LOGIC: action logic
            with torch.no_grad():
                action, logprob, _, value = agent.get_action_and_value(next_obs)
                values[step] = value.flatten()
            actions[step] = action
            logprobs[step] = logprob

            # TRY NOT TO MODIFY: execute the game and log data.
            # EDITED HERE
            next_obs, reward, done, info = yield action.cpu().numpy()
            rewards[step] = torch.tensor(reward).to(device).view(-1)
            next_obs, next_done = torch.Tensor(next_obs).to(device), torch.Tensor(done).to(device)

            for item in info:
                if "episode" in item.keys():
                    # EDITED HERE
                    pbar.set_description(f"Episodic return: {item['episode']['r']}")

                    writer.add_scalar("charts/episodic_return", item["episode"]["r"], global_step)
                    writer.add_scalar("charts/episodic_length", item["episode"]["l"], global_step)
                    break

            pbar.update(args.num_envs)

        # bootstrap value if not done
        with torch.no_grad():
            next_value = agent.get_value(next_obs).reshape(1, -1)
            if args.gae:
                advantages = torch.zeros_like(rewards).to(device)
                lastgaelam = 0
                for t in reversed(range(args.num_steps)):
                    if t == args.num_steps - 1:
                        nextnonterminal = 1.0 - next_done
                        nextvalues = next_value
                    else:
                        nextnonterminal = 1.0 - dones[t + 1]
                        nextvalues = values[t + 1]
                    delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t]
                    advantages[t] = lastgaelam = delta + args.gamma * args.gae_lambda * nextnonterminal * lastgaelam
                returns = advantages + values
            else:
                returns = torch.zeros_like(rewards).to(device)
                for t in reversed(range(args.num_steps)):
                    if t == args.num_steps - 1:
                        nextnonterminal = 1.0 - next_done
                        next_return = next_value
                    else:
                        nextnonterminal = 1.0 - dones[t + 1]
                        next_return = returns[t + 1]
                    returns[t] = rewards[t] + args.gamma * nextnonterminal * next_return
                advantages = returns - values

        # flatten the batch
        b_obs = obs.reshape((-1,) + envs.single_observation_space.shape)
        b_logprobs = logprobs.reshape(-1)
        b_actions = actions.reshape((-1,) + envs.single_action_space.shape)
        b_advantages = advantages.reshape(-1)
        b_returns = returns.reshape(-1)
        b_values = values.reshape(-1)

        # Optimizaing the policy and value network
        b_inds = np.arange(args.batch_size)
        clipfracs = []
        for epoch in range(args.update_epochs):
            np.random.shuffle(b_inds)
            for start in range(0, args.batch_size, args.minibatch_size):
                end = start + args.minibatch_size
                mb_inds = b_inds[start:end]

                _, newlogprob, entropy, newvalue = agent.get_action_and_value(b_obs[mb_inds], b_actions.long()[mb_inds])
                logratio = newlogprob - b_logprobs[mb_inds]
                ratio = logratio.exp()

                with torch.no_grad():
                    # calculate approx_kl http://joschu.net/blog/kl-approx.html
                    # old_approx_kl = (-logratio).mean()
                    approx_kl = ((ratio - 1) - logratio).mean()
                    clipfracs += [((ratio - 1.0).abs() > args.clip_coef).float().mean().item()]

                mb_advantages = b_advantages[mb_inds]
                if args.norm_adv:
                    mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)

                # Policy loss
                pg_loss1 = -mb_advantages * ratio
                pg_loss2 = -mb_advantages * torch.clamp(ratio, 1 - args.clip_coef, 1 + args.clip_coef)
                pg_loss = torch.max(pg_loss1, pg_loss2).mean()

                # Value loss
                newvalue = newvalue.view(-1)
                if args.clip_vloss:
                    v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2
                    v_clipped = b_values[mb_inds] + torch.clamp(
                        newvalue - b_values[mb_inds],
                        -args.clip_coef,
                        args.clip_coef,
                    )
                    v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2
                    v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped)
                    v_loss = 0.5 * v_loss_max.mean()
                else:
                    v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean()

                entropy_loss = entropy.mean()
                loss = pg_loss - args.ent_coef * entropy_loss + v_loss * args.vf_coef

                optimizer.zero_grad()
                loss.backward()
                nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm)
                optimizer.step()

            if args.target_kl is not None:
                if approx_kl > args.target_kl:
                    break

        y_pred, y_true = b_values.cpu().numpy(), b_returns.cpu().numpy()
        var_y = np.var(y_true)
        explained_var = np.nan if var_y == 0 else 1 - np.var(y_true - y_pred) / var_y

        # TRY NOT TO MODIFY: record rewards for plotting purposes
        writer.add_scalar("charts/learning_rate", optimizer.param_groups[0]["lr"], global_step)
        writer.add_scalar("losses/value_loss", v_loss.item(), global_step)
        writer.add_scalar("losses/policy_loss", pg_loss.item(), global_step)
        writer.add_scalar("losses/entropy", entropy_loss.item(), global_step)
        writer.add_scalar("losses/approx_kl", approx_kl.item(), global_step)
        writer.add_scalar("losses/clipfrac", np.mean(clipfracs), global_step)
        writer.add_scalar("losses/explained_variance", explained_var, global_step)
        # EDITED HERE
        # print("SPS:", int(global_step / (time.time() - start_time)))
        writer.add_scalar("charts/SPS", int(global_step / (time.time() - start_time)), global_step)

    torch.save(agent.state_dict(), f"trained_agents/{run_name}.pt")
    pbar.close()
    writer.close()


# EDITED HERE: make SlimeVolley a little bit more friendly to Wrappers
class SelfPlayWrapper(gym.Wrapper):

    def step(self, action):
        action1, action2 = action
        obs, reward, done, info = self.env.step(action1, action2)
        return (obs, info['otherObs']), reward, done, info


def make_env(gym_id, seed, run_name):
    env = gym.make(gym_id)

    env.single_observation_space = env.observation_space
    env.single_action_space = env.action_space

    # EDITED HERE: need to wrap, otherwise the rest of the wrapper
    # won't work (step only accepts a single argument)
    env = SelfPlayWrapper(env)

    env = gym.wrappers.RecordEpisodeStatistics(env)
    env.seed(seed)
    env.action_space.seed(seed)
    env.observation_space.seed(seed)
    return env


if __name__ == "__main__":
    args = parse_args()
    run_name = f"{args.gym_id}__{args.exp_name}__{args.seed}__{int(time.time())}"

    # TRY NOT TO MODIFY: seeding
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    torch.backends.cudnn.deterministic = args.torch_deterministic

    # env setup
    # EDITED HERE, SlimeVolley is hard with VecEnv :(
    envs = make_env(args.gym_id, args.seed, run_name)

    # RUNNING 1-vs-1 BASED ON NORMAL PPO LOOP
    # make loops (generators)
    loop_agent1 = train_loop(args, envs, run_name, "1")
    loop_agent2 = train_loop(args, envs, run_name, "2")

    # first yield
    next(loop_agent1), next(loop_agent2)
    next_obs = envs.reset()
    next_obs = np.expand_dims(next_obs, 0)
    action1 = loop_agent1.send(next_obs)
    action2 = loop_agent2.send(next_obs)

    # each following yield
    while True:
        (next_obs1, next_obs2), reward, done, info = envs.step((action1.squeeze(0), action2.squeeze(0)))
        try:
            action1 = loop_agent1.send((np.expand_dims(next_obs1,0), np.expand_dims(reward,0), np.array([done]), [info]))
            action2 = loop_agent2.send((np.expand_dims(next_obs2,0), np.expand_dims(-reward,0), np.array([done]), [info]))
        except StopIteration:
            break

    envs.close()