Issues with applying PPO Impala on Retro Env in regards to running multiple environment

[hlsafin]

So what I essentially need is to so have something like
"venv = ProcgenEnv(num_envs=" ... but for retro.make(). Running multiple retro environments is causing issues for me, and retrowrapper isn't helping. Thank you!

[vwxyzjn]

Hey would you mind posting more specifics here? It’s kind of hard to help you with just these descriptions. You should specify, for example, what issue or error it was having, what does your env initialization code looks like.

[hlsafin]

Yes, I understand. I really don't have a specific example; I just wanted to know some broad adjustments needed to be made to account for Retro Environment; Since they have issues with running multiple environments once. I apologize if this was a vague post.

[vwxyzjn]

No worries, I think the script below would work, using some Atari wrappers. If you are interested in running more retro experiments, feel free to join our discord channel at https://discord.gg/D6RCjA6sVT. I would be happy to feature your retro experiments in our Open RL Benchmark

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions.categorical import Categorical
from torch.utils.tensorboard import SummaryWriter

import argparse
from distutils.util import strtobool
import numpy as np
import gym
from gym.wrappers import TimeLimit, Monitor
import retro
from gym.spaces import Discrete, Box, MultiBinary, MultiDiscrete, Space
import time
import random
import os
from stable_baselines3.common.atari_wrappers import (
    NoopResetEnv, MaxAndSkipEnv, EpisodicLifeEnv, FireResetEnv, WarpFrame, ClipRewardEnv)
from stable_baselines3.common.vec_env import DummyVecEnv, SubprocVecEnv, VecEnvWrapper
from stable_baselines3.common.vec_env import VecFrameStack

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='PPO agent')
    # Common arguments
    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="Airstriker-Genesis",
                        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=10000,
                        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 not be enabled by default')
    parser.add_argument('--prod-mode', type=lambda x:bool(strtobool(x)), default=False, nargs='?', const=True,
                        help='run the script in production mode and use wandb to log outputs')
    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)')
    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")

    # Algorithm specific arguments
    parser.add_argument('--n-minibatch', type=int, default=4,
                        help='the number of mini batch')
    parser.add_argument('--num-envs', type=int, default=8,
                        help='the number of parallel game environment')
    parser.add_argument('--num-steps', type=int, default=128,
                        help='the number of steps per game environment')
    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('--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('--clip-coef', type=float, default=0.1,
                        help="the surrogate clipping coefficient")
    parser.add_argument('--update-epochs', type=int, default=4,
                         help="the K epochs to update the policy")
    parser.add_argument('--kle-stop', type=lambda x:bool(strtobool(x)), default=False, nargs='?', const=True,
                         help='If toggled, the policy updates will be early stopped w.r.t target-kl')
    parser.add_argument('--kle-rollback', type=lambda x:bool(strtobool(x)), default=False, nargs='?', const=True,
                         help='If toggled, the policy updates will roll back to previous policy if KL exceeds target-kl')
    parser.add_argument('--target-kl', type=float, default=0.03,
                         help='the target-kl variable that is referred by --kl')
    parser.add_argument('--gae', type=lambda x:bool(strtobool(x)), default=True, nargs='?', const=True,
                         help='Use GAE for advantage computation')
    parser.add_argument('--norm-adv', type=lambda x:bool(strtobool(x)), default=True, nargs='?', const=True,
                          help="Toggles advantages normalization")
    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('--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.')

    args = parser.parse_args()
    if not args.seed:
        args.seed = int(time.time())

args.batch_size = int(args.num_envs * args.num_steps)
args.minibatch_size = int(args.batch_size // args.n_minibatch)

class VecPyTorch(VecEnvWrapper):
    def __init__(self, venv, device):
        super(VecPyTorch, self).__init__(venv)
        self.device = device

    def reset(self):
        obs = self.venv.reset()
        obs = torch.from_numpy(obs).float().to(self.device)
        return obs

    def step_async(self, actions):
        actions = actions.cpu().numpy()
        self.venv.step_async(actions)

    def step_wait(self):
        obs, reward, done, info = self.venv.step_wait()
        obs = torch.from_numpy(obs).float().to(self.device)
        reward = torch.from_numpy(reward).unsqueeze(dim=1).float()
        return obs, reward, done, info

# TRY NOT TO MODIFY: setup the environment
experiment_name = f"{args.gym_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
writer = SummaryWriter(f"runs/{experiment_name}")
writer.add_text('hyperparameters', "|param|value|\n|-|-|\n%s" % (
        '\n'.join([f"|{key}|{value}|" for key, value in vars(args).items()])))
if args.prod_mode:
    import wandb
    wandb.init(project=args.wandb_project_name, entity=args.wandb_entity, sync_tensorboard=True, config=vars(args), name=experiment_name, monitor_gym=True, save_code=True)
    writer = SummaryWriter(f"/tmp/{experiment_name}")

# TRY NOT TO MODIFY: seeding
device = torch.device('cuda' if torch.cuda.is_available() and args.cuda else 'cpu')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = args.torch_deterministic
def make_env(gym_id, seed, idx):
    def thunk():
        env = retro.make(gym_id, use_restricted_actions=retro.Actions.DISCRETE)
        env = MaxAndSkipEnv(env, skip=4)
        env = gym.wrappers.RecordEpisodeStatistics(env)
        if args.capture_video:
            if idx == 0:
                env = Monitor(env, f'videos/{experiment_name}')
        env = WarpFrame(env, width=84, height=84)
        env = ClipRewardEnv(env)
        env.seed(seed)
        env.action_space.seed(seed)
        env.observation_space.seed(seed)
        return env
    return thunk
envs = VecPyTorch(VecFrameStack(
    SubprocVecEnv([make_env(args.gym_id, args.seed+i, i) for i in range(args.num_envs)], start_method=("fork")),
4), device)
assert isinstance(envs.action_space, Discrete), "only discrete action space is supported"

# ALGO LOGIC: initialize agent here:
class Scale(nn.Module):
    def __init__(self, scale):
        super().__init__()
        self.scale = scale

    def forward(self, x):
        return x * self.scale

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, frames=4):
        super(Agent, self).__init__()
        self.network = nn.Sequential(
            Scale(1/255),
            layer_init(nn.Conv2d(frames, 32, 8, stride=4)),
            nn.ReLU(),
            layer_init(nn.Conv2d(32, 64, 4, stride=2)),
            nn.ReLU(),
            layer_init(nn.Conv2d(64, 64, 3, stride=1)),
            nn.ReLU(),
            nn.Flatten(),
            layer_init(nn.Linear(3136, 512)),
            nn.ReLU()
        )
        self.actor = layer_init(nn.Linear(512, envs.action_space.n), std=0.01)
        self.critic = layer_init(nn.Linear(512, 1), std=1)

    def forward(self, x):
        return self.network(x.permute((0, 3, 1, 2))) # "bhwc" -> "bchw"

    def get_action(self, x, action=None):
        logits = self.actor(self.forward(x))
        probs = Categorical(logits=logits)
        if action is None:
            action = probs.sample()
        return action, probs.log_prob(action), probs.entropy()

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

agent = Agent(envs).to(device)
optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)
if args.anneal_lr:
    # https://github.com/openai/baselines/blob/ea25b9e8b234e6ee1bca43083f8f3cf974143998/baselines/ppo2/defaults.py#L20
    lr = lambda f: f * args.learning_rate

# ALGO Logic: Storage for epoch data
obs = torch.zeros((args.num_steps, args.num_envs) + envs.observation_space.shape).to(device)
actions = torch.zeros((args.num_steps, args.num_envs) + envs.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
start_time = time.time()
# Note how `next_obs` and `next_done` are used; their usage is equivalent to
# https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/84a7582477fb0d5c82ad6d850fe476829dddd2e1/a2c_ppo_acktr/storage.py#L60
next_obs = envs.reset()
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 = lr(frac)
        optimizer.param_groups[0]['lr'] = lrnow

    # TRY NOT TO MODIFY: prepare the execution of the game.
    for step in range(0, args.num_steps):
        global_step += 1 * args.num_envs
        obs[step] = next_obs
        dones[step] = next_done

        # ALGO LOGIC: put action logic here
        with torch.no_grad():
            values[step] = agent.get_value(obs[step]).flatten()
            action, logproba, _ = agent.get_action(obs[step])

            # visualization
            if args.capture_video:
                probs_list = np.array(Categorical(
                    logits=agent.actor(agent.forward(obs[step]))).probs[0:1].tolist())
                envs.env_method("set_probs", probs_list, indices=0)

        actions[step] = action
        logprobs[step] = logproba

        # TRY NOT TO MODIFY: execute the game and log data.
        next_obs, rs, ds, infos = envs.step(action)
        rewards[step], next_done = rs.view(-1), torch.Tensor(ds).to(device)

        for info in infos:
            if 'episode' in info.keys():
                print(f"global_step={global_step}, episode_reward={info['episode']['r']}")
                writer.add_scalar("charts/episode_reward", info['episode']['r'], global_step)
                break

    # bootstrap reward if not done. reached the batch limit
    with torch.no_grad():
        last_value = agent.get_value(next_obs.to(device)).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 = last_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 = last_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.observation_space.shape)
    b_logprobs = logprobs.reshape(-1)
    b_actions = actions.reshape((-1,)+envs.action_space.shape)
    b_advantages = advantages.reshape(-1)
    b_returns = returns.reshape(-1)
    b_values = values.reshape(-1)

    # Optimizaing the policy and value network
    target_agent = Agent(envs).to(device)
    inds = np.arange(args.batch_size,)
    for i_epoch_pi in range(args.update_epochs):
        np.random.shuffle(inds)
        target_agent.load_state_dict(agent.state_dict())
        for start in range(0, args.batch_size, args.minibatch_size):
            end = start + args.minibatch_size
            minibatch_ind = inds[start:end]
            mb_advantages = b_advantages[minibatch_ind]
            if args.norm_adv:
                mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)

            _, newlogproba, entropy = agent.get_action(b_obs[minibatch_ind], b_actions.long()[minibatch_ind])
            ratio = (newlogproba - b_logprobs[minibatch_ind]).exp()

            # Stats
            approx_kl = (b_logprobs[minibatch_ind] - newlogproba).mean()

            # 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()
            entropy_loss = entropy.mean()

            # Value loss
            new_values = agent.get_value(b_obs[minibatch_ind]).view(-1)
            if args.clip_vloss:
                v_loss_unclipped = ((new_values - b_returns[minibatch_ind]) ** 2)
                v_clipped = b_values[minibatch_ind] + torch.clamp(new_values - b_values[minibatch_ind], -args.clip_coef, args.clip_coef)
                v_loss_clipped = (v_clipped - b_returns[minibatch_ind])**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 * ((new_values - b_returns[minibatch_ind]) ** 2).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.kle_stop:
            if approx_kl > args.target_kl:
                break
        if args.kle_rollback:
            if (b_logprobs[minibatch_ind] - agent.get_action(b_obs[minibatch_ind], b_actions.long()[minibatch_ind])[1]).mean() > args.target_kl:
                agent.load_state_dict(target_agent.state_dict())
                break

    # 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.mean().item(), global_step)
    writer.add_scalar("losses/approx_kl", approx_kl.item(), global_step)
    print("SPS:", int(global_step / (time.time() - start_time)))
    if args.kle_stop or args.kle_rollback:
        writer.add_scalar("debug/pg_stop_iter", i_epoch_pi, global_step)

envs.close()
writer.close()

[hlsafin]

Above is the progress I've made on Mortal-Kombat3-genesis environment. And the top-performing agent is PPG (via your implementation), I am just trying to make more adjustments and compare them IMPALA_Cnn.

[vwxyzjn]

Awesome! The PPG implementation is a little unstable. I wasn't able to reproduce the authors' results so I'd be careful with that.

[hlsafin]

Yeah, it seems like it's not getting past 400 rewards no matter what I try, unfortunately.

[vwxyzjn]

How long did you train it for?

[hlsafin]

Well, the longest one was trained for 12 hours straight, and on average it stayed around 400. (PPG). Currently, i'm trying PPO Impala(8 workers), and it's getting the same results(trained on 10 million global_steps so far)

[hlsafin]

mk3 seems to be a hard environment I suppose

[vwxyzjn]

Could you share the run?

[hlsafin]

code or graph?

[hlsafin]

[vwxyzjn]

Graph. Or the url of your experiment

[hlsafin]

blue is ppo-impala

[vwxyzjn]

Looks like there are some learning happened in the beginning. So this might be just environment specific

[hlsafin]

Do you think bringing down the learning rate would make any difference? Or changing any other hyper-parameter would make any changes? Or is this a lost cause?

[vwxyzjn]

Short answer is I am not sure, but I think they should help

[hlsafin]

Well, I'll keep trying and if anything happens I'll post it here. I'll let PPO-impala go on until it reaches around 40~ ish million steps and sees if it improves. If anyone has any suggestions, please feel free to reply.

[hlsafin]

https://wandb.ai/hlsafin/mk3_project/reports/Shared-panel-21-05-12-22-05-85--Vmlldzo2ODU5NjI?accessToken=rthx4pxd7w8w8yn79rlb534ppa5h8729u36m59hit8zm9ooqkvf23j526y5qsvcm

[vwxyzjn]

Closing now.