Note: 실행을 위해 아래의 패키지들을 설치해주기 바랍니다.

!pip install tqdm numpy scikit-learn pyglet setuptools && \
!pip install gym asciinema pandas tabulate tornado==5.* PyBullet && \
!pip install git+https://github.com/pybox2d/pybox2d#egg=Box2D && \
!pip install git+https://github.com/mimoralea/gym-bandits#egg=gym-bandits && \
!pip install git+https://github.com/mimoralea/gym-walk#egg=gym-walk && \
!pip install git+https://github.com/mimoralea/gym-aima#egg=gym-aima && \
!pip install gym[atari]
!pip install torch torchvision

import warnings ; warnings.filterwarnings('ignore')
import os
os.environ['CUDA_DEVICE_ORDER']='PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES']=''
os.environ['OMP_NUM_THREADS'] = '1'

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.multiprocessing as mp
import threading

import numpy as np
from IPython.display import display
from collections import namedtuple, deque
import matplotlib.pyplot as plt
import matplotlib.pylab as pylab
from itertools import cycle, count
from textwrap import wrap

import matplotlib
import subprocess
import os.path
import tempfile
import random
import base64
import pprint
import glob
import time
import json
import sys
import gym
import io
import os
import gc
import platform

from gym import wrappers
from subprocess import check_output
from IPython.display import HTML

LEAVE_PRINT_EVERY_N_SECS = 30
ERASE_LINE = '\x1b[2K'
EPS = 1e-6
RESULTS_DIR = os.path.join('.', 'gym-results')
SEEDS = (12, 34, 56, 78, 90)

%matplotlib inline

plt.style.use('fivethirtyeight')
params = {
    'figure.figsize': (15, 8),
    'font.size': 24,
    'legend.fontsize': 20,
    'axes.titlesize': 28,
    'axes.labelsize': 24,
    'xtick.labelsize': 20,
    'ytick.labelsize': 20
}
pylab.rcParams.update(params)
np.set_printoptions(suppress=True)

torch.cuda.is_available()

False

def get_make_env_fn(**kargs):
    def make_env_fn(env_name, seed=None, render=None, record=False,
                    unwrapped=False, monitor_mode=None, 
                    inner_wrappers=None, outer_wrappers=None):
        mdir = tempfile.mkdtemp()
        env = None
        if render:
            try:
                env = gym.make(env_name, render=render)
            except:
                pass
        if env is None:
            env = gym.make(env_name)
        if seed is not None: env.seed(seed)
        env = env.unwrapped if unwrapped else env
        if inner_wrappers:
            for wrapper in inner_wrappers:
                env = wrapper(env)
        env = wrappers.Monitor(
            env, mdir, force=True, 
            mode=monitor_mode, 
            video_callable=lambda e_idx: record) if monitor_mode else env
        if outer_wrappers:
            for wrapper in outer_wrappers:
                env = wrapper(env)
        return env
    return make_env_fn, kargs

def get_videos_html(env_videos, title, max_n_videos=5):
    videos = np.array(env_videos)
    if len(videos) == 0:
        return
    
    n_videos = max(1, min(max_n_videos, len(videos)))
    idxs = np.linspace(0, len(videos) - 1, n_videos).astype(int) if n_videos > 1 else [-1,]
    videos = videos[idxs,...]

    strm = '<h2>{}</h2>'.format(title)
    for video_path, meta_path in videos:
        video = io.open(video_path, 'r+b').read()
        encoded = base64.b64encode(video)

        with open(meta_path) as data_file:    
            meta = json.load(data_file)

        html_tag = """
        <h3>{0}</h3>
        <video width="960" height="540" controls>
            <source src="data:video/mp4;base64,{1}" type="video/mp4" />
        </video>"""
        strm += html_tag.format('Episode ' + str(meta['episode_id']), encoded.decode('ascii'))
    return strm

platform.system()

'Linux'

def get_gif_html(env_videos, title, subtitle_eps=None, max_n_videos=4):
    videos = np.array(env_videos)
    if len(videos) == 0:
        return
    
    n_videos = max(1, min(max_n_videos, len(videos)))
    idxs = np.linspace(0, len(videos) - 1, n_videos).astype(int) if n_videos > 1 else [-1,]
    videos = videos[idxs,...]

    strm = '<h2>{}</h2>'.format(title)
    for video_path, meta_path in videos:
        basename = os.path.splitext(video_path)[0]
        gif_path = basename + '.gif'
        if not os.path.exists(gif_path):
            if platform.system() == 'Linux':
                ps = subprocess.Popen(
                    ('ffmpeg', 
                     '-i', video_path, 
                     '-r', '7',
                     '-f', 'image2pipe', 
                     '-vcodec', 'ppm',
                     '-crf', '20',
                     '-vf', 'scale=512:-1',
                     '-'), 
                    stdout=subprocess.PIPE,
                    universal_newlines=True)
                output = subprocess.check_output(
                    ('convert',
                     '-coalesce',
                     '-delay', '7',
                     '-loop', '0',
                     '-fuzz', '2%',
                     '+dither',
                     '-deconstruct',
                     '-layers', 'Optimize',
                     '-', gif_path), 
                    stdin=ps.stdout)
                ps.wait()
            else:
                ps = subprocess.Popen('ffmpeg -i {} -r 7 -f image2pipe \
                                      -vcodec ppm -crf 20 -vf scale=512:-1 - | \
                                      convert -coalesce -delay 7 -loop 0 -fuzz 2% \
                                      +dither -deconstruct -layers Optimize \
                                      - {}'.format(video_path, gif_path), 
                                      stdin=subprocess.PIPE, 
                                      shell=True)
                ps.wait()

        gif = io.open(gif_path, 'r+b').read()
        encoded = base64.b64encode(gif)
            
        with open(meta_path) as data_file:    
            meta = json.load(data_file)

        html_tag = """
        <h3>{0}</h3>
        <img src="data:image/gif;base64,{1}" />"""
        prefix = 'Trial ' if subtitle_eps is None else 'Episode '
        sufix = str(meta['episode_id'] if subtitle_eps is None \
                    else subtitle_eps[meta['episode_id']])
        strm += html_tag.format(prefix + sufix, encoded.decode('ascii'))
    return strm

Different types of Cart Pole environments

class DiscountedCartPole(gym.Wrapper):
    def __init__(self, env):
        gym.Wrapper.__init__(self, env)
    def reset(self, **kwargs):
        return self.env.reset(**kwargs)
    def step(self, a):
        o, r, d, _ = self.env.step(a)
        (x, x_dot, theta, theta_dot) = o
        pole_fell =  x < -self.env.unwrapped.x_threshold \
                    or x > self.env.unwrapped.x_threshold \
                    or theta < -self.env.unwrapped.theta_threshold_radians \
                    or theta > self.env.unwrapped.theta_threshold_radians
        r = -1 if pole_fell else 0
        return o, r, d, _

class MCCartPole(gym.Wrapper):
    def __init__(self, env):
        gym.Wrapper.__init__(self, env)
    def reset(self, **kwargs):
        return self.env.reset(**kwargs)
    def step(self, a):
        o, r, d, _ = self.env.step(a)
        (x, x_dot, theta, theta_dot) = o
        pole_fell =  x < -self.env.unwrapped.x_threshold \
                    or x > self.env.unwrapped.x_threshold \
                    or theta < -self.env.unwrapped.theta_threshold_radians \
                    or theta > self.env.unwrapped.theta_threshold_radians
        if d:
            if pole_fell:
                r = 0 # done, in failure
            else:
                r = self.env._max_episode_steps # done, but successfully
        return o, r, d, _

Monte-Carlo REINFORCE

class FCDAP(nn.Module):
    def __init__(self, 
                 input_dim, 
                 output_dim,
                 hidden_dims=(32,32), 
                 activation_fc=F.relu):
        super(FCDAP, self).__init__()
        self.activation_fc = activation_fc

        self.input_layer = nn.Linear(input_dim, hidden_dims[0])
        self.hidden_layers = nn.ModuleList()
        for i in range(len(hidden_dims)-1):
            hidden_layer = nn.Linear(hidden_dims[i], hidden_dims[i+1])
            self.hidden_layers.append(hidden_layer)
        self.output_layer = nn.Linear(hidden_dims[-1], output_dim)

    def _format(self, state):
        x = state
        if not isinstance(x, torch.Tensor):
            x = torch.tensor(x, 
                             dtype=torch.float32)
            x = x.unsqueeze(0)
        return x
        
    def forward(self, state):
        x = self._format(state)
        x = self.activation_fc(self.input_layer(x))
        for hidden_layer in self.hidden_layers:
            x = self.activation_fc(hidden_layer(x))
        return self.output_layer(x)

    def full_pass(self, state):
        logits = self.forward(state)
        dist = torch.distributions.Categorical(logits=logits)
        action = dist.sample()
        logpa = dist.log_prob(action).unsqueeze(-1)
        entropy = dist.entropy().unsqueeze(-1)
        is_exploratory = action != np.argmax(logits.detach().numpy())
        return action.item(), is_exploratory.item(), logpa, entropy

    def select_action(self, state):
        logits = self.forward(state)
        dist = torch.distributions.Categorical(logits=logits)
        action = dist.sample()
        return action.item()
    
    def select_greedy_action(self, state):
        logits = self.forward(state)
        return np.argmax(logits.detach().numpy())

class REINFORCE():
    def __init__(self, policy_model_fn, policy_optimizer_fn, policy_optimizer_lr):
        self.policy_model_fn = policy_model_fn
        self.policy_optimizer_fn = policy_optimizer_fn
        self.policy_optimizer_lr = policy_optimizer_lr

    def optimize_model(self):
        T = len(self.rewards)
        discounts = np.logspace(0, T, num=T, base=self.gamma, endpoint=False)
        returns = np.array([np.sum(discounts[:T-t] * self.rewards[t:]) for t in range(T)])

        discounts = torch.FloatTensor(discounts).unsqueeze(1)
        returns = torch.FloatTensor(returns).unsqueeze(1)
        self.logpas = torch.cat(self.logpas)

        policy_loss = -(discounts * returns * self.logpas).mean()
        self.policy_optimizer.zero_grad()
        policy_loss.backward()
        self.policy_optimizer.step()

    def interaction_step(self, state, env):
        action, is_exploratory, logpa, _ = self.policy_model.full_pass(state)
        new_state, reward, is_terminal, _ = env.step(action)

        self.logpas.append(logpa)
        self.rewards.append(reward)
        
        self.episode_reward[-1] += reward
        self.episode_timestep[-1] += 1
        self.episode_exploration[-1] += int(is_exploratory)

        return new_state, is_terminal

    def train(self, make_env_fn, make_env_kargs, seed, gamma, 
              max_minutes, max_episodes, goal_mean_100_reward):
        training_start, last_debug_time = time.time(), float('-inf')

        self.checkpoint_dir = tempfile.mkdtemp()
        self.make_env_fn = make_env_fn
        self.make_env_kargs = make_env_kargs
        self.seed = seed
        self.gamma = gamma
        
        env = self.make_env_fn(**self.make_env_kargs, seed=self.seed)
        torch.manual_seed(self.seed) ; np.random.seed(self.seed) ; random.seed(self.seed)
    
        nS, nA = env.observation_space.shape[0], env.action_space.n
        self.episode_timestep = []
        self.episode_reward = []
        self.episode_seconds = []
        self.episode_exploration = []
        self.evaluation_scores = []
        
        self.policy_model = self.policy_model_fn(nS, nA)
        self.policy_optimizer = self.policy_optimizer_fn(self.policy_model, 
                                                         self.policy_optimizer_lr)
                    
        result = np.empty((max_episodes, 5))
        result[:] = np.nan
        training_time = 0
        for episode in range(1, max_episodes + 1):
            episode_start = time.time()
            
            state, is_terminal = env.reset(), False
            self.episode_reward.append(0.0)
            self.episode_timestep.append(0.0)
            self.episode_exploration.append(0.0)

            # collect rollout
            self.logpas, self.rewards = [], []
            for step in count():
                state, is_terminal = self.interaction_step(state, env)
                if is_terminal:
                    gc.collect()
                    break

            self.optimize_model()
            
            # stats
            episode_elapsed = time.time() - episode_start
            self.episode_seconds.append(episode_elapsed)
            training_time += episode_elapsed
            evaluation_score, _ = self.evaluate(self.policy_model, env)
            self.save_checkpoint(episode-1, self.policy_model)

            total_step = int(np.sum(self.episode_timestep))
            self.evaluation_scores.append(evaluation_score)
            
            mean_10_reward = np.mean(self.episode_reward[-10:])
            std_10_reward = np.std(self.episode_reward[-10:])
            mean_100_reward = np.mean(self.episode_reward[-100:])
            std_100_reward = np.std(self.episode_reward[-100:])
            mean_100_eval_score = np.mean(self.evaluation_scores[-100:])
            std_100_eval_score = np.std(self.evaluation_scores[-100:])
            lst_100_exp_rat = np.array(
                self.episode_exploration[-100:])/np.array(self.episode_timestep[-100:])
            mean_100_exp_rat = np.mean(lst_100_exp_rat)
            std_100_exp_rat = np.std(lst_100_exp_rat)

            wallclock_elapsed = time.time() - training_start
            result[episode-1] = total_step, mean_100_reward, \
                mean_100_eval_score, training_time, wallclock_elapsed
            
            reached_debug_time = time.time() - last_debug_time >= LEAVE_PRINT_EVERY_N_SECS
            reached_max_minutes = wallclock_elapsed >= max_minutes * 60
            reached_max_episodes = episode >= max_episodes
            reached_goal_mean_reward = mean_100_eval_score >= goal_mean_100_reward
            training_is_over = reached_max_minutes or \
                               reached_max_episodes or \
                               reached_goal_mean_reward

            elapsed_str = time.strftime("%H:%M:%S", time.gmtime(time.time() - training_start))
            debug_message = 'el {}, ep {:04}, ts {:06}, '
            debug_message += 'ar 10 {:05.1f}\u00B1{:05.1f}, '
            debug_message += '100 {:05.1f}\u00B1{:05.1f}, '
            debug_message += 'ex 100 {:02.1f}\u00B1{:02.1f}, '
            debug_message += 'ev {:05.1f}\u00B1{:05.1f}'
            debug_message = debug_message.format(
                elapsed_str, episode-1, total_step, mean_10_reward, std_10_reward, 
                mean_100_reward, std_100_reward, mean_100_exp_rat, std_100_exp_rat,
                mean_100_eval_score, std_100_eval_score)
            print(debug_message, end='\r', flush=True)
            if reached_debug_time or training_is_over:
                print(ERASE_LINE + debug_message, flush=True)
                last_debug_time = time.time()
            if training_is_over:
                if reached_max_minutes: print(u'--> reached_max_minutes \u2715')
                if reached_max_episodes: print(u'--> reached_max_episodes \u2715')
                if reached_goal_mean_reward: print(u'--> reached_goal_mean_reward \u2713')
                break
                
        final_eval_score, score_std = self.evaluate(self.policy_model, env, n_episodes=100)
        wallclock_time = time.time() - training_start
        print('Training complete.')
        print('Final evaluation score {:.2f}\u00B1{:.2f} in {:.2f}s training time,'
              ' {:.2f}s wall-clock time.\n'.format(
                  final_eval_score, score_std, training_time, wallclock_time))
        env.close() ; del env
        self.get_cleaned_checkpoints()
        return result, final_eval_score, training_time, wallclock_time
    
    def evaluate(self, eval_policy_model, eval_env, n_episodes=1, greedy=True):
        rs = []
        for _ in range(n_episodes):
            s, d = eval_env.reset(), False
            rs.append(0)
            for _ in count():
                if greedy:
                    a = eval_policy_model.select_greedy_action(s)
                else: 
                    a = eval_policy_model.select_action(s)
                s, r, d, _ = eval_env.step(a)
                rs[-1] += r
                if d: break
        return np.mean(rs), np.std(rs)

    def get_cleaned_checkpoints(self, n_checkpoints=5):
        try: 
            return self.checkpoint_paths
        except AttributeError:
            self.checkpoint_paths = {}

        paths = glob.glob(os.path.join(self.checkpoint_dir, '*.tar'))
        paths_dic = {int(path.split('.')[-2]):path for path in paths}
        last_ep = max(paths_dic.keys())
        # checkpoint_idxs = np.geomspace(1, last_ep+1, n_checkpoints, endpoint=True, dtype=np.int)-1
        checkpoint_idxs = np.linspace(1, last_ep+1, n_checkpoints, endpoint=True, dtype=np.int)-1

        for idx, path in paths_dic.items():
            if idx in checkpoint_idxs:
                self.checkpoint_paths[idx] = path
            else:
                os.unlink(path)

        return self.checkpoint_paths

    def demo_last(self, title='Fully-trained {} Agent', n_episodes=3, max_n_videos=3):
        env = self.make_env_fn(**self.make_env_kargs, monitor_mode='evaluation', render=True, record=True)

        checkpoint_paths = self.get_cleaned_checkpoints()
        last_ep = max(checkpoint_paths.keys())
        self.policy_model.load_state_dict(torch.load(checkpoint_paths[last_ep]))

        self.evaluate(self.policy_model, env, n_episodes=n_episodes)
        env.close()
        data = get_gif_html(env_videos=env.videos, 
                            title=title.format(self.__class__.__name__),
                            max_n_videos=max_n_videos)
        del env
        return HTML(data=data)

    def demo_progression(self, title='{} Agent progression', max_n_videos=5):
        env = self.make_env_fn(**self.make_env_kargs, monitor_mode='evaluation', render=True, record=True)

        checkpoint_paths = self.get_cleaned_checkpoints()
        for i in sorted(checkpoint_paths.keys()):
            self.policy_model.load_state_dict(torch.load(checkpoint_paths[i]))
            self.evaluate(self.policy_model, env, n_episodes=1)

        env.close()
        data = get_gif_html(env_videos=env.videos, 
                            title=title.format(self.__class__.__name__),
                            subtitle_eps=sorted(checkpoint_paths.keys()),
                            max_n_videos=max_n_videos)
        del env
        return HTML(data=data)

    def save_checkpoint(self, episode_idx, model):
        torch.save(model.state_dict(), 
                   os.path.join(self.checkpoint_dir, 'model.{}.tar'.format(episode_idx)))

reinforce_results = []
best_agent, best_eval_score = None, float('-inf')
for seed in SEEDS:
    environment_settings = {
        'env_name': 'CartPole-v1',
        'gamma': 1.00,
        'max_minutes': 10,
        'max_episodes': 10000,
        'goal_mean_100_reward': 475
    }

    policy_model_fn = lambda nS, nA: FCDAP(nS, nA, hidden_dims=(128,64))
    policy_optimizer_fn = lambda net, lr: optim.Adam(net.parameters(), lr=lr)
    policy_optimizer_lr = 0.0005

    env_name, gamma, max_minutes, \
    max_episodes, goal_mean_100_reward = environment_settings.values()
    agent = REINFORCE(policy_model_fn, policy_optimizer_fn, policy_optimizer_lr)

    make_env_fn, make_env_kargs = get_make_env_fn(env_name=env_name)
    # make_env_fn, make_env_kargs = get_make_env_fn(env_name=env_name, unwrapped=True)
    # make_env_fn, make_env_kargs = get_make_env_fn(
    #     env_name=env_name, addon_wrappers=[MCCartPole,])
    result, final_eval_score, training_time, wallclock_time = agent.train(
        make_env_fn, make_env_kargs, seed, gamma, max_minutes, max_episodes, goal_mean_100_reward)
    reinforce_results.append(result)
    if final_eval_score > best_eval_score:
        best_eval_score = final_eval_score
        best_agent = agent
reinforce_results = np.array(reinforce_results)

el 00:00:00, ep 0000, ts 000020, ar 10 020.0±000.0, 100 020.0±000.0, ex 100 0.5±0.0, ev 012.0±000.0
el 00:00:30, ep 0403, ts 030072, ar 10 207.1±092.1, 100 164.4±086.7, ex 100 0.3±0.0, ev 296.0±132.1
el 00:01:00, ep 0604, ts 075023, ar 10 374.2±095.5, 100 237.9±112.4, ex 100 0.3±0.0, ev 306.8±145.2
el 00:01:21, ep 0703, ts 107979, ar 10 373.6±129.8, 100 334.2±127.6, ex 100 0.3±0.0, ev 475.5±056.7
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 500.00±0.00 in 64.53s training time, 86.24s wall-clock time.

el 00:00:00, ep 0000, ts 000019, ar 10 019.0±000.0, 100 019.0±000.0, ex 100 0.5±0.0, ev 010.0±000.0
el 00:00:30, ep 0485, ts 023911, ar 10 173.6±068.5, 100 110.1±075.4, ex 100 0.3±0.0, ev 239.3±128.7
el 00:01:00, ep 0694, ts 065145, ar 10 223.7±063.6, 100 197.9±093.0, ex 100 0.3±0.0, ev 301.2±134.1
el 00:01:30, ep 0856, ts 110977, ar 10 318.1±094.9, 100 319.8±123.0, ex 100 0.3±0.0, ev 447.4±103.9
el 00:02:00, ep 1058, ts 154511, ar 10 184.5±024.9, 100 301.9±124.2, ex 100 0.3±0.0, ev 385.2±138.5
el 00:02:29, ep 1196, ts 201289, ar 10 316.4±145.6, 100 389.7±115.9, ex 100 0.3±0.0, ev 475.4±058.3
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 500.00±0.00 in 119.75s training time, 154.79s wall-clock time.

el 00:00:00, ep 0000, ts 000016, ar 10 016.0±000.0, 100 016.0±000.0, ex 100 0.7±0.0, ev 014.0±000.0
el 00:00:30, ep 0460, ts 023972, ar 10 197.6±096.7, 100 113.7±081.2, ex 100 0.3±0.1, ev 264.3±146.3
el 00:01:00, ep 0716, ts 064889, ar 10 384.3±131.7, 100 258.2±189.5, ex 100 0.3±0.0, ev 321.4±196.5
el 00:01:10, ep 0754, ts 082591, ar 10 494.7±015.9, 100 412.8±123.1, ex 100 0.3±0.0, ev 476.9±076.8
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 500.00±0.00 in 56.44s training time, 75.48s wall-clock time.

el 00:00:00, ep 0000, ts 000041, ar 10 041.0±000.0, 100 041.0±000.0, ex 100 0.5±0.0, ev 011.0±000.0
el 00:00:30, ep 0397, ts 029474, ar 10 190.9±089.6, 100 144.0±073.6, ex 100 0.3±0.0, ev 289.7±125.7
el 00:01:00, ep 0601, ts 071386, ar 10 354.8±106.5, 100 264.6±129.0, ex 100 0.3±0.0, ev 426.7±091.1
el 00:01:30, ep 0805, ts 113800, ar 10 446.0±096.3, 100 263.2±146.0, ex 100 0.3±0.0, ev 320.8±162.1
el 00:01:43, ep 0853, ts 134368, ar 10 404.3±083.7, 100 398.4±108.8, ex 100 0.3±0.0, ev 475.0±063.1
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 435.82±62.11 in 82.74s training time, 107.79s wall-clock time.

el 00:00:00, ep 0000, ts 000029, ar 10 029.0±000.0, 100 029.0±000.0, ex 100 0.6±0.0, ev 019.0±000.0
el 00:00:30, ep 0459, ts 026790, ar 10 179.2±058.6, 100 119.9±065.1, ex 100 0.3±0.0, ev 205.7±124.6
el 00:01:00, ep 0639, ts 069009, ar 10 398.4±119.3, 100 263.9±153.6, ex 100 0.2±0.0, ev 319.6±169.6
el 00:01:30, ep 0783, ts 117147, ar 10 431.5±096.4, 100 357.9±128.2, ex 100 0.2±0.0, ev 391.5±127.9
el 00:01:37, ep 0812, ts 128183, ar 10 379.3±118.5, 100 409.9±095.6, ex 100 0.2±0.0, ev 475.1±053.5
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 500.00±0.00 in 78.07s training time, 102.37s wall-clock time.

best_agent.demo_progression()

best_agent.demo_last()

reinforce_max_t, reinforce_max_r, reinforce_max_s, \
    reinforce_max_sec, reinforce_max_rt = np.max(reinforce_results, axis=0).T
reinforce_min_t, reinforce_min_r, reinforce_min_s, \
    reinforce_min_sec, reinforce_min_rt = np.min(reinforce_results, axis=0).T
reinforce_mean_t, reinforce_mean_r, reinforce_mean_s, \
    reinforce_mean_sec, reinforce_mean_rt = np.mean(reinforce_results, axis=0).T
reinforce_x = np.arange(len(reinforce_mean_s))

# reinforce_max_t, reinforce_max_r, reinforce_max_s, \
#     reinforce_max_sec, reinforce_max_rt = np.nanmax(reinforce_results, axis=0).T
# reinforce_min_t, reinforce_min_r, reinforce_min_s, \
#     reinforce_min_sec, reinforce_min_rt = np.nanmin(reinforce_results, axis=0).T
# reinforce_mean_t, reinforce_mean_r, reinforce_mean_s, \
#     reinforce_mean_sec, reinforce_mean_rt = np.nanmean(reinforce_results, axis=0).T
# reinforce_x = np.arange(len(reinforce_mean_s))

# change convergence checks to episode only (not minutes, not mean reward 'float('inf')' can help)

fig, axs = plt.subplots(5, 1, figsize=(20,30), sharey=False, sharex=True)

# REINFORCE
axs[0].plot(reinforce_max_r, 'y', linewidth=1)
axs[0].plot(reinforce_min_r, 'y', linewidth=1)
axs[0].plot(reinforce_mean_r, 'y', label='REINFORCE', linewidth=2)
axs[0].fill_between(reinforce_x, reinforce_min_r, reinforce_max_r, facecolor='y', alpha=0.3)

axs[1].plot(reinforce_max_s, 'y', linewidth=1)
axs[1].plot(reinforce_min_s, 'y', linewidth=1)
axs[1].plot(reinforce_mean_s, 'y', label='REINFORCE', linewidth=2)
axs[1].fill_between(reinforce_x, reinforce_min_s, reinforce_max_s, facecolor='y', alpha=0.3)

axs[2].plot(reinforce_max_t, 'y', linewidth=1)
axs[2].plot(reinforce_min_t, 'y', linewidth=1)
axs[2].plot(reinforce_mean_t, 'y', label='REINFORCE', linewidth=2)
axs[2].fill_between(reinforce_x, reinforce_min_t, reinforce_max_t, facecolor='y', alpha=0.3)

axs[3].plot(reinforce_max_sec, 'y', linewidth=1)
axs[3].plot(reinforce_min_sec, 'y', linewidth=1)
axs[3].plot(reinforce_mean_sec, 'y', label='REINFORCE', linewidth=2)
axs[3].fill_between(reinforce_x, reinforce_min_sec, reinforce_max_sec, facecolor='y', alpha=0.3)

axs[4].plot(reinforce_max_rt, 'y', linewidth=1)
axs[4].plot(reinforce_min_rt, 'y', linewidth=1)
axs[4].plot(reinforce_mean_rt, 'y', label='REINFORCE', linewidth=2)
axs[4].fill_between(reinforce_x, reinforce_min_rt, reinforce_max_rt, facecolor='y', alpha=0.3)

# ALL
axs[0].set_title('Moving Avg Reward (Training)')
axs[1].set_title('Moving Avg Reward (Evaluation)')
axs[2].set_title('Total Steps')
axs[3].set_title('Training Time')
axs[4].set_title('Wall-clock Time')
plt.xlabel('Episodes')
axs[0].legend(loc='upper left')
plt.show()

reinforce_root_dir = os.path.join(RESULTS_DIR, 'reinforce')
not os.path.exists(reinforce_root_dir) and os.makedirs(reinforce_root_dir)

np.save(os.path.join(reinforce_root_dir, 'x'), reinforce_x)

np.save(os.path.join(reinforce_root_dir, 'max_r'), reinforce_max_r)
np.save(os.path.join(reinforce_root_dir, 'min_r'), reinforce_min_r)
np.save(os.path.join(reinforce_root_dir, 'mean_r'), reinforce_mean_r)

np.save(os.path.join(reinforce_root_dir, 'max_s'), reinforce_max_s)
np.save(os.path.join(reinforce_root_dir, 'min_s'), reinforce_min_s )
np.save(os.path.join(reinforce_root_dir, 'mean_s'), reinforce_mean_s)

np.save(os.path.join(reinforce_root_dir, 'max_t'), reinforce_max_t)
np.save(os.path.join(reinforce_root_dir, 'min_t'), reinforce_min_t)
np.save(os.path.join(reinforce_root_dir, 'mean_t'), reinforce_mean_t)

np.save(os.path.join(reinforce_root_dir, 'max_sec'), reinforce_max_sec)
np.save(os.path.join(reinforce_root_dir, 'min_sec'), reinforce_min_sec)
np.save(os.path.join(reinforce_root_dir, 'mean_sec'), reinforce_mean_sec)

np.save(os.path.join(reinforce_root_dir, 'max_rt'), reinforce_max_rt)
np.save(os.path.join(reinforce_root_dir, 'min_rt'), reinforce_min_rt)
np.save(os.path.join(reinforce_root_dir, 'mean_rt'), reinforce_mean_rt)

Monte-Carlo VPG

weight, probs, entropies = -0.001, [], []
for p in np.arange(0, 1.01, 0.01):
    probs.append(p)
    p = torch.FloatTensor([p, 1-p])
    d = torch.distributions.Categorical(probs=p)
    entropies.append(weight * d.entropy().item())
plt.plot(probs, entropies)
plt.xlabel('Probability of action A\np(B)=1-p(A)', labelpad=20)
plt.ylabel('Negative\nweighted\nentropy', labelpad=80, rotation=0)
plt.title('Entropy contribution to the loss function\n{}*entropy(π)'.format(weight), pad=30)
plt.show()

class FCV(nn.Module):
    def __init__(self, 
                 input_dim,
                 hidden_dims=(32,32), 
                 activation_fc=F.relu):
        super(FCV, self).__init__()
        self.activation_fc = activation_fc

        self.input_layer = nn.Linear(input_dim, hidden_dims[0])
        self.hidden_layers = nn.ModuleList()
        for i in range(len(hidden_dims)-1):
            hidden_layer = nn.Linear(hidden_dims[i], hidden_dims[i+1])
            self.hidden_layers.append(hidden_layer)
        self.output_layer = nn.Linear(hidden_dims[-1], 1)

    def _format(self, state):
        x = state
        if not isinstance(x, torch.Tensor):
            x = torch.tensor(x,
                             dtype=torch.float32)
            x = x.unsqueeze(0)
        return x

    def forward(self, state):
        x = self._format(state)
        x = self.activation_fc(self.input_layer(x))
        for hidden_layer in self.hidden_layers:
            x = self.activation_fc(hidden_layer(x))
        return self.output_layer(x)

class VPG():
    def __init__(self, 
                 policy_model_fn, 
                 policy_model_max_grad_norm, 
                 policy_optimizer_fn, 
                 policy_optimizer_lr,
                 value_model_fn, 
                 value_model_max_grad_norm, 
                 value_optimizer_fn, 
                 value_optimizer_lr, 
                 entropy_loss_weight):
        self.policy_model_fn = policy_model_fn
        self.policy_model_max_grad_norm = policy_model_max_grad_norm
        self.policy_optimizer_fn = policy_optimizer_fn
        self.policy_optimizer_lr = policy_optimizer_lr
        
        self.value_model_fn = value_model_fn
        self.value_model_max_grad_norm = value_model_max_grad_norm
        self.value_optimizer_fn = value_optimizer_fn
        self.value_optimizer_lr = value_optimizer_lr
        
        self.entropy_loss_weight = entropy_loss_weight

    def optimize_model(self):
        T = len(self.rewards)
        discounts = np.logspace(0, T, num=T, base=self.gamma, endpoint=False)
        returns = np.array([np.sum(discounts[:T-t] * self.rewards[t:]) for t in range(T)])
        discounts = torch.FloatTensor(discounts[:-1]).unsqueeze(1)
        returns = torch.FloatTensor(returns[:-1]).unsqueeze(1)

        self.logpas = torch.cat(self.logpas)
        self.entropies = torch.cat(self.entropies) 
        self.values = torch.cat(self.values)

        value_error = returns - self.values
        policy_loss = -(discounts * value_error.detach() * self.logpas).mean()
        entropy_loss = -self.entropies.mean()
        loss = policy_loss + self.entropy_loss_weight * entropy_loss
        self.policy_optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(self.policy_model.parameters(), 
                                       self.policy_model_max_grad_norm)
        self.policy_optimizer.step()

        value_loss = value_error.pow(2).mul(0.5).mean()
        self.value_optimizer.zero_grad()
        value_loss.backward()
        torch.nn.utils.clip_grad_norm_(self.value_model.parameters(), 
                                       self.value_model_max_grad_norm)
        self.value_optimizer.step()
        
    def interaction_step(self, state, env):
        action, is_exploratory, logpa, entropy = self.policy_model.full_pass(state)
        new_state, reward, is_terminal, info = env.step(action)
        is_truncated = 'TimeLimit.truncated' in info and info['TimeLimit.truncated']

        self.logpas.append(logpa)
        self.entropies.append(entropy)
        self.rewards.append(reward)
        self.values.append(self.value_model(state))

        self.episode_reward[-1] += reward
        self.episode_timestep[-1] += 1
        self.episode_exploration[-1] += int(is_exploratory)
        return new_state, is_terminal, is_truncated

    def train(self, make_env_fn, make_env_kargs, seed, gamma, 
              max_minutes, max_episodes, goal_mean_100_reward):
        training_start, last_debug_time = time.time(), float('-inf')

        self.checkpoint_dir = tempfile.mkdtemp()
        self.make_env_fn = make_env_fn
        self.make_env_kargs = make_env_kargs
        self.seed = seed
        self.gamma = gamma
        
        env = self.make_env_fn(**self.make_env_kargs, seed=self.seed)
        torch.manual_seed(self.seed) ; np.random.seed(self.seed) ; random.seed(self.seed)
    
        nS, nA = env.observation_space.shape[0], env.action_space.n
        self.episode_timestep = []
        self.episode_reward = []
        self.episode_seconds = []
        self.episode_exploration = []
        self.evaluation_scores = []

        self.policy_model = self.policy_model_fn(nS, nA)
        self.policy_optimizer = self.policy_optimizer_fn(self.policy_model, 
                                                         self.policy_optimizer_lr)
        
        self.value_model = self.value_model_fn(nS)
        self.value_optimizer = self.value_optimizer_fn(self.value_model, 
                                                       self.value_optimizer_lr)
        result = np.empty((max_episodes, 5))
        result[:] = np.nan
        training_time = 0
        
        for episode in range(1, max_episodes + 1):
            episode_start = time.time()
            
            state, is_terminal = env.reset(), False
            self.episode_reward.append(0.0)
            self.episode_timestep.append(0.0)
            self.episode_exploration.append(0.0)

            # collect rollout
            self.logpas, self.entropies, self.rewards, self.values = [], [], [], []
            for step in count():
                state, is_terminal, is_truncated = self.interaction_step(state, env)
                if is_terminal:
                    gc.collect()
                    break

            is_failure = is_terminal and not is_truncated
            next_value = 0 if is_failure else self.value_model(state).detach().item()
            self.rewards.append(next_value)
            self.optimize_model()

            # stats
            episode_elapsed = time.time() - episode_start
            self.episode_seconds.append(episode_elapsed)
            training_time += episode_elapsed
            evaluation_score, _ = self.evaluate(self.policy_model, env)
            self.save_checkpoint(episode-1, self.policy_model)

            total_step = int(np.sum(self.episode_timestep))
            self.evaluation_scores.append(evaluation_score)
            
            mean_10_reward = np.mean(self.episode_reward[-10:])
            std_10_reward = np.std(self.episode_reward[-10:])
            mean_100_reward = np.mean(self.episode_reward[-100:])
            std_100_reward = np.std(self.episode_reward[-100:])
            mean_100_eval_score = np.mean(self.evaluation_scores[-100:])
            std_100_eval_score = np.std(self.evaluation_scores[-100:])
            lst_100_exp_rat = np.array(
                self.episode_exploration[-100:])/np.array(self.episode_timestep[-100:])
            mean_100_exp_rat = np.mean(lst_100_exp_rat)
            std_100_exp_rat = np.std(lst_100_exp_rat)

            wallclock_elapsed = time.time() - training_start
            result[episode-1] = total_step, mean_100_reward, \
                mean_100_eval_score, training_time, wallclock_elapsed
            
            reached_debug_time = time.time() - last_debug_time >= LEAVE_PRINT_EVERY_N_SECS
            reached_max_minutes = wallclock_elapsed >= max_minutes * 60
            reached_max_episodes = episode >= max_episodes
            reached_goal_mean_reward = mean_100_eval_score >= goal_mean_100_reward
            training_is_over = reached_max_minutes or \
                               reached_max_episodes or \
                               reached_goal_mean_reward

            elapsed_str = time.strftime("%H:%M:%S", time.gmtime(time.time() - training_start))
            debug_message = 'el {}, ep {:04}, ts {:06}, '
            debug_message += 'ar 10 {:05.1f}\u00B1{:05.1f}, '
            debug_message += '100 {:05.1f}\u00B1{:05.1f}, '
            debug_message += 'ex 100 {:02.1f}\u00B1{:02.1f}, '
            debug_message += 'ev {:05.1f}\u00B1{:05.1f}'
            debug_message = debug_message.format(
                elapsed_str, episode-1, total_step, mean_10_reward, std_10_reward, 
                mean_100_reward, std_100_reward, mean_100_exp_rat, std_100_exp_rat,
                mean_100_eval_score, std_100_eval_score)
            print(debug_message, end='\r', flush=True)
            if reached_debug_time or training_is_over:
                print(ERASE_LINE + debug_message, flush=True)
                last_debug_time = time.time()
            if training_is_over:
                if reached_max_minutes: print(u'--> reached_max_minutes \u2715')
                if reached_max_episodes: print(u'--> reached_max_episodes \u2715')
                if reached_goal_mean_reward: print(u'--> reached_goal_mean_reward \u2713')
                break

        final_eval_score, score_std = self.evaluate(self.policy_model, env, n_episodes=100)
        wallclock_time = time.time() - training_start
        print('Training complete.')
        print('Final evaluation score {:.2f}\u00B1{:.2f} in {:.2f}s training time,'
              ' {:.2f}s wall-clock time.\n'.format(
                  final_eval_score, score_std, training_time, wallclock_time))
        env.close() ; del env
        self.get_cleaned_checkpoints()
        return result, final_eval_score, training_time, wallclock_time

    def evaluate(self, eval_policy_model, eval_env, n_episodes=1, greedy=True):
        rs = []
        for _ in range(n_episodes):
            s, d = eval_env.reset(), False
            rs.append(0)
            for _ in count():
                if greedy:
                    a = eval_policy_model.select_greedy_action(s)
                else: 
                    a = eval_policy_model.select_action(s)
                s, r, d, _ = eval_env.step(a)
                rs[-1] += r
                if d: break
        return np.mean(rs), np.std(rs)

    def get_cleaned_checkpoints(self, n_checkpoints=5):
        try: 
            return self.checkpoint_paths
        except AttributeError:
            self.checkpoint_paths = {}

        paths = glob.glob(os.path.join(self.checkpoint_dir, '*.tar'))
        paths_dic = {int(path.split('.')[-2]):path for path in paths}
        last_ep = max(paths_dic.keys())
        # checkpoint_idxs = np.geomspace(1, last_ep+1, n_checkpoints, endpoint=True, dtype=np.int)-1
        checkpoint_idxs = np.linspace(1, last_ep+1, n_checkpoints, endpoint=True, dtype=np.int)-1

        for idx, path in paths_dic.items():
            if idx in checkpoint_idxs:
                self.checkpoint_paths[idx] = path
            else:
                os.unlink(path)

        return self.checkpoint_paths

    def demo_last(self, title='Fully-trained {} Agent', n_episodes=3, max_n_videos=3):
        env = self.make_env_fn(**self.make_env_kargs, monitor_mode='evaluation', render=True, record=True)

        checkpoint_paths = self.get_cleaned_checkpoints()
        last_ep = max(checkpoint_paths.keys())
        self.policy_model.load_state_dict(torch.load(checkpoint_paths[last_ep]))

        self.evaluate(self.policy_model, env, n_episodes=n_episodes)
        env.close()
        data = get_gif_html(env_videos=env.videos, 
                            title=title.format(self.__class__.__name__),
                            max_n_videos=max_n_videos)
        del env
        return HTML(data=data)

    def demo_progression(self, title='{} Agent progression', max_n_videos=5):
        env = self.make_env_fn(**self.make_env_kargs, monitor_mode='evaluation', render=True, record=True)

        checkpoint_paths = self.get_cleaned_checkpoints()
        for i in sorted(checkpoint_paths.keys()):
            self.policy_model.load_state_dict(torch.load(checkpoint_paths[i]))
            self.evaluate(self.policy_model, env, n_episodes=1)

        env.close()
        data = get_gif_html(env_videos=env.videos, 
                            title=title.format(self.__class__.__name__),
                            subtitle_eps=sorted(checkpoint_paths.keys()),
                            max_n_videos=max_n_videos)
        del env
        return HTML(data=data)

    def save_checkpoint(self, episode_idx, model):
        torch.save(model.state_dict(), 
                   os.path.join(self.checkpoint_dir, 'model.{}.tar'.format(episode_idx)))

vpg_results = []
best_agent, best_eval_score = None, float('-inf')
for seed in SEEDS:
    environment_settings = {
        'env_name': 'CartPole-v1',
        'gamma': 1.00,
        'max_minutes': 10,
        'max_episodes': 10000,
        'goal_mean_100_reward': 475
    }

    policy_model_fn = lambda nS, nA: FCDAP(nS, nA, hidden_dims=(128,64))
    policy_model_max_grad_norm = 1
    policy_optimizer_fn = lambda net, lr: optim.Adam(net.parameters(), lr=lr)
    policy_optimizer_lr = 0.0005

    value_model_fn = lambda nS: FCV(nS, hidden_dims=(256,128))
    value_model_max_grad_norm = float('inf')
    value_optimizer_fn = lambda net, lr: optim.RMSprop(net.parameters(), lr=lr)
    value_optimizer_lr = 0.0007

    entropy_loss_weight = 0.001

    env_name, gamma, max_minutes, \
    max_episodes, goal_mean_100_reward = environment_settings.values()
    agent = VPG(policy_model_fn, 
                policy_model_max_grad_norm, 
                policy_optimizer_fn, 
                policy_optimizer_lr,
                value_model_fn, 
                value_model_max_grad_norm, 
                value_optimizer_fn, 
                value_optimizer_lr, 
                entropy_loss_weight)

    make_env_fn, make_env_kargs = get_make_env_fn(env_name=env_name)
    result, final_eval_score, training_time, wallclock_time = agent.train(
        make_env_fn, make_env_kargs, seed, gamma, max_minutes, max_episodes, goal_mean_100_reward)
    vpg_results.append(result)
    if final_eval_score > best_eval_score:
        best_eval_score = final_eval_score
        best_agent = agent
vpg_results = np.array(vpg_results)

el 00:00:00, ep 0000, ts 000015, ar 10 015.0±000.0, 100 015.0±000.0, ex 100 0.7±0.0, ev 022.0±000.0
el 00:00:30, ep 0272, ts 023994, ar 10 204.4±074.4, 100 152.8±083.0, ex 100 0.3±0.0, ev 384.5±108.0
el 00:01:00, ep 0392, ts 056416, ar 10 337.2±072.2, 100 281.3±126.7, ex 100 0.3±0.0, ev 426.8±109.9
el 00:01:30, ep 0475, ts 091685, ar 10 485.0±045.0, 100 412.3±115.2, ex 100 0.3±0.0, ev 469.8±070.5
el 00:01:36, ep 0490, ts 099102, ar 10 491.7±024.9, 100 433.8±106.0, ex 100 0.3±0.0, ev 475.6±064.9
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 470.15±48.91 in 80.51s training time, 100.92s wall-clock time.

el 00:00:00, ep 0000, ts 000016, ar 10 016.0±000.0, 100 016.0±000.0, ex 100 0.5±0.0, ev 010.0±000.0
el 00:00:30, ep 0271, ts 023015, ar 10 224.7±140.5, 100 157.4±102.9, ex 100 0.3±0.0, ev 348.5±127.1
el 00:01:00, ep 0378, ts 053590, ar 10 375.6±109.5, 100 289.5±108.4, ex 100 0.3±0.0, ev 470.0±066.1
el 00:01:00, ep 0381, ts 054412, ar 10 364.8±130.6, 100 291.2±109.8, ex 100 0.3±0.0, ev 475.7±059.7
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 491.57±29.58 in 49.83s training time, 65.84s wall-clock time.

el 00:00:00, ep 0000, ts 000012, ar 10 012.0±000.0, 100 012.0±000.0, ex 100 0.2±0.0, ev 009.0±000.0
el 00:00:30, ep 0255, ts 021235, ar 10 178.6±065.7, 100 145.2±081.3, ex 100 0.3±0.0, ev 392.6±116.4
el 00:01:00, ep 0367, ts 053476, ar 10 319.4±129.0, 100 298.6±132.4, ex 100 0.3±0.0, ev 456.4±076.7
el 00:01:10, ep 0398, ts 065780, ar 10 434.8±127.8, 100 352.9±128.5, ex 100 0.3±0.0, ev 475.2±052.5
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 497.89±9.70 in 56.72s training time, 75.84s wall-clock time.

el 00:00:00, ep 0000, ts 000015, ar 10 015.0±000.0, 100 015.0±000.0, ex 100 0.3±0.0, ev 011.0±000.0
el 00:00:30, ep 0261, ts 023823, ar 10 235.1±082.2, 100 169.4±095.7, ex 100 0.3±0.0, ev 381.2±123.0
el 00:00:54, ep 0344, ts 049251, ar 10 359.2±111.2, 100 285.9±113.9, ex 100 0.3±0.0, ev 476.1±053.5
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 499.79±2.09 in 44.73s training time, 59.04s wall-clock time.

el 00:00:00, ep 0000, ts 000019, ar 10 019.0±000.0, 100 019.0±000.0, ex 100 0.3±0.0, ev 009.0±000.0
el 00:00:30, ep 0252, ts 021122, ar 10 233.9±078.6, 100 151.6±082.6, ex 100 0.3±0.0, ev 393.2±115.6
el 00:01:00, ep 0357, ts 053059, ar 10 403.8±121.6, 100 305.5±119.6, ex 100 0.3±0.0, ev 470.1±062.5
el 00:01:03, ep 0366, ts 056699, ar 10 412.9±111.0, 100 321.4±123.6, ex 100 0.3±0.0, ev 476.0±056.4
--> reached_goal_mean_reward ✓
Training complete.
Final evaluation score 500.00±0.00 in 50.90s training time, 68.78s wall-clock time.

best_agent.demo_progression()

Policy-Gradient and Actor-Critic Methods

Different types of Cart Pole environments

Monte-Carlo REINFORCE

REINFORCE Agent progression

Episode 0

Episode 175

Episode 351

Episode 527

Episode 703

Fully-trained REINFORCE Agent

Trial 0

Trial 1

Trial 2

Monte-Carlo VPG

VPG Agent progression

Episode 0

Episode 91

Episode 183