def agent(obs_dict, config_dict):
global prev_direction
env = make('hungry_geese')
# agent = QAgent(rows=11, columns=11, num_actions=3)
agent = PPOAgent(rows=11, columns=11, num_actions=3)
model_name = ''
agent.load_model_weights('models/' + model_name + '.h5')
state = preprocess_state(obs_dict, prev_direction)
action = agent.select_action(state)
direction = get_direction(prev_direction, action)
prev_direction = direction
return env.specification.action.enum[direction]
def test_run():
"""
use this to test runner
"""
env = BillardsEnv()
task = AvoidanceTask(env)
actor_critic = Policy(env.get_obs_shape(), task.get_action_space()) # TODO: get action space not yet implemented in env
agent = PPOAgent(actor_critic, clip_param=0.1,
ppo_epoch=4, num_mini_batch=32,
value_loss_coef=0.5, entropy_coef=0.01,
lr=2e-4, eps=1e-5, max_grad_norm=40)
runner = PPORunner(env=env, task=task, device='cuda', summary_path='./summary/5', agent=agent, actor_critic=actor_critic)
num_batches = 2000
try:
for i in range(num_batches):
runner.run_batch()
if i % 100 == 0:
torch.save(actor_critic, './models/model_5_' + str(i) + '.pt' )
except:
torch.save(actor_critic, './models/model_5_crash.pt')
exit(0)
torch.save(actor_critic, './models/model_5_end.pt')
def main():
seaborn.set()
# create agents with LSTM policy network
agents = [PPOAgent(game.actions,
LSTMPolicy(game.state_shape()[0], game.actions),
lr=5e-5, discount=0.99, eps=0.1)
for _ in range(game.num_players)]
# load agents if resuming
for i, a in enumerate(agents):
path = find_latest(args.agents, 'agent_{}_*.pt'.format(i))
print(f'Resuming agent {i} from path "{path}"')
a.load(path)
# load generator
path = find_latest(args.generator, 'generator_[0-9]*.pt')
print(f'Resuming generator from path "{path}"')
generator = RaceTrackGenerator.from_file(path)
latent = generator.latent_size
# agents on own boards
own_boards = torch.zeros(args.num_boards, RaceConfig.max_segments, 2, device=device)
for i in range(0, RaceConfig.max_segments, 16):
own_boards[:, i: i + 16, 0] = 2 * ((i // 16) % 2) - 1
run_evaluation(agents, game, own_boards, name='Own')
# # agents on random boards
val = 1.
random_boards = torch.zeros(args.num_boards, RaceConfig.max_segments, 2, device=device)
random_boards[:, :, 0].uniform_(-val, val)
run_evaluation(agents, game, random_boards, name='Random')
# generated dummy boards
dummy_generator = RaceTrackGenerator(latent, lr=1e-5, asynchronous=True)
generated_boards = dummy_generator.generate(RaceConfig.max_segments, args.num_boards)
run_evaluation(agents, game, generated_boards, name='Dummy')
# generated boards
generated_boards = generator.generate(RaceConfig.max_segments, args.num_boards, t=10.)
run_evaluation(agents, game, generated_boards, name='Generated')
def ppo_train(model_name, load_model=False, actor_filename=None, critic_filename=None, optimizer_filename=None):
print("PPO -- Training")
env = make('hungry_geese')
trainer = env.train(['greedy', None, 'agents/boilergoose.py', 'agents/handy_rl.py'])
agent = PPOAgent(rows=11, columns=11, num_actions=3)
memory = Memory()
if load_model:
agent.load_model_weights(actor_filename, critic_filename)
agent.load_optimizer_weights(optimizer_filename)
episode = 0
start_episode = 0
end_episode = 50000
reward_threshold = None
threshold_reached = False
epochs = 4
batch_size = 128
current_frame = 0
training_rewards = []
evaluation_rewards = []
last_1000_ep_reward = []
for episode in range(start_episode + 1, end_episode + 1):
obs_dict = trainer.reset()
ep_reward, ep_steps, done = 0, 0, False
prev_direction = 0
while not done:
current_frame += 1
ep_steps += 1
state = preprocess_state(obs_dict, prev_direction)
action = agent.select_action(state, training=True)
direction = get_direction(prev_direction, action)
next_obs_dict, _, done, _ = trainer.step(env.specification.action.enum[direction])
reward = calculate_reward(obs_dict, next_obs_dict)
next_state = preprocess_state(next_obs_dict, direction)
memory.add(state, action, reward, next_state, float(done))
obs_dict = next_obs_dict
prev_direction = direction
ep_reward += reward
if current_frame % batch_size == 0:
for _ in range(epochs):
states, actions, rewards, next_states, dones = memory.get_all_samples()
agent.fit(states, actions, rewards, next_states, dones)
memory.clear()
agent.update_networks()
print("EPISODE " + str(episode) + " - REWARD: " + str(ep_reward) + " - STEPS: " + str(ep_steps))
if len(last_1000_ep_reward) == 1000:
last_1000_ep_reward = last_1000_ep_reward[1:]
last_1000_ep_reward.append(ep_reward)
if reward_threshold:
if len(last_1000_ep_reward) == 1000:
if np.mean(last_1000_ep_reward) >= reward_threshold:
print("You solved the task after" + str(episode) + "episodes")
agent.save_model_weights('models/ppo_actor_' + model_name + '_' + str(episode) + '.h5',
'models/ppo_critic_' + model_name + '_' + str(episode) + '.h5')
threshold_reached = True
break
if episode % 1000 == 0:
print('Episode ' + str(episode) + '/' + str(end_episode))
last_1000_ep_reward_mean = np.mean(last_1000_ep_reward).round(3)
training_rewards.append(last_1000_ep_reward_mean)
print('Average reward in last 1000 episodes: ' + str(last_1000_ep_reward_mean))
print()
if episode % 1000 == 0:
eval_reward = 0
for i in range(100):
obs_dict = trainer.reset()
done = False
prev_direction = 0
while not done:
state = preprocess_state(obs_dict, prev_direction)
action = agent.select_action(state)
direction = get_direction(prev_direction, action)
next_obs_dict, _, done, _ = trainer.step(env.specification.action.enum[direction])
reward = calculate_reward(obs_dict, next_obs_dict)
obs_dict = next_obs_dict
prev_direction = direction
eval_reward += reward
eval_reward /= 100
evaluation_rewards.append(eval_reward)
print("Evaluation reward: " + str(eval_reward))
print()
if episode % 5000 == 0:
agent.save_model_weights('models/ppo_actor_' + model_name + '_' + str(episode) + '.h5',
'models/ppo_critic_' + model_name + '_' + str(episode) + '.h5')
agent.save_optimizer_weights('models/ppo_' + model_name + '_' + str(episode) + '_optimizer.npy')
agent.save_model_weights('models/ppo_actor_' + model_name + '_' + str(end_episode) + '.h5',
'models/ppo_critic_' + model_name + '_' + str(end_episode) + '.h5')
agent.save_optimizer_weights('models/ppo_' + model_name + '_' + str(end_episode) + '_optimizer.npy')
if threshold_reached:
plt.plot([i for i in range(start_episode + 1000, episode, 1000)], training_rewards)
else:
plt.plot([i for i in range(start_episode + 1000, end_episode + 1, 1000)], training_rewards)
plt.title("Reward")
plt.show()
plt.plot([i for i in range(start_episode + 1000, end_episode + 1, 1000)], evaluation_rewards)
plt.title('Evaluation rewards')
plt.show()
def main():
run_path = args.resume_path if args.resume_path else find_next_run_dir(
'experiments')
print(f'Running experiment {run_path}')
episode = 0
finish_mean = 0.
# create agents with LSTM policy network
agents = [
PPOAgent(game.actions,
LSTMPolicy(game.state_shape()[0], game.actions),
lr=5e-5,
discount=0.99,
eps=0.1) for _ in range(game.num_players)
]
# load agents if resuming
for i, a in enumerate(agents):
path = find_latest(args.agents, f'agent_{i}_*.pt')
print(f'Resuming agent {i} from path "{path}"')
a.load(path)
# create discriminator
discriminator = RaceWinnerDiscriminator(game.num_players,
lr=1e-5,
betas=(0.5, 0.9))
# create generator
generator = RaceTrackGenerator(args.latent, lr=1e-5, betas=(0.3, 0.9))
if args.resume_path:
path = find_latest(args.resume_path, 'discriminator_[0-9]*.pt')
print(f'Resuming discriminator from path "{path}"')
discriminator.load(path)
path = find_latest(args.resume_path, 'generator_[0-9]*.pt')
print(f'Resuming generator from path "{path}"')
generator.load(path)
path = find_latest(args.resume_path, 'params_*.pt')
print(f'Resuming params from path "{path}"')
params = torch.load(path)
episode = params['episode']
finish_mean = params['finish_mean']
summary_writer = SummaryWriter(os.path.join(run_path, 'summary'),
purge_step=episode)
result = {}
while True:
if episode % 30 == 0:
print(f'-- episode {episode}')
# -- training discriminator
boards = generator.generate(RaceConfig.max_segments,
args.batch_size).detach()
boards = torch.cat((boards, predefined_tracks()), dim=0)
boards = torch.cat(
(boards, -boards),
dim=0) # mirror levels to train more robust discriminator
rboards = boards.repeat(args.trials, 1, 1)
states, any_valid = game.reset(rboards)
game.record(0)
# run agents to find who wins
with torch.no_grad():
while any_valid and not game.finished():
actions = torch.stack(
[a.act(s, training=False) for a, s in zip(agents, states)],
dim=0)
states, rewards = game.step(actions)
for a in agents:
a.reset()
cur_mean = game.finishes.float().mean().item()
finish_mean = 0.9 * finish_mean + 0.1 * cur_mean
result['game/finishes'] = cur_mean
# discriminator calculate loss and perform backward pass
winners = one_hot(game.winners() + 1, num_classes=game.num_players + 1)
winners = winners.view(args.trials, -1,
*winners.shape[1:]).float().mean(0)
dloss, dacc = discriminator.train(boards.detach(), winners)
result['discriminator/loss'] = dloss
result['discriminator/accuracy'] = dacc
# -- train generator
for _ in range(args.generator_train_steps):
generated = generator.generate(RaceConfig.max_segments,
args.generator_batch_size)
pred_winners = discriminator.forward(generated)
gloss, galoss = generator.train(pred_winners, args.generator_beta)
result['generator/loss'] = gloss
if galoss:
result['generator/aux_loss'] = galoss
# log data
for p in range(game.num_players):
result[f'game/win_rates/player_{p}'] = winners[:, p +
1].mean().item()
result['game/invalid'] = winners[:, 0].mean().item()
# save episode
if episode % 100 == 0:
game.record_episode(
os.path.join(run_path, 'videos', f'episode_{episode}'))
# save boards as images in tensorboard
for i, img in enumerate(game.tracks_images(top_n=args.batch_size)):
result[f'game/boards_{i}'] = np.transpose(img, axes=(2, 0, 1))
# save networks
if episode % 500 == 0:
discriminator.save(
os.path.join(run_path, f'discriminator_{episode}.pt'))
generator.save(os.path.join(run_path, f'generator_{episode}.pt'))
torch.save({
'episode': episode,
'finish_mean': finish_mean
}, os.path.join(run_path, f'params_{episode}.pt'))
# save data to tensorboard
for tag, data in result.items():
if isinstance(data, np.ndarray):
summary_writer.add_image(tag, data, global_step=episode)
else:
summary_writer.add_scalar(tag, data, global_step=episode)
# -----
if episode % 1000 == 0:
gc.collect()
result.clear()
episode += 1
from agents import CommandLineAgent, DeepQLearningAgent, PPOAgent, RandomAgent
from environments.tictactoe import TicTacToeGameState
from runners import run_for_n_games_and_print_stats, run_step
if __name__ == "__main__":
gs = TicTacToeGameState()
agent0 = PPOAgent(state_space_size=gs.get_vectorized_state().shape[0],
action_space_size=gs.get_action_space_size())
agent1 = RandomAgent()
for i in range(100):
run_for_n_games_and_print_stats([agent0, agent1], gs, 5000)
run_for_n_games_and_print_stats([agent0, agent1], gs, 100)
gs_clone = gs.clone()
while not gs_clone.is_game_over():
run_step([agent0, CommandLineAgent()], gs_clone)
print(gs_clone)
gs_clone = gs.clone()
while not gs_clone.is_game_over():
run_step([CommandLineAgent(), agent1], gs_clone)
print(gs_clone)
"MsPacman-v0",
"SpaceInvaders-v0",
"Seaquest-v0",
"LunarLanderV2",
"Reacher-v2",
"FrozenLake-v0"
]
env = gym.make('MountainCar-v0')
obs, rew, done, ep_ret, ep_len = env.reset(), 0, False, 0, 0
epochs = 10
local_steps_per_epoch = 1000
# tf.set_random_seed(22222)
agent = PPOAgent(env.observation_space, env.action_space)
buffer = Buffer(env.observation_space.shape, env.action_space.shape, size=local_steps_per_epoch)
rewards = [0]
for epoch in tqdm(range(epochs)):
# print("Epoch {} Reward {}".format(epoch, rewards[-1]))
for t in range(local_steps_per_epoch):
act, v_t, logp_pi = agent.get_action(obs)
buffer.store(obs, act, rew, v_t, logp_pi) # Last var is logpi (not used in vpg)
obs, rew, done, _ = env.step(act[0])
ep_ret += rew
ep_len += 1
if done or (t==local_steps_per_epoch-1):
'action_size': brain.vector_action_space_size,
'number_of_agents': len(env_info.agents)
},
'pytorch': {
'device':
torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
},
'hyperparameters': {
'discount_rate': 0.99,
'tau': 0.95,
'gradient_clip': 5,
'rollout_length': 2048,
'optimization_epochs': 10,
'ppo_clip': 0.2,
'log_interval': 2048,
'max_steps': 1e5,
'mini_batch_number': 32,
'entropy_coefficent': 0.01,
'episode_count': 250,
'hidden_size': 512,
'adam_learning_rate': 3e-4,
'adam_epsilon': 1e-5
}
}
policy = PPOPolicyNetwork(config)
optimizier = optim.Adam(policy.parameters(),
config['hyperparameters']['adam_learning_rate'],
eps=config['hyperparameters']['adam_epsilon'])
agent = PPOAgent(env, brain_name, policy, optimizier, config)
agent.train(1500)
reset_pcnt = True
using_pcnt = True
bn_flag = 2
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
actor_critic = ActorCriticConv(num_dwell, input_channel, input_size,
std_scale, bn_flag)
actor_critic.to(device)
agent = PPOAgent(actor_critic,
num_dwell,
horizon,
max_grad_norm,
coef_value_loss,
coef_entropy_loss,
clip_param,
training_batch,
ppo_epochs,
constraint,
data_folder="./pg_data")
if load_pretrain_model:
print("Loading trained model from ", model_save_dir,
'{}-all-{}.ckpt'.format(total_supposed_steps, model_name))
model_path = os.path.join(
model_save_dir, '{}-all-{}.ckpt'.format(total_supposed_steps,
model_name))
state_dict = torch.load(model_path)
agent.model.load_state_dict(state_dict, strict=False)
import random
from agents import DeepQLearningAgent, RandomAgent, TabQLearningAgent, DeepQLearningExperienceReplayAgent, PPOAgent
from environments.battle_royale import BattleRoyalGameWorldTerminal, BattleRoyale
from runners import run_for_n_games_and_print_stats, run_step
import tensorflow as tf
if __name__ == "__main__":
tf.compat.v1.disable_eager_execution()
state_space_size = 19
list_agent = [
PPOAgent(state_space_size=state_space_size,
action_space_size=48,
episodes_count_between_training=1)
if i < 1 else RandomAgent() for i in range(6)
]
for i in range(100):
#random.shuffle(list_agent)
gs = BattleRoyalGameWorldTerminal(i,
numberofPlayer=6,
list_agent=list_agent)
gs.run()
#list_agent[0].epsilon = -1
#list_agent[1].epsilon = -1
gs2 = BattleRoyale(numberofPlayer=6, list_agent=list_agent)
gs2.run()