def main():
# create config
config = Config()
config.game = args.game
config.algo = 'ppo'
config.max_steps = int(2e6)
config.num_envs = 1
config.optimizer = 'RMSprop'
config.lr = 0.0003
config.discount = 0.99
config.use_gae = True
config.gae_lambda = 0.95
config.use_grad_clip = True
config.max_grad_norm = 0.5
config.rollout_length = 2048
config.value_loss_coef = 0.5
config.entropy_coef = 0
config.ppo_epoch = 10
config.ppo_clip_param = 0.2
config.num_mini_batch = 32
config.use_gpu = True
config.seed = args.seed
config.num_frame_stack = 1
config.after_set()
print(config)
# prepare env, model and logger
env = make_vec_envs(config.game,
num_envs=config.num_envs,
seed=config.seed,
num_frame_stack=config.num_frame_stack)
model = GaussianSeparatedPolicy(env.observation_space.shape[0],
action_dim=get_action_dim(
env.action_space)).to(config.device)
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes)
# create agent and run
agent = PPOAgent(config, env, model, logger)
agent.run()
def main():
# create config with basic parameters for a2c
config = Config()
config.game = 'BreakoutNoFrameskip-v4'
config.algo = 'a2c'
config.max_steps = int(2e7)
config.num_envs = 16
config.optimizer = 'RMSprop'
config.lr = 0.0001
config.discount = 0.99
config.use_gae = True
config.gae_lambda = 0.95
config.use_grad_clip = True
config.max_grad_norm = 5
config.rollout_length = 5
config.value_loss_coef = 0.5
config.entropy_coef = 0.01
config.use_gpu = True
config.num_frame_stack = 4
config.seed = 1
# update config with argparse object (pass game and seed from command line)
config.update(args)
config.tag = '%s-%s-%d' % (config.game, config.algo, config.seed)
config.after_set()
print(config)
# prepare env, model and logger
env = make_vec_envs(config.game,
num_envs=config.num_envs,
seed=config.seed,
num_frame_stack=config.num_frame_stack)
model = CatACConvNet(input_channels=env.observation_space.shape[0],
action_dim=get_action_dim(env.action_space)).to(
config.device)
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes)
# create agent and run
agent = A2CAgent(config, env, model, logger)
agent.run()
def main():
# create config
config = Config()
config.game = 'BreakoutNoFrameskip-v4'
config.algo = 'dqn'
config.max_steps = int(1e7)
config.num_envs = 4
config.optimizer = 'RMSprop'
config.lr = 0.00025
config.discount = 0.99
config.use_grad_clip = True
config.max_grad_norm = 5
config.replay_size = int(1e5)
config.replay_batch = 32
config.replay_on_gpu = False
config.exploration_threshold_start = 1
config.exploration_threshold_end = 0.01
config.exploration_steps = int(1e6)
config.target_update_interval = int(1e4)
config.learning_start = int(5e4)
config.intermediate_eval = True
config.eval_interval = int(1e5)
config.use_gpu = True
config.num_frame_stack = 4
config.seed = 0
config.log_episodes_avg_window = 10000
# update config with argparse object (pass game and seed from command line)
config.update(args)
config.tag = '%s-%s-%d' % (config.game, config.algo, config.seed)
config.after_set()
print(config)
# prepare env, model and logger
env = make_vec_envs(config.game, num_envs = config.num_envs, seed = config.seed, num_frame_stack= config.num_frame_stack)
eval_env = make_vec_envs(config.game, num_envs = 1, seed = config.seed, num_frame_stack= config.num_frame_stack)
model = CatQConvNet(input_channels = env.observation_space.shape[0], action_dim = get_action_dim(env.action_space)).to(config.device)
target_model = CatQConvNet(input_channels = env.observation_space.shape[0], action_dim = get_action_dim(env.action_space)).to(config.device)
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes, config.log_episodes_avg_window)
# create agent and run
agent = DQNAgent(config, env, eval_env, model, target_model, logger)
agent.run()
def main():
# create config with basic parameters for td3
config = Config()
config.game = 'HalfCheetah-v2'
config.algo = 'td3'
config.max_steps = int(1e6)
config.num_envs = 1
config.optimizer = 'Adam'
config.lr = 0.001
config.discount = 0.99
config.replay_size = int(1e6)
config.replay_batch = 100
config.replay_on_gpu = True
config.warmup_steps = 10000
config.action_noise = 0.1
config.target_noise = 0.2
config.target_noise_clip = 0.5
config.policy_delay = 2
config.soft_update_rate = 0.005
config.intermediate_eval = True
config.eval_interval = int(1e4)
config.use_gpu = True
config.seed = 0
# update config with argparse object (pass game and seed from command line)
config.update(args)
config.tag = '%s-%s-%d' % (config.game, config.algo, config.seed)
config.after_set()
print(config)
# prepare env, model and logger
env = make_vec_envs(config.game, num_envs = config.num_envs, seed = config.seed)
eval_env = make_vec_envs(config.game, num_envs = 1, seed = config.seed)
model = ConDetADCLinearNet(input_dim = env.observation_space.shape[0], action_dim = get_action_dim(env.action_space)).to(config.device)
target_model = ConDetADCLinearNet(input_dim = env.observation_space.shape[0], action_dim = get_action_dim(env.action_space)).to(config.device)
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes)
# create agent and run
agent = TD3Agent(config, env, eval_env, model, target_model, logger)
agent.run()
def main():
# create config
config = Config()
config.game = 'starpilot'
config.algo = 'ppo'
config.max_steps = int(1e8)
config.num_envs = 256
config.optimizer = 'Adam'
config.lr = 0.0005
config.discount = 0.999
config.use_gae = True
config.gae_lambda = 0.95
config.use_grad_clip = True
config.max_grad_norm = 0.5
config.rollout_length = 256
config.value_loss_coef = 0.5
config.entropy_coef = 0.01
config.ppo_epoch = 3
config.ppo_clip_param = 0.2
config.num_mini_batch = 8
config.mini_batch_size = 2048
config.use_gpu = True
config.num_frame_stack = 1
# update config with argparse object
config.update(args)
config.tag = '%s-%s' % (config.game, config.algo)
if args.tag:
config.tag += '-' + args.tag
config.after_set()
print(config)
# prepare env, model and logger
env = make_vec_envs_procgen(env_name=config.game,
num_envs=config.num_envs,
start_level=config.start_level,
num_levels=config.num_levels,
distribution_mode=config.distribution_mode,
num_frame_stack=config.num_frame_stack)
Model = SeparateImpalaCNN if args.separate_actor_critic else ImpalaCNN
model = Model(input_channels=env.observation_space.shape[0],
action_dim=get_action_dim(env.action_space)).to(
config.device)
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes)
# create agent and run
agent = PPOAgent(config, env, model, logger)
agent.run()
def main():
# prepare env
encoder = None
if args.use_encoder:
encoder = Encoder(latent_size=args.latent_size)
weights = torch.load(args.encoder_path,
map_location=torch.device('cpu'))
for k in list(weights.keys()):
if k not in encoder.state_dict().keys():
del weights[k]
encoder.load_state_dict(weights)
carla_env = gym.make('carla-v0', params=params)
env = VecGymCarla(carla_env, args.action_repeat, encoder)
# prepare config
config = Config()
config.game = 'weather%d' % args.weather
config.algo = 'ppo'
config.max_steps = args.max_steps
config.num_envs = 1
config.optimizer = 'Adam'
config.lr = args.lr
config.discount = 0.99
config.use_gae = True
config.gae_lambda = 0.95
config.use_grad_clip = True
config.max_grad_norm = 0.5
config.rollout_length = 128
config.value_loss_coef = 1
config.entropy_coef = 0.01
config.ppo_epoch = 4
config.ppo_clip_param = 0.2
config.num_mini_batch = 4
config.use_gpu = True
config.save_interval = 10000
config.memory_on_gpu = True
config.after_set()
print(config)
# prepare model
if args.use_encoder:
Model = CarlaLatentPolicy
input_dim = args.latent_size + 1 # 16+1 in paper
else:
Model = CarlaImgPolicy
input_dim = args.latent_size + 1 # 128+1 in paper (16 is too small)
model = Model(input_dim, 2).to(config.device)
# create ppo agent and run
logger = Logger(SummaryWriter(config.save_path), config.num_echo_episodes)
# create agent and run
agent = PPOAgent(config, env, model, logger)
agent.run()
torch.save(model.state_dict(), args.model_save_path)