def __init__(self,
hp=None,
input_size=None,
output_size=None,
policy=None,
normalizer=None,
monitor_dir=None):
self.hp = hp or HyperParams() #HyperParameters used
np.random.seed(self.hp.seed) # Random seed initialize
self.env = pybullet_envs.make(self.hp.env_name) #Making Environment
# Setting Monitor to observe how agent plays
if monitor_dir is not None:
should_record = lambda i: self.record_video
self.env = wrappers.Monitor(self.env,
monitor_dir,
video_callable=should_record,
force=True)
self.in_size = input_size or self.env.observation_space.shape[
0] # Input size
self.out_size = output_size or self.env.action_space.shape[
0] # Output size
self.policy = policy or Policy(self.in_size, self.out_size,
self.hp) # Policy to be used
self.normalizer = normalizer or Normalizer(
self.in_size) # Normalizer to use
self.record_video = False
def train_eval_reacher_dqn(num_train: int,
num_test: int,
device: str,
config: Optional[Dict] = None) -> Tuple[Agent, float, float]:
assert device in ('cpu', 'cuda'), "The device must be either cpu or cuda"
env = pe.make('ReacherBulletEnv-v0')
print('Training the agent...')
agent = te.train_dqn_agent(env, num_train, config, device, show=True)
print()
print('\nTraining finished. Evaluating performance.')
mean_score, success_rate = te.evaluate_model(agent, num_test, show=True)
return agent, mean_score, success_rate
import time
import sys
sys.path.insert(0, '../..')
import libs_agents
from libs_common.Training import *
import models.ddpg_noisy.model.src.model_critic as ModelCritic
import models.ddpg_noisy.model.src.model_actor as ModelActor
import models.ddpg_noisy.model.src.config as Config
path = "models/ddpg_noisy/model/"
env = pybullet_envs.make("AntBulletEnv-v0")
#env.render()
agent = libs_agents.AgentDDPG(env, ModelCritic, ModelActor, Config)
max_iterations = 4*(10**6)
trainig = TrainingIterations(env, agent, max_iterations, path, 10000)
trainig.run()
'''
agent.load(path)
agent.disable_training()
while True:
reward, done = agent.main()
env.render()
time.sleep(0.01)
'discard_time_limit': True,
'alpha': 0.1,
'actor_update_interval': 2,
'critic_target_update_interval': 2,
'warmup_step': 1000,
'q_regularisation_k': 2,
'max_env_step': 200000,
'testing_gap': 10000,
'testing_episodes': 10,
'saving_gap': 100000,
}
seeds = [11]
seed_returns = []
path = os.path.dirname(os.path.realpath(__file__))
for seed in seeds:
env = pybullet_envs.make("Walker2DBulletEnv-v0")
# call render() before training to visualize (pybullet-gym-specific)
# env.render()
seed_path = path + '/seed'+str(seed)
agent = SACDrQ(algo_params=drq_params, env=env, path=seed_path, seed=seed)
agent.run(test=False)
# the sleep argument pause the rendering for a while at every env step, useful for slowing down visualization
# agent.run(test=True, load_network_ep=50, sleep=0.05)
seed_returns.append(agent.statistic_dict['episode_return'])
del env, agent
import time
import sys
sys.path.insert(0, '../..')
import libs_agents
from libs_common.Training import *
import models.ddpg_noisy.model.src.model_critic as ModelCritic
import models.ddpg_noisy.model.src.model_actor as ModelActor
import models.ddpg_noisy.model.src.config as Config
path = "models/ddpg_noisy/model/"
env = pybullet_envs.make("HalfCheetahBulletEnv-v0")
#env.render()
agent = libs_agents.AgentDDPG(env, ModelCritic, ModelActor, Config)
max_iterations = 4 * (10**6)
trainig = TrainingIterations(env, agent, max_iterations, path, 10000)
trainig.run()
'''
agent.load(path)
agent.disable_training()
while True:
reward, done = agent.main()
env.render()
time.sleep(0.01)
'''
'Q_weight_decay': 0.0,
'update_interval': 1,
'batch_size': 100,
'optimization_steps': 1,
'tau': 0.005,
'discount_factor': 0.99,
'discard_time_limit': True,
'warmup_step': 2500,
'observation_normalization': False,
'training_episodes': 101,
'testing_gap': 10,
'testing_episodes': 10,
'saving_gap': 50,
}
seeds = [11, 22, 33, 44, 55, 66]
seed_returns = []
path = os.path.dirname(os.path.realpath(__file__))
for seed in seeds:
env = pybullet_envs.make("InvertedPendulumSwingupBulletEnv-v0")
# call render() before training to visualize (pybullet-gym-specific)
# env.render()
seed_path = path + '/seed' + str(seed)
agent = DDPG(algo_params=ddpg_params, env=env, path=seed_path, seed=seed)
agent.run(test=False)
# the sleep argument pause the rendering for a while at every env step, useful for slowing down visualization
# agent.run(test=True, load_network_ep=50, sleep=0.05)
seed_returns.append(agent.statistic_dict['episode_return'])
del env, agent