action='store_true',
default=False)
args = parser.parse_args()
env, environment_params, environment_wrappers, environment_wrapper_arguments = choose_env(
ENV_NAME)
prefix = ENV_NAME + str(len(environment_params["parameters_to_randomise"])
) # number of randomised parameters
model_path = '../../../../../data/epi/model/' + prefix + '_epi'
if TASK_POLICY_ALG == 'td3':
# load td3 hyper parameters
[action_range, batch_size, explore_steps, update_itr, explore_noise_scale, eval_noise_scale, reward_scale, \
hidden_dim, noise_decay, policy_target_update_interval, q_lr, policy_lr, replay_buffer_size, DETERMINISTIC] = \
load_params('td3', ['action_range', 'batch_size', 'explore_steps', 'update_itr', 'explore_noise_scale',\
'eval_noise_scale', 'reward_scale', 'hidden_dim', 'noise_decay', \
'policy_target_update_interval', 'q_lr', 'policy_lr','replay_buffer_size', 'deterministic'] )
# load replay buffer when off-policy
BaseManager.register('ReplayBuffer', ReplayBuffer)
manager = BaseManager()
manager.start()
replay_buffer = manager.ReplayBuffer(
replay_buffer_size) # share the replay buffer through manager
elif TASK_POLICY_ALG == 'ppo':
[batch_size] = load_params('ppo', ['batch_size'])
epi = EPI(env, data_path='./data/' + ENV_NAME, model_path=model_path)
if args.epi:
epi.embedding_learn(env,
from sim2real_policies.utils.envs import make_env
from sim2real_policies.utils.evaluate import evaluate
from sim2real_policies.utils.optimizers import SharedAdam, ShareParameters
from sim2real_policies.utils.load_params import load_params
##### hyper-parameters for RL training ############
ENV_NAME = ['SawyerReach', 'SawyerPush', 'SawyerSlide'][0] # environment name
EP_MAX = 100000 # total number of episodes for training
EP_LEN = 200 # total number of steps for each episode
prefix = ''
MODEL_PATH = '../../../../data/ppo/model/' + prefix + 'ppo'
NUM_WORKERS = 1 # or: mp.cpu_count()
EVAL_INTERVAL = 100
[ACTION_RANGE, BATCH, GAMMA, RANDOMSEED, A_UPDATE_STEPS, C_UPDATE_STEPS, EPS,\
A_LR, C_LR, METHOD] = load_params('ppo', ['action_range', 'batch_size', 'gamma', 'random_seed', 'actor_update_steps', \
'critic_update_steps', 'eps', 'actor_lr', 'critic_lr', 'method' ])
############################### PPO ####################################
class PPO(object):
'''
PPO class
'''
def __init__(self, state_space, action_space, hidden_dim=512):
self.actor = PPO_PolicyNetwork(state_space,
action_space,
hidden_dim,
action_range=ACTION_RANGE)
self.critic = ValueNetwork(state_space, hidden_dim)
self.actor_optimizer = SharedAdam(self.actor.parameters(), lr=A_LR)