def __init__(self):
self.model = DDPG()
self.model.load_weights()
self.memory = Memory()
self.queue = fifo
self.model_queue = model_queue
self.warmup = config.get(MODEL_WARMUP)
self.modelsavefreq = config.get(MODEL_SAVE_FREQ) * 1000
def __init__(self):
self.model = DDPG()
self.num_episodes = config.get(EVALUATOR_NUM_EPISODES)
self.max_step = config.get(EVALUATOR_MAX_STEP)
self.visualize = config.get(EVALUATOR_VISABLE)
self.save = config.get(EVALUATOR_DRAW)
self.save_path = config.get(EVALUATOR_DRAW_PATH)
self.queue = fifo
self.model_queue = model_queue
self.model.load_weights()
def __init__(self):
limit = config.get(REPLAY_BUFFER_SIZE)
self.actions = RingBuffer(limit)
self.rewards = RingBuffer(limit)
self.terminals = RingBuffer(limit)
self.current_observations = RingBuffer(limit)
self.next_observations = RingBuffer(limit)
import logging
from utils.configsupport import config
from constfile.constkey import *
__fileout = config.get(LOG_ISFILEOUT)
__filepath = config.get(LOG_FILEPATH)
__level = config.get(LOG_LEVEL)
__formated = "%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s"
__datefmt = '%a, %d %b %Y %H:%M:%S'
if not __fileout:
logging.basicConfig(level=__level, format=__formated, datefmt=__datefmt)
else:
logging.basicConfig(level=__level,
format=__formated,
datefmt=__datefmt,
filename=__filepath)
log = logging.getLogger("ddpg")
def __init__(self):
# random seed for torch
__seed = config.get(MODEL_SEED)
self.policy_loss = []
self.critic_loss = []
if __seed > 0:
self.seed(__seed)
self.nb_states = config.get(MODEL_STATE_COUNT)
self.nb_actions = config.get(MODEL_ACTION_COUNT)
# Create Actor and Critic Network
actor_net_cfg = {
'hidden1': config.get(MODEL_ACTOR_HIDDEN1),
'hidden2': config.get(MODEL_ACTOR_HIDDEN2),
'init_w': config.get(MODEL_INIT_WEIGHT)
}
critic_net_cfg = {
'hidden1': config.get(MODEL_CRITIC_HIDDEN1),
'hidden2': config.get(MODEL_CRITIC_HIDDEN2),
'init_w': config.get(MODEL_INIT_WEIGHT)
}
self.actor = Actor(self.nb_states, self.nb_actions, **actor_net_cfg)
self.actor_target = Actor(self.nb_states, self.nb_actions,
**actor_net_cfg)
self.actor_optim = Adam(
self.actor.parameters(),
lr=config.get(MODEL_ACTOR_LR),
weight_decay=config.get(MODEL_ACTOR_WEIGHT_DECAY))
self.critic = Critic(self.nb_states, self.nb_actions, **critic_net_cfg)
self.critic_target = Critic(self.nb_states, self.nb_actions,
**critic_net_cfg)
self.critic_optim = Adam(
self.critic.parameters(),
lr=config.get(MODEL_CRITIC_LR),
weight_decay=config.get(MODEL_CRITIC_WEIGHT_DECAY))
hard_update(self.actor_target, self.actor)
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = Memory()
self.random_process = OrnsteinUhlenbeckProcess(
size=self.nb_actions,
theta=config.get(RANDOM_THETA),
mu=config.get(RANDOM_MU),
sigma=config.get(RANDOM_SIGMA))
# Hyper-parameters
self.batch_size = config.get(MODEL_BATCH_SIZE)
self.tau = config.get(MODEL_TARGET_TAU)
self.discount = config.get(MODEL_DISCOUNT)
self.depsilon = 1.0 / config.get(MODEL_EPSILON)
self.model_path = config.get(MODEL_SAVE_PATH)
#
self.epsilon = 1.0
# init device
self.device_init()