latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if latest_checkpoint:
print("Loading model checkpoint {}...\n".format(latest_checkpoint))
# saver.restore(sess, latest_checkpoint)
try:
total_t = sess.run(tf.train.get_global_step())
except:
total_t = 0
epsilon_start = 1.0
epsilon_end = 0.1
epsilon_decay_steps = 500000
epsilons = np.linspace(epsilon_start, epsilon_end, epsilon_decay_steps)
policy_angle = dqn_utils.make_epsilon_greedy_policy(
angle_estimator, len(valid_angles))
policy_taptime = dqn_utils.make_epsilon_greedy_policy(
taptime_estimator, len(valid_taptimes))
########################################
##### Populating replay memory (size: N)
# 원래는 랜덤하게 N번의 shot을 해서 replay_memory를 채워야 하지만...
# 각 레벨별로, 0도부터 90도까지 쏜 데이터를 replay_memory로 함.
# pre_train을 넣어서, 이 replay_memory로 학습을 한 weight를 가져와서 시작하는 것도 고려.
Transition = namedtuple(
"Transition",
["state", "action", "reward", "next_state", "game_state"])
replay_memory_size = 500000
print('Populating replay memory...')
replay_memory = []
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir) # path를 반환
if latest_checkpoint:
print("Loading model checkpoint {}...\n".format(latest_checkpoint))
saver.restore(sess, latest_checkpoint)
total_t = sess.run(tf.train.get_global_step())
# 처음에 안됐었던 이유는, global_step이란 tensor 변수를 안만들어서임
## user parameter
epsilon_start = 1.0
epsilon_end = 0.1
epsilon_decay_steps = 500000
epsilons = np.linspace(epsilon_start, epsilon_end, epsilon_decay_steps)
policy = dqn_utils.make_epsilon_greedy_policy(
estimator, [len(valid_angles), len(valid_taptimes)])
########################################
##### Populating replay memory (size: N)
# 원래는 랜덤하게 N번의 shot을 해서 replay_memory를 채워야 하지만...
# 각 레벨별로, 0도부터 90도까지 쏜 데이터를 replay_memory로 함.
batch_size = 6
discount_factor = 0.99
Transition = namedtuple(
"Transition",
["state", "action", "reward", "next_state", "game_state"])
replay_memory_size = 500000
print('Populating replay memory...')
# replay_memory = [] # pretrain_memory가 사실 replay_memory가 되는게 아닌가?
# with open(os.path.join(EXP_PATH, 'pretrain_memory_5'), 'rb') as f:
with open(os.path.join(EXP_PATH, 'replay_memoryAll'), 'rb') as f: