def __init__(self, *args, **kwargs):
'''
inputs:
a_dim: action spaces
base_dir: the directory that store data, like model, logs, and other data
'''
super().__init__()
self.no_save = bool(kwargs.get('no_save', False))
self.base_dir = base_dir = kwargs.get('base_dir')
tf_dtype = str(kwargs.get('tf_dtype'))
self._tf_data_type = tf.float32 if tf_dtype == 'float32' else tf.float64
tf.keras.backend.set_floatx(tf_dtype)
self.device = get_device()
self.cp_dir, self.log_dir, self.excel_dir = [os.path.join(base_dir, i) for i in ['model', 'log', 'excel']]
if not self.no_save:
check_or_create(self.cp_dir, 'checkpoints(models)')
if 1 == 0: # Not used
import pandas as pd
check_or_create(self.excel_dir, 'excel')
self.excel_writer = pd.ExcelWriter(self.excel_dir + '/data.xlsx')
self.global_step = tf.Variable(0, name="global_step", trainable=False, dtype=tf.int64) # in TF 2.x must be tf.int64, because function set_step need args to be tf.int64.
self._worker_params_dict = {}
self._all_params_dict = dict(global_step=self.global_step)
self.writer = self._create_writer(self.log_dir) # TODO: Annotation
if bool(kwargs.get('logger2file', False)):
set_log_file(log_file=os.path.join(self.log_dir, 'log.txt'))
def __init__(self, log_dir, ids=list(), *args, **kwargs):
self._ids = ids
if len(ids) == 0:
self._is_multi_logger = False
check_or_create(log_dir, 'logs(summaries)')
self._writer = SummaryWriter(log_dir)
else:
self._is_multi_logger = True
self._writer = {}
for id in ids:
_log_dir = log_dir + f'_{id}'
check_or_create(_log_dir, 'logs(summaries)')
self._writer['id'] = SummaryWriter(_log_dir)
def __init__(self, *args, **kwargs):
'''
inputs:
a_dim: action spaces
is_continuous: action type, refer to whether this control problem is continuous(True) or discrete(False)
base_dir: the directory that store data, like model, logs, and other data
'''
super().__init__()
base_dir = kwargs.get('base_dir')
tf_dtype = str(kwargs.get('tf_dtype'))
self._tf_data_type = tf.float32 if tf_dtype == 'float32' else tf.float64
tf.keras.backend.set_floatx(tf_dtype)
tf.random.set_seed(int(kwargs.get('seed', 0)))
self.device = get_device()
self.cp_dir, self.log_dir, self.excel_dir = [
os.path.join(base_dir, i) for i in ['model', 'log', 'excel']
]
self.logger = create_logger(name='rls.algos.base',
logger2file=bool(
kwargs.get('logger2file', False)),
file_name=self.log_dir + 'log.txt')
check_or_create(self.cp_dir, 'checkpoints')
check_or_create(self.log_dir, 'logs(summaries)')
if 1 == 0: # Not used
import pandas as pd
check_or_create(self.excel_dir, 'excel')
self.excel_writer = pd.ExcelWriter(self.excel_dir + '/data.xlsx')
self.global_step = tf.Variable(
0, name="global_step", trainable=False, dtype=tf.int64
) # in TF 2.x must be tf.int64, because function set_step need args to be tf.int64.
def __init__(
self,
n_copies=1,
is_save=True,
base_dir='',
device: str = 'cpu',
max_train_step=sys.maxsize,
max_frame_step=sys.maxsize,
max_train_episode=sys.maxsize,
save_frequency=100,
save2single_file=False,
n_step_value=4,
gamma=0.999,
logger_types=['none'],
decay_lr=False,
normalize_vector_obs=False,
obs_with_pre_action=False,
oplr_params=dict(),
rep_net_params={
'vector_net_params': {
'h_dim': 16,
'network_type': 'adaptive' # rls.nn.represents.vectors
},
'visual_net_params': {
'h_dim': 128,
'network_type': 'simple' # rls.nn.represents.visuals
},
'encoder_net_params': {
'h_dim': 16,
'network_type': 'identity' # rls.nn.represents.encoders
},
'memory_net_params': {
'rnn_units': 16,
'network_type': 'lstm'
}
},
**kwargs):
"""
inputs:
a_dim: action spaces
base_dir: the directory that store data, like model, logs, and other data
"""
self.n_copies = n_copies
self._is_save = is_save
self._base_dir = base_dir
self._training_name = os.path.split(self._base_dir)[-1]
self.device = device
logger.info(colorize(f"PyTorch Tensor Device: {self.device}"))
self._max_train_step = max_train_step
self._should_learn_cond_train_step = Until(max_train_step)
self._should_learn_cond_frame_step = Until(max_frame_step)
self._should_learn_cond_train_episode = Until(max_train_episode)
self._should_save_model = Every(save_frequency)
self._save2single_file = save2single_file
self.gamma = gamma
self._logger_types = logger_types
self._n_step_value = n_step_value
self._decay_lr = decay_lr # TODO: implement
self._normalize_vector_obs = normalize_vector_obs # TODO: implement
self._obs_with_pre_action = obs_with_pre_action
self._rep_net_params = dict(rep_net_params)
self._oplr_params = dict(oplr_params)
super().__init__()
self.memory_net_params = rep_net_params.get('memory_net_params', {
'rnn_units': 16,
'network_type': 'lstm'
})
self.use_rnn = self.memory_net_params.get('network_type',
'identity') != 'identity'
self.cp_dir, self.log_dir = [
os.path.join(base_dir, i) for i in ['model', 'log']
]
if self._is_save:
check_or_create(self.cp_dir, 'checkpoints(models)')
self._cur_interact_step = th.tensor(0).long().to(self.device)
self._cur_train_step = th.tensor(0).long().to(self.device)
self._cur_frame_step = th.tensor(0).long().to(self.device)
self._cur_episode = th.tensor(0).long().to(self.device)
self._trainer_modules = {
'_cur_train_step': self._cur_train_step,
'_cur_frame_step': self._cur_frame_step,
'_cur_episode': self._cur_episode
}
self._buffer = self._build_buffer()
self._loggers = self._build_loggers() if self._is_save else list()
def _create_writer(self, log_dir: str) -> tf.summary.SummaryWriter:
check_or_create(log_dir, 'logs(summaries)')
return tf.summary.create_file_writer(log_dir)