class TensorboardCallback(Callback):
def __init__(self,
path,
args=None,
events_dir=None,
max_step=None,
save_period=10):
self.save_period = save_period
self.path = path
train_dir = os.path.join(path, 'training')
if not os.path.exists(train_dir): os.makedirs(train_dir)
self.train_logger = Logger(train_dir)
valid_dir = os.path.join(path, 'validation')
if not os.path.exists(valid_dir): os.makedirs(valid_dir)
self.valid_logger = Logger(valid_dir)
if args:
text = 'Parameters\n---------\n'
for (key, val) in args.items():
text += '- ' + key + ' = ' + str(val) + '\n'
self.train_logger.log_text('Description', text)
self.valid_logger.log_text('Description', text)
if events_dir and max_step:
events_files = [
F for F in scan_dir(events_dir, '')[1]
if os.path.basename(F).startswith('events')
]
for events_file in events_files:
parent_dir = os.path.dirname(events_file).split(os.sep)[-1]
if 'training' == parent_dir:
train_events_file = events_file
elif 'validation' == parent_dir:
valid_events_file = events_file
self.train_logger.copyFrom(train_events_file, max_step=max_step)
self.valid_logger.copyFrom(valid_events_file, max_step=max_step)
def on_epoch_begin(self, epoch, logs={}):
self.starttime = time()
def on_epoch_end(self, epoch, logs={}):
self.train_logger.log_scalar("Speed", time() - self.starttime, epoch)
self.train_logger.log_scalar("sparse_categorical_accuracy_%",
logs['sparse_categorical_accuracy'] * 100,
epoch)
self.train_logger.log_scalar("loss", logs['loss'], epoch)
self.valid_logger.log_scalar("Speed", time() - self.starttime, epoch)
self.valid_logger.log_scalar(
"sparse_categorical_accuracy_%",
logs['val_sparse_categorical_accuracy'] * 100, epoch)
self.valid_logger.log_scalar("loss", logs['val_loss'], epoch)
# Model save
if ((epoch + 1) % self.save_period) == 0:
self.model.save(
os.path.join(self.path, 'save_' + str(epoch) + '.h5'))
_, oldsaves = scan_dir(self.path, '.h5')
for save in oldsaves:
try:
if int(save.split('.')[-2].split('_')[-1]) < epoch:
os.remove(save)
except:
continue
sparse=args.sparse)
eval_env = create_environment(args.env,
n_env=args.n_proc,
seed=42,
size=args.size,
sparse=args.sparse)
is_mario = True if 'Mario' in args.env else False
norm_input = True
# Logger
TB_LOGGER = Logger(sett.LOGPATH)
print('Torch Device: %s' % sett.device)
# Store HYPER in the log
for key, value in args._get_kwargs():
TB_LOGGER.log_text(tag=str(key), value=[str(value)], step=0)
obs = env.reset()
# Setup Model
n_actions = env.action_space.n if env.action_space.shape == (
) else env.action_space.shape[0]
n_state = env.observation_space.n if env.observation_space.shape == (
) else env.observation_space.shape
conv = True if isinstance(n_state, tuple) else False
if args.use_baseline:
dqn = DQN(state_dim=n_state,
tau=args.tau,
action_dim=n_actions,
