def LogSetup(self):
Logger().init(
config_or_config_dict=GlobalConfig().DEFAULT_LOG_CONFIG_DICT,
log_path=GlobalConfig().DEFAULT_LOG_PATH,
log_level=GlobalConfig().DEFAULT_LOG_LEVEL)
ConsoleLogger().init(logger_name='console_logger',
to_file_flag=True,
level=GlobalConfig().DEFAULT_LOG_LEVEL,
to_file_name=os.path.join(Logger().log_dir,
'console.log'))
self.assertTrue(ConsoleLogger().inited_flag)
self.assertTrue(Logger().inited_flag)
def save(self, save_path, global_step, name=None, default_save_param=None, *args, **kwargs):
if default_save_param is None:
default_save_param = dict(_parameters=self._parameters, _source_config=self._source_config.config_dict)
if not name:
name = self.name
Logger().out_to_file(file_path=save_path,
file_name='{}-{}.json'.format(name, global_step),
content=default_save_param)
def setUp(self):
BaseTestCase.setUp(self)
try:
shutil.rmtree(GlobalConfig().DEFAULT_LOG_PATH)
except FileNotFoundError:
pass
os.makedirs(GlobalConfig().DEFAULT_LOG_PATH)
self.assertFalse(ConsoleLogger().inited_flag)
self.assertFalse(Logger().inited_flag)
def setUp(self):
BaseTestCase.setUp(self)
try:
shutil.rmtree(GlobalConfig().DEFAULT_LOG_PATH)
except FileNotFoundError:
pass
# os.makedirs(GlobalConfig().DEFAULT_LOG_PATH)
# self.assertFalse(ConsoleLogger().inited_flag)
# self.assertFalse(Logger().inited_flag)
Logger().init(config_or_config_dict=GlobalConfig().DEFAULT_LOG_CONFIG_DICT,
log_path=GlobalConfig().DEFAULT_LOG_PATH,
log_level=GlobalConfig().DEFAULT_LOG_LEVEL)
ConsoleLogger().init(logger_name='console_logger',
to_file_flag=True,
level=GlobalConfig().DEFAULT_LOG_LEVEL,
to_file_name=os.path.join(Logger().log_dir, 'console.log'))
self.assertTrue(ConsoleLogger().inited_flag)
self.assertTrue(Logger().inited_flag)
def single_exp_runner(task_fn,
auto_choose_gpu_flag=False,
gpu_id: int = 0,
seed=None,
del_if_log_path_existed=False,
keep_session=False,
**task_fn_kwargs):
"""
:param task_fn: task function defined bu users
:type task_fn: method
:param auto_choose_gpu_flag: auto choose gpu, default False
:type auto_choose_gpu_flag: bool
:param gpu_id: gpu id, default 0
:type gpu_id: int
:param seed: seed generated by system time
:type seed: int
:param del_if_log_path_existed:delete obsolete log file path if existed, by default False
:type del_if_log_path_existed: bool
:param task_fn_kwargs:
:type task_fn_kwargs:
:param keep_session: Whether to keep default session & graph
:type keep_session:
:return:
:rtype:
"""
os.environ['CUDA_DEVICE_ORDER'] = "PCI_BUS_ID"
if auto_choose_gpu_flag is True:
DEVICE_ID_LIST = Gpu.getFirstAvailable()
DEVICE_ID = DEVICE_ID_LIST[0]
os.environ["CUDA_VISIBLE_DEVICES"] = str(DEVICE_ID)
else:
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
if not seed:
seed = int(round(time.time() * 1000)) % (2**32 - 1)
_reset_global_seed(seed, keep_session)
print("create log path at {}".format(GlobalConfig().DEFAULT_LOG_PATH),
flush=True)
file.create_path(path=GlobalConfig().DEFAULT_LOG_PATH,
del_if_existed=del_if_log_path_existed)
Logger().init(config_or_config_dict=dict(),
log_path=GlobalConfig().DEFAULT_LOG_PATH,
log_level=GlobalConfig().DEFAULT_LOG_LEVEL)
ConsoleLogger().init(
to_file_flag=GlobalConfig().DEFAULT_WRITE_CONSOLE_LOG_TO_FILE_FLAG,
to_file_name=os.path.join(
GlobalConfig().DEFAULT_LOG_PATH,
GlobalConfig().DEFAULT_CONSOLE_LOG_FILE_NAME),
level=GlobalConfig().DEFAULT_LOG_LEVEL,
logger_name=GlobalConfig().DEFAULT_CONSOLE_LOGGER_NAME)
task_fn(**task_fn_kwargs)
def test_integration_with_dqn(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name='mlp_q',
name_scope='mlp_q',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
dqn = DQN(env_spec=env_spec,
name='dqn_test',
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=1000,
GAMMA=0.99,
BATCH_SIZE=10,
LEARNING_RATE=0.001,
TRAIN_ITERATION=1,
DECAY=0.5),
value_func=mlp_q)
dqn.init()
st = env.reset()
from baconian.common.sampler.sample_data import TransitionData
a = TransitionData(env_spec)
res = []
for i in range(100):
ac = dqn.predict(obs=st, sess=self.sess, batch_flag=False)
st_new, re, done, _ = env.step(action=ac)
a.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
dqn.append_to_memory(a)
res.append(
dqn.train(batch_data=a,
train_iter=10,
sess=None,
update_target=True)['average_loss'])
res.append(
dqn.train(batch_data=None,
train_iter=10,
sess=None,
update_target=True)['average_loss'])
self.assertTrue(dqn in dqn.recorder._obj_log)
self.assertTrue('average_loss' in dqn.recorder._obj_log[dqn])
self.assertTrue(len(dqn.recorder._obj_log[dqn]['average_loss']) == 2)
self.assertTrue(
np.equal(np.array(res), [
x['log_val']
for x in dqn.recorder._obj_log[dqn]['average_loss']
]).all())
self.assertTrue(len(Logger()._registered_recorders) > 0)
self.assertTrue(dqn.recorder in Logger()._registered_recorders)
Logger().flush_recorder()
def tearDown(self):
Logger().reset()
ConsoleLogger().reset()
BaseTestCase.tearDown(self)
self.assertFalse(ConsoleLogger().inited_flag)
self.assertFalse(Logger().inited_flag)
def test_integration_with_dqn(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
mlp_q = MLPQValueFunction(env_spec=env_spec,
name='mlp_q',
name_scope='mlp_q',
mlp_config=[
{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
},
{
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}
])
dqn = DQN(env_spec=env_spec,
name='dqn_test',
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=1000,
GAMMA=0.99,
BATCH_SIZE=10,
LEARNING_RATE=0.001,
TRAIN_ITERATION=1,
DECAY=0.5),
value_func=mlp_q)
agent = Agent(env=env, env_spec=env_spec,
algo=dqn,
name='agent')
agent.init()
# dqn.init()
st = env.reset()
from baconian.common.sampler.sample_data import TransitionData
a = TransitionData(env_spec)
res = []
agent.sample(env=env,
sample_count=100,
in_which_status='TRAIN',
store_flag=True,
sample_type='transition')
agent.sample(env=env,
sample_count=100,
in_which_status='TRAIN',
store_flag=True,
sample_type='transition')
res.append(dqn.train(batch_data=a, train_iter=10, sess=None, update_target=True)['average_loss'])
res.append(dqn.train(batch_data=None, train_iter=10, sess=None, update_target=True)['average_loss'])
self.assertTrue(dqn in dqn.recorder._obj_log)
self.assertTrue('average_loss' in dqn.recorder._obj_log[dqn])
self.assertTrue(len(dqn.recorder._obj_log[dqn]['average_loss']) == 2)
self.assertTrue(
np.equal(np.array(res), [x['value'] for x in dqn.recorder._obj_log[dqn]['average_loss']]).all())
self.assertTrue(len(Logger()._registered_recorders) > 0)
self.assertTrue(dqn.recorder in Logger()._registered_recorders)
res = dqn.recorder.get_log(attr_name='average_loss', filter_by_status=dict())
self.assertEqual(len(res), 2)
res = agent.recorder.get_log(attr_name='sum_reward', filter_by_status={'status': 'TRAIN'})
self.assertEqual(len(res), 2)
res = agent.recorder.get_log(attr_name='sum_reward', filter_by_status={'status': 'TEST'})
self.assertEqual(len(res), 0)
Logger().flush_recorder()
