def test_saving_scheduler_on_all_model_free_algo(self):
to_test_algo_func = (self.create_ppo, self.create_dqn, self.create_ddpg)
for func in to_test_algo_func:
self.setUp()
single_exp_runner(_saving_scheduler(self, func), auto_choose_gpu_flag=False, gpu_id=0,
del_if_log_path_existed=True)
self.tearDown()
def test_exp_with_scheduler(self, algo=None, locals=None):
def wrap_algo(algo=None, locals=None):
def func(algo=algo, locals=locals):
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT', dict(TOTAL_AGENT_TRAIN_SAMPLE_COUNT=500,
TOTAL_AGENT_TEST_SAMPLE_COUNT=None,
TOTAL_AGENT_UPDATE_COUNT=None))
if not algo:
algo, locals = self.create_dqn()
env_spec = locals['env_spec']
env = locals['env']
agent = self.create_agent(env=locals['env'],
algo=algo,
name='agent',
eps=self.create_eps(env_spec)[0],
env_spec=env_spec)[0]
exp = self.create_exp(name='model_free', env=env, agent=agent)
algo.parameters.set_scheduler(param_key='LEARNING_RATE',
to_tf_ph_flag=True,
scheduler=LinearSchedule(
t_fn=exp.TOTAL_ENV_STEP_TRAIN_SAMPLE_COUNT,
schedule_timesteps=GlobalConfig().DEFAULT_EXPERIMENT_END_POINT[
'TOTAL_AGENT_TRAIN_SAMPLE_COUNT'],
final_p=0.0001,
initial_p=0.01))
exp.run()
self.assertEqual(exp.TOTAL_AGENT_TEST_SAMPLE_COUNT(), exp.TOTAL_ENV_STEP_TEST_SAMPLE_COUNT())
self.assertEqual(exp.TOTAL_AGENT_TRAIN_SAMPLE_COUNT(), exp.TOTAL_ENV_STEP_TRAIN_SAMPLE_COUNT(), 500)
return func
single_exp_runner(wrap_algo(algo, locals), auto_choose_gpu_flag=False, gpu_id=0,
del_if_log_path_existed=True)
def test_saving_scheduler_on_all_model_free_algo(self):
to_test_algo_func = (self.create_ppo, self.create_dqn,
self.create_ddpg)
sample_traj_flag = (True, False, False)
for i, func in enumerate(to_test_algo_func):
self.setUp()
single_exp_runner(_saving_scheduler(
self, func, sample_traj_flag=sample_traj_flag[i]),
auto_choose_gpu_flag=False,
gpu_id=0,
del_if_log_path_existed=True)
self.tearDown()
def test_experiment(self):
def func():
GlobalConfig().set('DEFAULT_EXPERIMENT_END_POINT', dict(TOTAL_AGENT_TRAIN_SAMPLE_COUNT=200,
TOTAL_AGENT_TEST_SAMPLE_COUNT=None,
TOTAL_AGENT_UPDATE_COUNT=None))
dqn, locals = self.create_dqn()
env_spec = locals['env_spec']
env = locals['env']
agent = self.create_agent(env=locals['env'],
algo=dqn,
name='agent',
eps=self.create_eps(env_spec)[0],
env_spec=env_spec)[0]
exp = self.create_exp(name='model_free', env=env, agent=agent)
exp.run()
single_exp_runner(func, auto_choose_gpu_flag=False, gpu_id=0, del_if_log_path_existed=True)
start_test_after_sample_count=5,
start_train_after_sample_count=5,
train_func_and_args=(agent.train, (), dict()),
test_func_and_args=(agent.test, (), dict(sample_count=10)),
sample_func_and_args=(agent.sample, (), dict(sample_count=100,
env=agent.env,
store_flag=True))
)
experiment = Experiment(
tuner=None,
env=env,
agent=agent,
flow=flow,
name=name + 'experiment_debug'
)
dqn.parameters.set_scheduler(param_key='LEARNING_RATE',
scheduler=LinearScheduler(
t_fn=experiment.TOTAL_AGENT_TRAIN_SAMPLE_COUNT,
schedule_timesteps=GlobalConfig().DEFAULT_EXPERIMENT_END_POINT[
'TOTAL_AGENT_TRAIN_SAMPLE_COUNT'],
final_p=0.0001,
initial_p=0.01))
experiment.run()
from baconian.core.experiment_runner import single_exp_runner
GlobalConfig().set('DEFAULT_LOG_PATH', './log_path')
single_exp_runner(task_fn, del_if_log_path_existed=True)
noise_adder=noise_adder,
name=name + '_agent')
flow = create_train_test_flow(env=env,
cyber=cyber,
agent=agent,
num_test=num_test,
total_steps=total_steps,
max_step_per_episode=max_step_per_episode,
train_after_step=train_after_step,
test_after_step=test_after_step,
train_every_step=train_every_step,
test_every_step=test_every_step,
train_func_and_args=(agent.train, (),
dict()),
test_func_and_args=(agent.test, (), dict()),
sample_func_and_args=(agent.sample, (),
dict()),
flow_type='DDPG_TrainTestFlow')
experiment = Experiment(tuner=None,
env=env,
agent=agent,
flow=flow,
name=name)
experiment.run()
if __name__ == '__main__':
single_exp_runner(task_fn, del_if_log_path_existed=True, keep_session=True)
