def main():
'''Main script function.'''
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
parser = argparse.ArgumentParser()
parser.add_argument("path", help="The path to save the results.")
parser.add_argument("--alg", help="The algorithm to use.", default='PPO')
parser.add_argument("--env_name", help="The gamma to use.",
default='MultiOptLRs-v0')
parser.add_argument('--total_timesteps', default=int(1e5), type=int,
help="Number of timesteps per training session")
parser.add_argument('--data_set', help="The data set to use.",
default='iris')
parser.add_argument('--trials', help="The number of trials to run.",
default=10, type=int)
args = parser.parse_args()
parameters = vars(args).copy()
del parameters['trials']
path = Path(parameters['path'])
if not path.exists():
path.mkdir()
parameters['kwargs'] = {'data_set': parameters['data_set']}
utils_file.save_json(parameters, path / 'parameters.json')
else:
if (path / 'study.db').exists():
print('Directory exists. Using existing study and parameters.')
parameters = utils_file.load_json(path / 'parameters.json')
else:
raise FileExistsError(('Directory already exists and is not a '
'study.'))
storage = 'sqlite:///' + str(path / 'study.db')
study = optuna.create_study(
study_name=str(path.name),
storage=storage, load_if_exists=True,
pruner=optuna.pruners.MedianPruner()
)
print(study.best_value)
print(study.best_trial.number)
print(study.best_params)
#params = study.best_params
param_file = path / str(study.best_trial.number) / 'parameters.json'
params = utils_file.load_json(param_file)
kwargs = params['kwargs']
parameters.update({
'path': path / str(study.best_trial.number),
'gamma': params['gamma'],
'learning_rate': params['learning_rate'],
'kwargs': {
#'batch_size': int(params['batch_size']),
'max_batches': int(kwargs['max_batches']),
'max_history': int(kwargs['max_history']),
'max_batches': 1e3
}
})
live_plot(run_experiment(parameters), 'test')
def test_load_json(save_path):
'''Test load_json function.'''
path = create_test_file(save_path, name='test_load_json0')
dictionary = utils_file.load_json(path)
for i, (name, value) in enumerate(dictionary.items()):
assert i == value
assert name == chr(97 + i)
path = create_test_file(save_path, name='test_load_json1', version=1)
dictionary = utils_file.load_json(path, line_no=5)
assert len(dictionary) == 26 - 5
for i, (name, value) in enumerate(dictionary.items()):
assert i == value
assert name == chr(97 + i)
def main():
'''Main script function.'''
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
parser = argparse.ArgumentParser()
parser.add_argument("path", help="The path to save the results.")
parser.add_argument("--alg", help="The algorithm to use.",
default='MADDPG', type=str.upper)
parser.add_argument("--env_name", help="The gamma to use.",
default='simple_spread')
parser.add_argument('--total_timesteps', default=int(1e6), type=int,
help="Number of timesteps per training session")
parser.add_argument('--trials', help="The number of trials to run.",
default=10, type=int)
args = parser.parse_args()
parameters = vars(args).copy()
del parameters['trials']
path = Path(parameters['path'])
if not path.exists():
path.mkdir()
utils_file.save_json(parameters, path / 'parameters.json')
else:
if (path / 'study.db').exists():
print('Directory exists. Using existing study and parameters.')
parameters = utils_file.load_json(path / 'parameters.json')
else:
raise FileExistsError(('Directory already exists and is not a '
'study.'))
objective = partial(run_experiment, parameters)
storage = 'sqlite:///' + str(path / 'study.db')
study = optuna.create_study(study_name=str(path.name),
storage=storage, load_if_exists=True,
pruner=optuna.pruners.MedianPruner())
study.optimize(objective, n_trials=args.trials)
def test_save_json(save_path):
'''Test save_json function.'''
data = {chr(97 + i): i for i in range(26)}
path = save_path / 'test_save_json.json'
utils_file.save_json(data, path)
dictionary = utils_file.load_json(path)
for i, (name, value) in enumerate(dictionary.items()):
assert i == value
assert name == chr(97 + i)
def main():
'''Evaluate a trained model against logistic regression.'''
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
parser = argparse.ArgumentParser()
parser.add_argument("path", help="The path to save the results.")
parser.add_argument("--trials",
help="The number of trials.",
type=int,
default=10)
parser.add_argument("--total_epochs",
help="The number of epochs.",
type=int,
default=40)
parser.add_argument("--data_set",
help="The data set to trial against.",
type=str,
default='iris')
args = parser.parse_args()
parameters = vars(args).copy()
del parameters['trials']
path = Path(parameters['path'])
if not path.exists():
path.mkdir()
utils_file.save_json(parameters, path / 'parameters.json')
else:
if (path / 'study.db').exists():
print('Directory exists. Using existing study and parameters.')
parameters = utils_file.load_json(path / 'parameters.json')
else:
raise FileExistsError(('Directory already exists and is not a '
'study.'))
objective = partial(train_model, parameters)
storage = 'sqlite:///' + str(path / 'study.db')
study = optuna.create_study(study_name=str(path.name),
storage=storage,
load_if_exists=True,
pruner=optuna.pruners.MedianPruner())
study.optimize(objective, n_trials=args.trials)
def task(path, seed, batch_size=None, total_epochs=40, data_set='mnist'):
'''
Run the agent on a data set.
'''
parameters = utils_file.load_json(path / 'hyperparams.json')
alg, *_ = path.name.split('-')
save_path = path / 'model.pkl'
sequence = load_data(data_set)
num_of_samples = len(sequence.features)
steps_per_epoch = ceil(num_of_samples / batch_size) if batch_size else 1
kwargs = parameters.get('kwargs', {})
kwargs['data_set'] = data_set
kwargs['batch_size'] = batch_size
kwargs['max_batches'] = steps_per_epoch * total_epochs
env = partial(gym.make, parameters['env_name'], **kwargs)
vec_env = OptVecEnv([env])
if alg == 'PPO':
with open(save_path, 'rb') as pkl:
model = PPO2.load(pkl) # , env=vec_env)
elif alg == 'A2C':
with open(save_path, 'rb') as pkl:
model = A2C.load(pkl, env=vec_env)
elif alg == 'DDPG':
model = DDPG.load(save_path, env=vec_env)
states = vec_env.reset()
info_list = []
cumulative_reward = 0
for epoch_no in trange(total_epochs, leave=False):
for step in trange(steps_per_epoch, leave=False):
actions = model.predict(states, deterministic=False)[0]
states, rewards, _, infos = vec_env.step(actions)
cumulative_reward = cumulative_reward + rewards[0]
info = infos[0]
info['step'] = epoch_no * steps_per_epoch + step
info['cumulative_reward'] = cumulative_reward
info['seed'] = seed
info['epoch'] = epoch_no
info_list.append(info)
return info_list