def example_function(doodad_config, variant):
x = variant['x']
y = variant['y']
z = variant['z']
with open(doodad_config.output_directory + '/function_output.txt',
"w") as f:
f.write('sum = {}'.format(x + y + z))
print('x, y, z = ', x, y, z)
save_doodad_config(doodad_config)
def offline_experiment(doodad_config, variant):
save_doodad_config(doodad_config)
parser = argparse.ArgumentParser()
# parser.add_argument('--env-type', default='gridworld')
# parser.add_argument('--env-type', default='point_robot_sparse')
# parser.add_argument('--env-type', default='cheetah_vel')
parser.add_argument('--env-type', default='ant_semicircle_sparse')
args, rest_args = parser.parse_known_args(args=[])
env = args.env_type
# --- GridWorld ---
if env == 'gridworld':
args = args_gridworld.get_args(rest_args)
# --- PointRobot ---
elif env == 'point_robot_sparse':
args = args_point_robot_sparse.get_args(rest_args)
# --- Mujoco ---
elif env == 'cheetah_vel':
args = args_cheetah_vel.get_args(rest_args)
elif env == 'ant_semicircle_sparse':
args = args_ant_semicircle_sparse.get_args(rest_args)
set_gpu_mode(torch.cuda.is_available() and args.use_gpu)
vae_args = config_utl.load_config_file(
os.path.join(args.vae_dir, args.env_name, args.vae_model_name,
'online_config.json'))
args = config_utl.merge_configs(
vae_args, args) # order of input to this function is important
# Transform data BAMDP (state relabelling)
if args.transform_data_bamdp:
# load VAE for state relabelling
vae_models_path = os.path.join(args.vae_dir, args.env_name,
args.vae_model_name, 'models')
vae = VAE(args)
off_utl.load_trained_vae(vae, vae_models_path)
# load data and relabel
save_data_path = os.path.join(args.main_data_dir, args.env_name,
args.relabelled_data_dir)
os.makedirs(save_data_path)
dataset, goals = off_utl.load_dataset(data_dir=args.data_dir,
args=args,
arr_type='numpy')
bamdp_dataset = off_utl.transform_mdps_ds_to_bamdp_ds(
dataset, vae, args)
# save relabelled data
off_utl.save_dataset(save_data_path, bamdp_dataset, goals)
learner = OfflineMetaLearner(args)
learner.train()
def run_doodad_experiment(doodad_config: DoodadConfig, params):
# print(params)
# import ipdb; ipdb.set_trace()
save_doodad_config(doodad_config)
log_dir = doodad_config.output_directory
save_variant(log_dir, params)
log_dir = doodad_config.output_directory
exp_name = log_dir.split('/')[-2]
setup_logger(logger,
variant=params,
base_log_dir=None,
exp_name=exp_name,
log_dir=log_dir)
run(log_dir, **params)
def collect_buffers(doodad_config, variant):
save_doodad_config(doodad_config)
random.seed(17)
np.random.seed(17)
tf.set_random_seed(17)
args = get_args(default=True)
args.task_idx = int(variant['task_idx'])
args.task = None
args.full_buffer_size = variant['full_buffer_size']
args.replay_buffer_size = variant['replay_buffer_size']
args.alg = variant['alg']
args.outer_policy_lr = variant['outer_policy_lr']
args.env = variant['env']
args.task_path = '/task_data'
args.log_dir = doodad_config.output_directory
main(args)
def train_vae(doodad_config: DoodadConfig, params):
save_doodad_config(doodad_config)
_train_vae(doodad_config.output_directory, **params)
def borel(doodad_config: DoodadConfig, params):
save_doodad_config(doodad_config)
log_dir = doodad_config.output_directory
exp_name = log_dir.split('/')[-2]
setup_logger(logger, variant=params, base_log_dir=None, exp_name=exp_name, log_dir=log_dir)
_borel(log_dir, **params)