from Environments.env_utils import make_env_utils
from NeuralNetwork.neural_net import NeuralNet
from PPO_trainer import PPO_trainer, evaluate
# Create parser and parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--parameters', type=str)
parser.add_argument('--logdir', type=str, default=None)
args = parser.parse_args()
# Open the json parameters file
with open(args.parameters) as f:
parameters = json.load(f)
# Create env and env_utils
env_utils, env = make_env_utils(**parameters)
# Get network kwargs from env_utils
network_kwargs = env_utils.get_network_kwargs()
# Add additional kwargs from parameter file
network_kwargs.update(parameters['neural_network'])
ac = NeuralNet(**network_kwargs)
if parameters['mode'] == 'training':
# Create the directories for logs and saved models
parent_dir = os.path.dirname(args.logdir)
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
models_dir = os.path.dirname(args.logdir + 'saved_models/')
if not os.path.exists(models_dir):
def __init__(
self,
starting_map_size=(10., 10., 2.),
cell_scale=(1., 1., 1.),
starting_position=(0., 0., 0.),
buffer_distance=(10., 10., 0.),
local_map_dim=(16, 16, 1),
vision_range=1,
fov_angle=np.pi / 2,
map_keys=[
'unknown', 'visible', 'visited', 'obstacle', 'object', 'position'
],
thresholds={
'visible': 1,
'visited': 1,
'obstacle': 1,
'object': 1
},
map_idx=0,
interactive_plot=False,
REWARD_FAILURE=-10,
REWARD_STEP=-0.5,
**parameters,
):
super().__init__(
starting_map_size=starting_map_size,
cell_scale=cell_scale,
starting_position=starting_position,
buffer_distance=buffer_distance,
local_map_dim=local_map_dim,
vision_range=vision_range,
fov_angle=fov_angle,
map_keys=map_keys,
thresholds=thresholds,
map_idx=map_idx,
interactive_plot=interactive_plot,
REWARD_FAILURE=REWARD_FAILURE,
REWARD_STEP=REWARD_STEP,
)
parameters_airsim = {'env_str': 'AirSim'}
parameters_airsim['AirSim'] = parameters['AirSim']
self.env_utils_airsim, self.env_airsim = make_env_utils(
**parameters_airsim)
network_kwargs = self.env_utils_airsim.get_network_kwargs()
network_kwargs.update(
parameters['Exploration']['local_navigation']
['neural_network']) # add additional kwargs from parameter file
self.local_navigator = NeuralNet(**network_kwargs)
self.local_navigator.load_state_dict(
torch.load(
parameters['Exploration']['local_navigation']['weights']))
self.navigator_max_steps = parameters['navigator_max_steps']
self.object_detection_frequency = parameters[
'object_detection_frequency']
self.obstacle_detection_frequency = parameters[
'obstacle_detection_frequency']
self.env_airsim.setup_object_detection(
**parameters['object_detection'])
self.reward_scaling = (self.vision_range / self.cell_scale[0]) * \
(self.vision_range / self.cell_scale[1]) * \
(self.vision_range / self.cell_scale[2]) * fov_angle / 2 # divide by area