def test_pusher_slider_keypoint_dataset():
project_root = get_project_root()
config_file = os.path.join(project_root, "experiments/02/config.yaml")
config = load_yaml(config_file)
config["n_history"] = 1
config["n_roll"] = 0
# new dataset loading approach
episodes = load_episodes_from_config(config)
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
dataset = MultiEpisodeDataset(config,
action_function=action_function,
observation_function=observation_function,
episodes=episodes,
phase="train")
# dataset, config = create_pusher_slider_keypoint_dataset(config=config)
episode_names = dataset.get_episode_names()
episode_names.sort()
episode_name = episode_names[0]
episode = dataset.episode_dict[episode_name]
obs_raw = episode.get_observation(0)
obs_raw['slider']['angle'] = 0
dataset.observation_function(obs_raw)
print("20 degrees\n\n\n\n")
obs_raw['slider']['angle'] = np.deg2rad(90)
dataset.observation_function(obs_raw)
quit()
data = dataset[0] # test the getitem
print("type(data)", type(data))
print("data.keys()", data.keys())
print(type(data["observations"]))
print("observations.shape", data["observations"].shape)
print("actions.shape", data["actions"].shape)
print("observations", data["observations"])
print("actions", data["actions"])
def eval_dynamics(config,
train_dir,
eval_dir,
state_dict_path=None,
keypoint_observation=False,
debug=False,
render_human=False):
# set random seed for reproduction
set_seed(config['train']['random_seed'])
tee = Tee(os.path.join(eval_dir, 'eval.log'), 'w')
print(config)
use_gpu = torch.cuda.is_available()
'''
model
'''
model_dy = DynaNetMLP(config)
# print model #params
print("model #params: %d" % count_trainable_parameters(model_dy))
if state_dict_path is None:
if config['eval']['eval_dy_epoch'] == -1:
state_dict_path = os.path.join(train_dir, 'net_best_dy.pth')
else:
state_dict_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d.pth' % \
(config['eval']['eval_dy_epoch'], config['eval']['eval_dy_iter']))
print("Loading saved ckp from %s" % state_dict_path)
model_dy.load_state_dict(torch.load(state_dict_path))
model_dy.eval()
if use_gpu:
model_dy.cuda()
criterionMSE = nn.MSELoss()
bar = ProgressBar()
st_idx = config['eval']['eval_st_idx']
ed_idx = config['eval']['eval_ed_idx']
# load the data
episodes = load_episodes_from_config(config)
# generate action/observation functions
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
dataset = MultiEpisodeDataset(config,
action_function=action_function,
observation_function=observation_function,
episodes=episodes,
phase="valid")
episode_names = dataset.get_episode_names()
episode_names.sort()
num_episodes = None
# for backwards compatibility
if "num_episodes" in config["eval"]:
num_episodes = config["eval"]["num_episodes"]
else:
num_episodes = 10
episode_list = []
if debug:
episode_list = [episode_names[0]]
else:
episode_list = episode_names[:num_episodes]
for roll_idx, episode_name in enumerate(episode_list):
print("episode_name", episode_name)
if keypoint_observation:
eval_episode_keypoint_observations(config,
dataset,
episode_name,
roll_idx,
model_dy,
eval_dir,
start_idx=9,
n_prediction=30,
render_human=render_human)
else:
eval_episode(config,
dataset,
episode_name,
roll_idx,
model_dy,
eval_dir,
start_idx=9,
n_prediction=30,
render_human=render_human)
from key_dynam.utils import dev_utils
from key_dynam.dataset.episode_dataset import MultiEpisodeDataset
from key_dynam.dataset.function_factory import ObservationFunctionFactory, ActionFunctionFactory
multi_episode_dict = dev_utils.load_drake_pusher_slider_episodes()
config = dev_utils.load_simple_config()
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(config)
dataset = MultiEpisodeDataset(config,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase="train")
episode_name = dataset.get_episode_names()[0]
episode = dataset.episode_dict[episode_name]
idx = 5
data = dataset._getitem(
episode,
idx,
rollout_length=5,
n_history=2,
visual_observation=False,
)
print("\n\ndata.keys()", data.keys())
data_w_vision = dataset._getitem(
episode,
idx,