def pack_condition_features(episode_data, idx):
"""Pack previous episode data into condition_ep* features/labels.
Args:
episode_data: List of (obs, action, rew, new_obs, done, debug) tuples.
idx: Index of the conditioning episode. 0 for demo, 1 for first trial,
etc.
"""
# Conditioning Context (The assumption is that policy is always adapting
# from some conditioning data, whether it is demo and/or trials.
episode_data = episode_to_transitions.make_fixed_length(
episode_data, fixed_length)
# Condition features.
batch_obs = np.stack([t[0].image for t in episode_data])
batch_gripper = np.stack([t[0].pose for t in episode_data])
meta_features['condition/features/image/%d' % idx] = batch_obs.astype(
np.uint8)
meta_features['condition/features/gripper_pose/%d' %
idx] = batch_gripper.astype(np.float32)
# Condition label.
batch_action = np.stack([t[1] for t in episode_data])
meta_features['condition/labels/action/%d' %
idx] = batch_action.astype(np.float32)
def pack_condition_features(episode_data, idx):
"""Pack previous episode data into condition_ep* features/labels.
Args:
episode_data: List of (obs, action, rew, new_obs, done, debug) tuples.
idx: Index of the conditioning episode. 0 for demo, 1 for first trial,
etc.
"""
# Conditioning Context (The assumption is that policy is always adapting
# from some conditioning data, whether it is demo and/or trials.
episode_data = episode_to_transitions.make_fixed_length(
episode_data, fixed_length, randomized=not deterministic_condition)
# Condition features.
if vision:
batch_obs = np.stack([t[0].image for t in episode_data])
batch_gripper = np.stack([t[0].pose for t in episode_data])
meta_features['condition/features/image/%d' %
idx] = batch_obs.astype(np.uint8)
meta_features['condition/features/gripper_pose/%d' %
idx] = batch_gripper.astype(np.float32)
else:
batch_full_state = np.stack(
[t[0].full_state_pose for t in episode_data])
meta_features['condition/features/full_state_pose/%d' %
idx] = batch_full_state.astype(np.float32)
# Condition label.
batch_action = np.stack([t[1] for t in episode_data])
meta_features['condition/labels/action/%d' %
idx] = batch_action.astype(np.float32)
cumulative_return = np.sum([t[2] for t in episode_data])
success = float(cumulative_return > 0) * np.ones(
(40, 1), dtype=np.float32)
meta_features['condition/labels/success/%d' % idx] = success
def test_make_fixed_length(self):
fixed_length = 10
dummy_feature_dict_lists = [
[{'dummy_feature': i} for i in range(5)],
[{'dummy_feature': i} for i in range(20)],
]
for feature_dict_list in dummy_feature_dict_lists:
filtered_feature_dict_list = episode_to_transitions.make_fixed_length(
feature_dict_list,
fixed_length=fixed_length,
always_include_endpoints=True)
self.assertLen(filtered_feature_dict_list, fixed_length)
# The first and last entries of the original list should be present in
# the filtered list.
self.assertEqual(feature_dict_list[0], filtered_feature_dict_list[0])
self.assertEqual(feature_dict_list[-1], filtered_feature_dict_list[-1])
