def get_metric_dictionary(self):
metrics_dict = {}
class_recall_list = [] # used for calculating mean pixel accuracy.
class_iou_list = [] # used for calculating mean iou.
for c in self.class_range:
tp = self.true_positive_metrics[c].result()
fp = self.false_positive_metrics[c].result()
fn = self.false_negative_metrics[c].result()
class_recall = tp / (tp + fn)
class_precision = tf.where(
tf.greater(tp + fn, 0.0), _safe_div(tp, (tp + fp)),
tf.constant(np.NaN))
class_iou = tf.where(
tf.greater(tp + fn, 0.0), tp / (tp + fn + fp), tf.constant(np.NaN))
class_recall_list.append(class_recall)
class_iou_list.append(class_iou)
class_name = _get_class_name(class_id=c, label_map=self.label_map)
metrics_dict[self.eval_prefix +
'_recall/{}'.format(class_name)] = class_recall
metrics_dict[self.eval_prefix +
'_precision/{}'.format(class_name)] = class_precision
metrics_dict[self.eval_prefix + '_iou/{}'.format(class_name)] = class_iou
mean_pixel_accuracy = _non_nan_mean(class_recall_list)
mean_iou = _non_nan_mean(class_iou_list)
metrics_dict[self.eval_prefix +
'_avg/mean_pixel_accuracy'] = mean_pixel_accuracy
metrics_dict[self.eval_prefix + '_avg/mean_iou'] = mean_iou
return metrics_dict
def _observation_cost(obs):
c_theta, s_theta, d_theta = obs[:, :1], obs[:, 1:2], obs[:, 2:3]
theta = tf.math.atan2(s_theta, c_theta)
cost = tf.reduce_sum(tf.square(theta) + 0.1 * tf.square(d_theta),
axis=1)
cost = tf.where(tf.math.is_nan(cost), 1e6 * tf.ones_like(cost), cost)
return cost
def select_slate_greedy(slate_size, s_no_click, s, q):
"""Selects the slate using the adaptive greedy algorithm.
This algorithm corresponds to the method "GS" in
Ie et al. https://arxiv.org/abs/1905.12767.
Args:
slate_size: int, the size of the recommendation slate.
s_no_click: float tensor, the score for not clicking any document.
s: [num_of_documents] tensor, the scores for clicking documents.
q: [num_of_documents] tensor, the predicted q values for documents.
Returns:
[slate_size] tensor, the selected slate.
"""
def argmax(v, mask):
return tf.argmax((v - tf.reduce_min(v) + 1) * mask, axis=0)
numerator = tf.constant(0.)
denominator = tf.constant(0.) + s_no_click
mask = tf.ones(tf.shape(q)[0])
def set_element(v, i, x):
mask = tf.one_hot(i, tf.shape(v)[0])
v_new = tf.ones_like(v) * x
return tf.where(tf.equal(mask, 1), v_new, v)
for _ in range(slate_size):
k = argmax((numerator + s * q) / (denominator + s), mask)
mask = set_element(mask, k, 0)
numerator = numerator + tf.gather(s * q, k)
denominator = denominator + tf.gather(s, k)
output_slate = tf.where(tf.equal(mask, 0))
return output_slate
def class_specific_data(onehot_labels, data, num_classes, axis=0):
# TODO(eringrant): Deal with case of no data for a class in [1...num_classes].
data_shape = [s for i, s in enumerate(data.shape) if i != axis]
labels = tf.argmax(onehot_labels, axis=-1)
class_idx = [tf.where(tf.equal(labels, i)) for i in range(num_classes)]
return [
tf.reshape(tf.gather(data, idx, axis=axis), [-1] + data_shape)
for idx in class_idx
]
def filter_before_first_step(time_steps, actions=None):
flat_time_steps = tf.nest.flatten(time_steps)
flat_time_steps = [tf.unstack(time_step, axis=1) for time_step in
flat_time_steps]
time_steps = [tf.nest.pack_sequence_as(time_steps, time_step) for time_step in
zip(*flat_time_steps)]
if actions is None:
actions = [None] * len(time_steps)
else:
actions = tf.unstack(actions, axis=1)
assert len(time_steps) == len(actions)
time_steps = list(reversed(time_steps))
actions = list(reversed(actions))
filtered_time_steps = []
filtered_actions = []
for t, (time_step, action) in enumerate(zip(time_steps, actions)):
if t == 0:
reset_mask = tf.equal(time_step.step_type, ts.StepType.FIRST)
else:
time_step = tf.nest.map_structure(lambda x, y: tf.where(reset_mask, x, y),
last_time_step, time_step)
action = tf.where(reset_mask, tf.zeros_like(action),
action) if action is not None else None
filtered_time_steps.append(time_step)
filtered_actions.append(action)
reset_mask = tf.logical_or(
reset_mask,
tf.equal(time_step.step_type, ts.StepType.FIRST))
last_time_step = time_step
filtered_time_steps = list(reversed(filtered_time_steps))
filtered_actions = list(reversed(filtered_actions))
filtered_flat_time_steps = [tf.nest.flatten(time_step) for time_step in
filtered_time_steps]
filtered_flat_time_steps = [tf.stack(time_step, axis=1) for time_step in
zip(*filtered_flat_time_steps)]
filtered_time_steps = tf.nest.pack_sequence_as(filtered_time_steps[0],
filtered_flat_time_steps)
if action is None:
return filtered_time_steps
else:
actions = tf.stack(filtered_actions, axis=1)
return filtered_time_steps, actions
def __init__(self, observations, env_spec):
with tf.name_scope('fully_conv_model'):
spatial_streams = {
name: spatial_stream(observations[name], spec)
for name, spec in env_spec.observation_spec.items()
if spec.is_spatial
}
fc = Concatenate()(
[Flatten()(x) for x in spatial_streams.values()])
fc = Dense(
256,
activation='relu',
name='fc',
kernel_initializer=tf.keras.initializers.Orthogonal())(fc)
with tf.name_scope('policy'):
self.policy = {}
for name, spec in env_spec.action_spec.items():
with tf.name_scope(name):
if spec.obs_space:
logits = Conv2D(
1,
1,
activation='linear',
data_format='channels_first',
kernel_initializer=tf.keras.initializers.
Orthogonal(gain=0.1))(
spatial_streams[spec.obs_space])
logits = Flatten()(logits)
else:
logits = Dense(
np.prod(spec.sizes),
activation='linear',
kernel_initializer=tf.keras.initializers.
Orthogonal(gain=0.1))(fc)
if name == 'function_id':
logits = tf.where(
observations['available_actions'] > 0,
logits,
-1000 * tf.ones_like(logits),
name='mask_unavailable_functions')
self.policy[name] = tfp.distributions.Categorical(
logits=logits)
with tf.name_scope('actions'):
self.actions = {
name: dist.sample(name=name + '_sample')
for name, dist in self.policy.items()
}
with tf.name_scope('value'):
self.value = value_output(fc)
def selu(x):
"""
SELU activation
https://arxiv.org/abs/1706.02515
:param x:
:return:
"""
with tf.name_scope('elu') as scope:
alpha = 1.6732632423543772848170429916717
scale = 1.0507009873554804934193349852946
return scale * tf.where(x >= 0.0, x, alpha * tf.nn.elu(x))
def historgram_loss(y, y_hat, k=100., sigma=1 / 2):
raise NotImplementedError()
ps = 0.
w = 1 / k
y = tf.squeeze(y, axis=2)
# y_hat = tf.layers.flatten(y_hat)
k = np.linspace(0., 1., k)
s = (tf.erf((1. - y) / (tf.sqrt(2.) * sigma)) - tf.erf((0. - y) / (tf.sqrt(2.) * sigma)))
for idx, j in enumerate(k):
u = tf.erf(((j + w - y) / (tf.sqrt(2.) * sigma)))
l = tf.erf(((j - y) / (tf.sqrt(2.) * sigma)))
p = (u - l) / (2 * s + 1e-6)
f_x = tf.log(y_hat[:, :, idx])
ps += p * tf.where(tf.is_nan(f_x), tf.zeros_like(f_x), f_x)
return tf.reduce_mean(-ps)
def _build_train_op(self):
"""Builds a training op.
Returns:
An op performing one step of training from replay data.
"""
# click_indicator: [B, S]
# q_values: [B, A]
# actions: [B, S]
# slate_q_values: [B, S]
# replay_click_q: [B]
click_indicator = self._replay.rewards[:, :,
self._click_response_index]
slate_q_values = tf.batch_gather(
self._replay_net_outputs.q_values,
tf.cast(self._replay.actions, dtype=tf.int32))
# Only get the Q from the clicked document.
replay_click_q = tf.reduce_sum(input_tensor=slate_q_values *
click_indicator,
axis=1,
name='replay_click_q')
target = tf.stop_gradient(self._build_target_q_op())
clicked = tf.reduce_sum(input_tensor=click_indicator, axis=1)
clicked_indices = tf.squeeze(tf.where(tf.equal(clicked, 1)), axis=1)
# clicked_indices is a vector and tf.gather selects the batch dimension.
q_clicked = tf.gather(replay_click_q, clicked_indices)
target_clicked = tf.gather(target, clicked_indices)
def get_train_op():
loss = tf.reduce_mean(input_tensor=tf.square(q_clicked -
target_clicked))
if self.summary_writer is not None:
with tf.variable_scope('Losses'):
tf.summary.scalar('Loss', loss)
return loss
loss = tf.cond(pred=tf.greater(tf.reduce_sum(input_tensor=clicked), 0),
true_fn=get_train_op,
false_fn=lambda: tf.constant(0.),
name='')
return self.optimizer.minimize(loss)
def policy_output(state, available_actions, action_spec):
def logits_output(num_categories, name):
return Dense(num_categories,
activation='linear',
name=name + '_logits')(state)
logits = [
logits_output(np.prod(spec.sizes), name)
for name, spec in action_spec.items()
]
logits[0] = tf.where(available_actions > 0,
logits[0],
-1000 * tf.ones_like(logits[0]),
name='mask_unavailable_functions')
dists = {
name: tfp.distributions.Categorical(logits=logits[spec.id],
name=name + '_dist')
for name, spec in action_spec.items()
}
return dists
def set_element(v, i, x):
mask = tf.one_hot(i, tf.shape(input=v)[0])
v_new = tf.ones_like(v) * x
return tf.where(tf.equal(mask, 1), v_new, v)
def compute_motion_labels(scene,
frame0,
frame1,
frame_start_index,
points_key,
box_margin=0.1):
"""Compute motion label for each point.
Args:
scene: dict of tensor containing scene.
frame0: dict of tensor containing points and objects.
frame1: dict of tensor containing points and objects.
frame_start_index: starting frame index.
points_key: A string corresponding to the tensor of point positions in
inputs.
box_margin: A margin value to enlarge box, so that surrounding points are
included.
Returns:
A motion tensor of [N, 3] shape.
"""
point_positions = frame0[points_key]
frame0_object_names = frame0['objects/name']
frame1_object_names = frame1['objects/name']
bool_matrix = tf.math.equal(
tf.expand_dims(frame0_object_names, axis=1),
tf.expand_dims(frame1_object_names, axis=0))
match_indices = tf.where(bool_matrix)
# object box level
box_dimension = tf.gather(
frame0['objects/shape/dimension'], match_indices[:, 0], axis=0)
boxes_length = box_dimension[:, 0:1]
boxes_width = box_dimension[:, 1:2]
boxes_height = box_dimension[:, 2:3]
boxes_rotation_matrix = tf.gather(
frame0['objects/pose/R'], match_indices[:, 0], axis=0)
boxes_center = tf.gather(
frame0['objects/pose/t'], match_indices[:, 0], axis=0)
frame1_box_rotation_matrix = tf.gather(
frame1['objects/pose/R'], match_indices[:, 1], axis=0)
frame1_box_center = tf.gather(
frame1['objects/pose/t'], match_indices[:, 1], axis=0)
# frame level
frame0_rotation = scene['frames/pose/R'][frame_start_index]
frame1_rotation = scene['frames/pose/R'][frame_start_index + 1]
frame0_translation = scene['frames/pose/t'][frame_start_index]
frame1_translation = scene['frames/pose/t'][frame_start_index + 1]
frame1_box_center_global = tf.tensordot(
frame1_box_center, frame1_rotation, axes=(1, 1)) + frame1_translation
frame1_box_center_in_frame0 = tf.tensordot(
frame1_box_center_global - frame0_translation,
frame0_rotation,
axes=(1, 0))
# only find index on boxes that are matched between two frames
points_box_index = box_utils.map_points_to_boxes(
points=point_positions,
boxes_length=boxes_length,
boxes_height=boxes_height,
boxes_width=boxes_width,
boxes_rotation_matrix=boxes_rotation_matrix,
boxes_center=boxes_center,
box_margin=box_margin)
# TODO(huangrui): disappered object box have 0 motion.
# Probably consider set to nan or ignore_label.
# 1. gather points in surviving matched box only,
# and replicate rotation/t to same length;
# 2. get points in box frame, apply new rotation/t per point;
# 3. new location minus old location -> motion vector;
# 4. scatter it to a larger motion_vector with 0 for
# points ouside of matched boxes.
# Need to limit boxes to those matched boxes.
# otherwise the points_box_index will contain useless box.
# index in all point array, of points that are inside the box.
points_inside_box_index = tf.where(points_box_index + 1)[:, 0]
box_index = tf.gather(points_box_index, points_inside_box_index)
points_inside_box = tf.gather(point_positions, points_inside_box_index)
box_rotation_per_point = tf.gather(boxes_rotation_matrix, box_index)
box_center_per_point = tf.gather(boxes_center, box_index)
# Tensor [N, 3, 3] and [N, 3]. note we are transform points reversely.
points_in_box_frame = tf.einsum('ikj,ik->ij', box_rotation_per_point,
points_inside_box - box_center_per_point)
# Transform rotation of box from frame1 coordinate to frame0 coordinate
# note, transpose is implemented via changing summation axis
frame1_box_rotation_matrix_global = tf.transpose(
tf.tensordot(frame1_rotation, frame1_box_rotation_matrix, axes=(1, 1)),
perm=(1, 0, 2))
frame1_box_rotation_matrix_in_frame0 = tf.transpose(
tf.tensordot(
frame0_rotation, frame1_box_rotation_matrix_global, axes=(0, 1)),
perm=(1, 0, 2))
# this is the points_position_after_following_frame1_box's motion.
frame1_box_rotation_in_frame0_per_point = tf.gather(
frame1_box_rotation_matrix_in_frame0, box_index)
frame1_box_center_in_frame0_per_point = tf.gather(frame1_box_center_in_frame0,
box_index)
points_in_box_frame1 = tf.einsum(
'ijk,ik->ij', frame1_box_rotation_in_frame0_per_point,
points_in_box_frame) + frame1_box_center_in_frame0_per_point
motion_vector = points_in_box_frame1 - points_inside_box
scattered_vector = tf.scatter_nd(
indices=tf.expand_dims(points_inside_box_index, axis=1),
updates=motion_vector,
shape=tf.shape(point_positions, out_type=tf.dtypes.int64))
return scattered_vector
def prepare_lidar_images_and_correspondences(
inputs,
resized_image_height,
resized_image_width,
camera_names=('front', 'front_left', 'front_right', 'side_left',
'side_right'),
lidar_names=('top', 'front', 'side_left', 'side_right', 'rear')):
"""Integrates and returns the lidars, cameras and their correspondences.
Args:
inputs: A dictionary containing the images and point / pixel
correspondences.
resized_image_height: Target height of the images.
resized_image_width: Target width of the images.
camera_names: List of cameras to include images from.
lidar_names: List of lidars to include point clouds from.
Returns:
A tf.float32 tensor of size [num_points, 3] containing point positions.
A tf.float32 tensor of size [num_points, 1] containing point intensities.
A tf.float32 tensor of size [num_points, 1] containing point elongations.
A tf.float32 tensor of size [num_points, 3] containing point normals.
A tf.float32 tensor of size [num_images, resized_image_height,
resized_image_width, 3].
A tf.int32 tensor of size [num_images, num_points, 2].
Raises:
ValueError: If camera_names or lidar_names are empty lists.
"""
if not camera_names:
raise ValueError('camera_names should contain at least one name.')
if not lidar_names:
raise ValueError('lidar_names should contain at least one name.')
(points_position, points_intensity, points_elongation, points_normal,
points_in_image_frame_yx, points_in_image_frame_id) = _prepare_lidar_points(
inputs=inputs, lidar_names=lidar_names)
images = []
points_in_image_frame = []
for camera_name in camera_names:
image_key = ('cameras/%s/image' % camera_name)
image_height = tf.shape(inputs[image_key])[0]
image_width = tf.shape(inputs[image_key])[1]
height_ratio = tf.cast(
resized_image_height, dtype=tf.float32) / tf.cast(
image_height, dtype=tf.float32)
width_ratio = tf.cast(
resized_image_width, dtype=tf.float32) / tf.cast(
image_width, dtype=tf.float32)
if tf.executing_eagerly():
resize_method = tf.image.ResizeMethod.NEAREST_NEIGHBOR
else:
resize_method = tf.image.ResizeMethod.BILINEAR
if inputs[image_key].dtype in [
tf.int8, tf.uint8, tf.int16, tf.uint16, tf.int32, tf.int64
]:
resize_method = tf.image.ResizeMethod.NEAREST_NEIGHBOR
images.append(
tf.image.resize(
images=inputs[image_key],
size=[resized_image_height, resized_image_width],
method=resize_method,
antialias=True))
camera_id = tf.cast(inputs[('cameras/%s/id' % camera_name)], dtype=tf.int32)
valid_points = tf.equal(points_in_image_frame_id, camera_id)
valid_points = tf.tile(valid_points, [1, 2])
point_coords = tf.cast(
tf.cast(points_in_image_frame_yx, dtype=tf.float32) *
tf.stack([height_ratio, width_ratio]),
dtype=tf.int32)
points_in_image_frame_camera = tf.where(
valid_points, point_coords, -tf.ones_like(valid_points, dtype=tf.int32))
points_in_image_frame.append(points_in_image_frame_camera)
num_images = len(images)
images = tf.stack(images, axis=0)
images.set_shape([num_images, resized_image_height, resized_image_width, 3])
points_in_image_frame = tf.stack(points_in_image_frame, axis=0)
return {
'points_position': points_position,
'points_intensity': points_intensity,
'points_elongation': points_elongation,
'points_normal': points_normal,
'view_images': {'rgb_view': images},
'view_indices_2d': {'rgb_view': points_in_image_frame}
}
def train_eval(
load_root_dir,
env_load_fn=None,
gym_env_wrappers=[],
monitor=False,
env_name=None,
agent_class=None,
train_metrics_callback=None,
# SacAgent args
actor_fc_layers=(256, 256),
critic_joint_fc_layers=(256, 256),
# Safety Critic training args
safety_critic_joint_fc_layers=None,
safety_critic_lr=3e-4,
safety_critic_bias_init_val=None,
safety_critic_kernel_scale=None,
n_envs=None,
target_safety=0.2,
fail_weight=None,
# Params for train
num_global_steps=10000,
batch_size=256,
# Params for eval
run_eval=False,
eval_metrics=[],
num_eval_episodes=10,
eval_interval=1000,
# Params for summaries and logging
train_checkpoint_interval=10000,
summary_interval=1000,
monitor_interval=5000,
summaries_flush_secs=10,
debug_summaries=False,
seed=None):
if isinstance(agent_class, str):
assert agent_class in ALGOS, 'trainer.train_eval: agent_class {} invalid'.format(
agent_class)
agent_class = ALGOS.get(agent_class)
train_ckpt_dir = osp.join(load_root_dir, 'train')
rb_ckpt_dir = osp.join(load_root_dir, 'train', 'replay_buffer')
py_env = env_load_fn(env_name, gym_env_wrappers=gym_env_wrappers)
tf_env = tf_py_environment.TFPyEnvironment(py_env)
if monitor:
vid_path = os.path.join(load_root_dir, 'rollouts')
monitor_env_wrapper = misc.monitor_freq(1, vid_path)
monitor_env = gym.make(env_name)
for wrapper in gym_env_wrappers:
monitor_env = wrapper(monitor_env)
monitor_env = monitor_env_wrapper(monitor_env)
# auto_reset must be False to ensure Monitor works correctly
monitor_py_env = gym_wrapper.GymWrapper(monitor_env, auto_reset=False)
if run_eval:
eval_dir = os.path.join(load_root_dir, 'eval')
n_envs = n_envs or num_eval_episodes
eval_summary_writer = tf.compat.v2.summary.create_file_writer(
eval_dir, flush_millis=summaries_flush_secs * 1000)
eval_metrics = [
tf_metrics.AverageReturnMetric(prefix='EvalMetrics',
buffer_size=num_eval_episodes,
batch_size=n_envs),
tf_metrics.AverageEpisodeLengthMetric(
prefix='EvalMetrics',
buffer_size=num_eval_episodes,
batch_size=n_envs)
] + [
tf_py_metric.TFPyMetric(m, name='EvalMetrics/{}'.format(m.name))
for m in eval_metrics
]
eval_tf_env = tf_py_environment.TFPyEnvironment(
parallel_py_environment.ParallelPyEnvironment([
lambda: env_load_fn(env_name,
gym_env_wrappers=gym_env_wrappers)
] * n_envs))
if seed:
seeds = [seed * n_envs + i for i in range(n_envs)]
try:
eval_tf_env.pyenv.seed(seeds)
except:
pass
global_step = tf.compat.v1.train.get_or_create_global_step()
time_step_spec = tf_env.time_step_spec()
observation_spec = time_step_spec.observation
action_spec = tf_env.action_spec()
actor_net = actor_distribution_network.ActorDistributionNetwork(
observation_spec,
action_spec,
fc_layer_params=actor_fc_layers,
continuous_projection_net=agents.normal_projection_net)
critic_net = agents.CriticNetwork(
(observation_spec, action_spec),
joint_fc_layer_params=critic_joint_fc_layers)
if agent_class in SAFETY_AGENTS:
safety_critic_net = agents.CriticNetwork(
(observation_spec, action_spec),
joint_fc_layer_params=critic_joint_fc_layers)
tf_agent = agent_class(time_step_spec,
action_spec,
actor_network=actor_net,
critic_network=critic_net,
safety_critic_network=safety_critic_net,
train_step_counter=global_step,
debug_summaries=False)
else:
tf_agent = agent_class(time_step_spec,
action_spec,
actor_network=actor_net,
critic_network=critic_net,
train_step_counter=global_step,
debug_summaries=False)
collect_data_spec = tf_agent.collect_data_spec
replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
collect_data_spec, batch_size=1, max_length=1000000)
replay_buffer = misc.load_rb_ckpt(rb_ckpt_dir, replay_buffer)
tf_agent, _ = misc.load_agent_ckpt(train_ckpt_dir, tf_agent)
if agent_class in SAFETY_AGENTS:
target_safety = target_safety or tf_agent._target_safety
loaded_train_steps = global_step.numpy()
logging.info("Loaded agent from %s trained for %d steps", train_ckpt_dir,
loaded_train_steps)
global_step.assign(0)
tf.summary.experimental.set_step(global_step)
thresholds = [target_safety, 0.5]
sc_metrics = [
tf.keras.metrics.AUC(name='safety_critic_auc'),
tf.keras.metrics.BinaryAccuracy(name='safety_critic_acc',
threshold=0.5),
tf.keras.metrics.TruePositives(name='safety_critic_tp',
thresholds=thresholds),
tf.keras.metrics.FalsePositives(name='safety_critic_fp',
thresholds=thresholds),
tf.keras.metrics.TrueNegatives(name='safety_critic_tn',
thresholds=thresholds),
tf.keras.metrics.FalseNegatives(name='safety_critic_fn',
thresholds=thresholds)
]
if seed:
tf.compat.v1.set_random_seed(seed)
summaries_flush_secs = 10
timestamp = datetime.utcnow().strftime('%Y-%m-%d-%H-%M-%S')
offline_train_dir = osp.join(train_ckpt_dir, 'offline', timestamp)
config_saver = gin.tf.GinConfigSaverHook(offline_train_dir,
summarize_config=True)
tf.function(config_saver.after_create_session)()
sc_summary_writer = tf.compat.v2.summary.create_file_writer(
offline_train_dir, flush_millis=summaries_flush_secs * 1000)
sc_summary_writer.set_as_default()
if safety_critic_kernel_scale is not None:
ki = tf.compat.v1.variance_scaling_initializer(
scale=safety_critic_kernel_scale,
mode='fan_in',
distribution='truncated_normal')
else:
ki = tf.compat.v1.keras.initializers.VarianceScaling(
scale=1. / 3., mode='fan_in', distribution='uniform')
if safety_critic_bias_init_val is not None:
bi = tf.constant_initializer(safety_critic_bias_init_val)
else:
bi = None
sc_net_off = agents.CriticNetwork(
(observation_spec, action_spec),
joint_fc_layer_params=safety_critic_joint_fc_layers,
kernel_initializer=ki,
value_bias_initializer=bi,
name='SafetyCriticOffline')
sc_net_off.create_variables()
target_sc_net_off = common.maybe_copy_target_network_with_checks(
sc_net_off, None, 'TargetSafetyCriticNetwork')
optimizer = tf.keras.optimizers.Adam(safety_critic_lr)
sc_net_off_ckpt_dir = os.path.join(offline_train_dir, 'safety_critic')
sc_checkpointer = common.Checkpointer(
ckpt_dir=sc_net_off_ckpt_dir,
safety_critic=sc_net_off,
target_safety_critic=target_sc_net_off,
optimizer=optimizer,
global_step=global_step,
max_to_keep=5)
sc_checkpointer.initialize_or_restore()
resample_counter = py_metrics.CounterMetric('ActionResampleCounter')
eval_policy = agents.SafeActorPolicyRSVar(
time_step_spec=time_step_spec,
action_spec=action_spec,
actor_network=actor_net,
safety_critic_network=sc_net_off,
safety_threshold=target_safety,
resample_counter=resample_counter,
training=True)
dataset = replay_buffer.as_dataset(num_parallel_calls=3,
num_steps=2,
sample_batch_size=batch_size //
2).prefetch(3)
data = iter(dataset)
full_data = replay_buffer.gather_all()
fail_mask = tf.cast(full_data.observation['task_agn_rew'], tf.bool)
fail_step = nest_utils.fast_map_structure(
lambda *x: tf.boolean_mask(*x, fail_mask), full_data)
init_step = nest_utils.fast_map_structure(
lambda *x: tf.boolean_mask(*x, full_data.is_first()), full_data)
before_fail_mask = tf.roll(fail_mask, [-1], axis=[1])
after_init_mask = tf.roll(full_data.is_first(), [1], axis=[1])
before_fail_step = nest_utils.fast_map_structure(
lambda *x: tf.boolean_mask(*x, before_fail_mask), full_data)
after_init_step = nest_utils.fast_map_structure(
lambda *x: tf.boolean_mask(*x, after_init_mask), full_data)
filter_mask = tf.squeeze(tf.logical_or(before_fail_mask, fail_mask))
filter_mask = tf.pad(
filter_mask, [[0, replay_buffer._max_length - filter_mask.shape[0]]])
n_failures = tf.reduce_sum(tf.cast(filter_mask, tf.int32)).numpy()
failure_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer(
collect_data_spec,
batch_size=1,
max_length=n_failures,
dataset_window_shift=1)
data_utils.copy_rb(replay_buffer, failure_buffer, filter_mask)
sc_dataset_neg = failure_buffer.as_dataset(num_parallel_calls=3,
sample_batch_size=batch_size //
2,
num_steps=2).prefetch(3)
neg_data = iter(sc_dataset_neg)
get_action = lambda ts: tf_agent._actions_and_log_probs(ts)[0]
eval_sc = log_utils.eval_fn(before_fail_step, fail_step, init_step,
after_init_step, get_action)
losses = []
mean_loss = tf.keras.metrics.Mean(name='mean_ep_loss')
target_update = train_utils.get_target_updater(sc_net_off,
target_sc_net_off)
with tf.summary.record_if(
lambda: tf.math.equal(global_step % summary_interval, 0)):
while global_step.numpy() < num_global_steps:
pos_experience, _ = next(data)
neg_experience, _ = next(neg_data)
exp = data_utils.concat_batches(pos_experience, neg_experience,
collect_data_spec)
boundary_mask = tf.logical_not(exp.is_boundary()[:, 0])
exp = nest_utils.fast_map_structure(
lambda *x: tf.boolean_mask(*x, boundary_mask), exp)
safe_rew = exp.observation['task_agn_rew'][:, 1]
if fail_weight:
weights = tf.where(tf.cast(safe_rew, tf.bool),
fail_weight / 0.5, (1 - fail_weight) / 0.5)
else:
weights = None
train_loss, sc_loss, lam_loss = train_step(
exp,
safe_rew,
tf_agent,
sc_net=sc_net_off,
target_sc_net=target_sc_net_off,
metrics=sc_metrics,
weights=weights,
target_safety=target_safety,
optimizer=optimizer,
target_update=target_update,
debug_summaries=debug_summaries)
global_step.assign_add(1)
global_step_val = global_step.numpy()
losses.append(
(train_loss.numpy(), sc_loss.numpy(), lam_loss.numpy()))
mean_loss(train_loss)
with tf.name_scope('Losses'):
tf.compat.v2.summary.scalar(name='sc_loss',
data=sc_loss,
step=global_step_val)
tf.compat.v2.summary.scalar(name='lam_loss',
data=lam_loss,
step=global_step_val)
if global_step_val % summary_interval == 0:
tf.compat.v2.summary.scalar(name=mean_loss.name,
data=mean_loss.result(),
step=global_step_val)
if global_step_val % summary_interval == 0:
with tf.name_scope('Metrics'):
for metric in sc_metrics:
if len(tf.squeeze(metric.result()).shape) == 0:
tf.compat.v2.summary.scalar(name=metric.name,
data=metric.result(),
step=global_step_val)
else:
fmt_str = '_{}'.format(thresholds[0])
tf.compat.v2.summary.scalar(
name=metric.name + fmt_str,
data=metric.result()[0],
step=global_step_val)
fmt_str = '_{}'.format(thresholds[1])
tf.compat.v2.summary.scalar(
name=metric.name + fmt_str,
data=metric.result()[1],
step=global_step_val)
metric.reset_states()
if global_step_val % eval_interval == 0:
eval_sc(sc_net_off, step=global_step_val)
if run_eval:
results = metric_utils.eager_compute(
eval_metrics,
eval_tf_env,
eval_policy,
num_episodes=num_eval_episodes,
train_step=global_step,
summary_writer=eval_summary_writer,
summary_prefix='EvalMetrics',
)
if train_metrics_callback is not None:
train_metrics_callback(results, global_step_val)
metric_utils.log_metrics(eval_metrics)
with eval_summary_writer.as_default():
for eval_metric in eval_metrics[2:]:
eval_metric.tf_summaries(
train_step=global_step,
step_metrics=eval_metrics[:2])
if monitor and global_step_val % monitor_interval == 0:
monitor_time_step = monitor_py_env.reset()
monitor_policy_state = eval_policy.get_initial_state(1)
ep_len = 0
monitor_start = time.time()
while not monitor_time_step.is_last():
monitor_action = eval_policy.action(
monitor_time_step, monitor_policy_state)
action, monitor_policy_state = monitor_action.action, monitor_action.state
monitor_time_step = monitor_py_env.step(action)
ep_len += 1
logging.debug(
'saved rollout at timestep %d, rollout length: %d, %4.2f sec',
global_step_val, ep_len,
time.time() - monitor_start)
if global_step_val % train_checkpoint_interval == 0:
sc_checkpointer.save(global_step=global_step_val)
def preprocess(inputs,
output_keys=None,
is_training=False,
using_sequence_dataset=False,
num_frame_to_load=1,
transform_points_fn=None,
image_preprocess_fn_dic=None,
images_points_correspondence_fn=None,
compute_semantic_labels_fn=None,
compute_motion_labels_fn=None,
view_names=(),
points_key='points',
colors_key='colors',
normals_key='normals',
intensities_key='intensities',
elongations_key='elongations',
semantic_labels_key='semantic_labels',
motion_labels_key='motion_labels',
spin_coords_key=None,
points_in_image_frame_key=None,
num_points_to_randomly_sample=None,
x_min_degree_rotation=None,
x_max_degree_rotation=None,
y_min_degree_rotation=None,
y_max_degree_rotation=None,
z_min_degree_rotation=None,
z_max_degree_rotation=None,
points_pad_or_clip_size=None,
voxels_pad_or_clip_size=None,
voxel_grid_cell_size=(0.1, 0.1, 0.1),
num_offset_bins_x=4,
num_offset_bins_y=4,
num_offset_bins_z=4,
point_feature_keys=('point_offsets', ),
point_to_voxel_segment_func=tf.math.unsorted_segment_mean,
x_random_crop_size=None,
y_random_crop_size=None,
min_scale_ratio=None,
max_scale_ratio=None,
semantic_labels_offset=0,
ignore_labels=(),
remove_unlabeled_images_and_points=False,
labeled_view_name=None,
only_keep_first_return_lidar_points=False):
"""Preprocesses a dictionary of `Tensor` inputs.
If is_training=True, it will randomly rotate the points around the z axis,
and will randomly flip the points with respect to x and/or y axis.
Note that the preprocessor function does not correct normal vectors if they
exist in the inputs.
Note that the preprocessing effects all values of `inputs` that are `Tensors`.
Args:
inputs: A dictionary of inputs. Each value must be a `Tensor`.
output_keys: Either None, or a list of strings containing the keys in the
dictionary that is returned by the preprocess function.
is_training: Whether we're training or testing.
using_sequence_dataset: if true, the inputs will contain scene and multiple
frames data.
num_frame_to_load: If greater than 1, load multiframe point cloud point
positions and its correspondence.
transform_points_fn: Fn to transform other frames to a specific frame's
coordinate.
image_preprocess_fn_dic: Image preprocessing function. Maps view names to
their image preprocessing functions. Set it to None, if there are no
images to preprocess or you are not interested in preprocesing images.
images_points_correspondence_fn: The function that computes correspondence
between images and points.
compute_semantic_labels_fn: If not None, semantic labels will be computed
using this function.
compute_motion_labels_fn: If not None, motion labels will be computed using
this function.
view_names: Names corresponding to 2d views of the scene.
points_key: The key used for `points` in the inputs.
colors_key: The key used for `colors` in the inputs.
normals_key: The key used for 'normals' in the inputs.
intensities_key: The key used for 'intensities' in the inputs.
elongations_key: The key used for 'elongations' in the inputs.
semantic_labels_key: The key used for 'semantic_labels' in the inputs.
motion_labels_key: The key used for 'motion_labels' in the inputs.
spin_coords_key: The key used for 'spin_coords' in the inputs. In Waymo
data, spin_coords is a [num_points, 3] tensor that contains scan_index,
shot_index, return_index. In Waymo data, return_index of the first return
points is 0.
points_in_image_frame_key: A string that identifies the tensor that contains
the points_in_image_frame tensor. If None, it won't be used.
num_points_to_randomly_sample: Number of points to randomly sample. If None,
it will keep the original points and does not perform sampling.
x_min_degree_rotation: Min degree of rotation around the x axis.
x_max_degree_rotation: Max degree of ratation around the x axis.
y_min_degree_rotation: Min degree of rotation around the y axis.
y_max_degree_rotation: Max degree of ratation around the y axis.
z_min_degree_rotation: Min degree of rotation around the z axis.
z_max_degree_rotation: Max degree of ratation around the z axis.
points_pad_or_clip_size: Number of target points to pad or clip to. If None,
it will not perform the point padding.
voxels_pad_or_clip_size: Number of target voxels to pad or clip to. If None,
it will not perform the voxel padding.
voxel_grid_cell_size: A three dimensional tuple determining the voxel grid
size.
num_offset_bins_x: Number of bins for point offsets in x direction.
num_offset_bins_y: Number of bins for point offsets in y direction.
num_offset_bins_z: Number of bins for point offsets in z direction.
point_feature_keys: The keys used to form the voxel features.
point_to_voxel_segment_func: The function used to aggregate the features
of the points that fall in the same voxel.
x_random_crop_size: Size of the random crop in x dimension. If None, random
crop will not take place on x dimension.
y_random_crop_size: Size of the random crop in y dimension. If None, random
crop will not take place on y dimension.
min_scale_ratio: Minimum scale ratio. Used for scaling point cloud.
max_scale_ratio: Maximum scale ratio. Used for scaling point cloud.
semantic_labels_offset: An integer offset that will be added to labels.
ignore_labels: A tuple containing labels that should be ignored when
computing the loss and metrics.
remove_unlabeled_images_and_points: If True, removes the images that are not
labeled and also removes the points that are associated with those images.
labeled_view_name: The name of the view that is labeled, otherwise None.
only_keep_first_return_lidar_points: If True, we only keep the first return
lidar points.
Returns:
The mean subtracted points with an optional rotation applied.
Raises:
ValueError: if `inputs` doesn't contain the points_key.
ValueError: if `points_in_image_frame` does not have rank 3.
"""
inputs = dict(inputs)
if using_sequence_dataset:
all_frame_inputs = inputs
scene = all_frame_inputs['scene']
frame1 = all_frame_inputs['frame1']
frame_start_index = all_frame_inputs['frame_start_index']
inputs = dict(
all_frame_inputs['frame0']
) # so that the following processing code can be unchanged.
# Initializing empty dictionary for mesh, image, indices_2d and non tensor
# inputs.
non_tensor_inputs = {}
view_image_inputs = {}
view_indices_2d_inputs = {}
mesh_inputs = {}
if image_preprocess_fn_dic is None:
image_preprocess_fn_dic = {}
# Convert all float64 to float32 and all int64 to int32.
for key in sorted(inputs):
if isinstance(inputs[key], tf.Tensor):
if inputs[key].dtype == tf.float64:
inputs[key] = tf.cast(inputs[key], dtype=tf.float32)
if inputs[key].dtype == tf.int64:
inputs[key] = tf.cast(inputs[key], dtype=tf.int32)
if points_key in inputs:
inputs[standard_fields.InputDataFields.
point_positions] = inputs[points_key]
if colors_key is not None and colors_key in inputs:
inputs[
standard_fields.InputDataFields.point_colors] = inputs[colors_key]
if normals_key is not None and normals_key in inputs:
inputs[standard_fields.InputDataFields.
point_normals] = inputs[normals_key]
if intensities_key is not None and intensities_key in inputs:
inputs[standard_fields.InputDataFields.
point_intensities] = inputs[intensities_key]
if elongations_key is not None and elongations_key in inputs:
inputs[standard_fields.InputDataFields.
point_elongations] = inputs[elongations_key]
if semantic_labels_key is not None and semantic_labels_key in inputs:
inputs[standard_fields.InputDataFields.
object_class_points] = inputs[semantic_labels_key]
if motion_labels_key is not None and motion_labels_key in inputs:
inputs[standard_fields.InputDataFields.
object_flow_points] = inputs[motion_labels_key]
if spin_coords_key is not None and spin_coords_key in inputs:
inputs[standard_fields.InputDataFields.
point_spin_coordinates] = inputs[spin_coords_key]
# Acquire point / image correspondences.
if images_points_correspondence_fn is not None:
fn_outputs = images_points_correspondence_fn(inputs)
if 'points_position' in fn_outputs:
inputs[standard_fields.InputDataFields.
point_positions] = fn_outputs['points_position']
if 'points_intensity' in fn_outputs and intensities_key is not None:
inputs[standard_fields.InputDataFields.
point_intensities] = fn_outputs['points_intensity']
if 'points_elongation' in fn_outputs and elongations_key is not None:
inputs[standard_fields.InputDataFields.
point_elongations] = fn_outputs['points_elongation']
if 'points_label' in fn_outputs and semantic_labels_key is not None:
inputs[standard_fields.InputDataFields.
object_class_points] = fn_outputs['points_label']
if 'view_images' in fn_outputs:
for key in sorted(fn_outputs['view_images']):
if len(fn_outputs['view_images'][key].shape) != 4:
raise ValueError(('%s image should have rank 4.' % key))
view_image_inputs = fn_outputs['view_images']
if 'view_indices_2d' in fn_outputs:
for key in sorted(fn_outputs['view_indices_2d']):
if len(fn_outputs['view_indices_2d'][key].shape) != 3:
raise ValueError(
('%s indices_2d should have rank 3.' % key))
view_indices_2d_inputs = fn_outputs['view_indices_2d']
else:
if points_in_image_frame_key is not None:
inputs['rgb_view/features'] = inputs['image']
inputs['rgb_view/indices_2d'] = inputs[points_in_image_frame_key]
if len(inputs['rgb_view/indices_2d'].shape) != 3:
raise ValueError('`points_in_image_frame` should have rank 3.')
frame0 = inputs.copy()
if num_frame_to_load > 1:
point_positions_list = [
frame0[standard_fields.InputDataFields.point_positions]
]
if view_indices_2d_inputs:
view_indices_2d_list = [view_indices_2d_inputs[view_names[0]]]
frame_source_list = [
tf.zeros([
tf.shape(
frame0[standard_fields.InputDataFields.point_positions])[0]
], tf.int32)
]
for i in range(1, num_frame_to_load):
target_frame_key = 'frame' + str(i)
if images_points_correspondence_fn is not None:
frame_i = images_points_correspondence_fn(
all_frame_inputs[target_frame_key])
else:
raise ValueError(
'images_points_correspondence_fn is needed for loading multi-frame pointclouds.'
)
transformed_point_positions = transform_points_fn(
scene, frame_i['points_position'], frame_start_index,
i + frame_start_index)
point_positions_list.append(transformed_point_positions)
if view_indices_2d_inputs:
view_indices_2d_list.append(
frame_i['view_indices_2d'][view_names[0]])
frame_source_list.append(
tf.ones([tf.shape(transformed_point_positions)[0]], tf.int32) *
i)
# add multi-frame info to override inputs and view_indices_2d_inputs
inputs[standard_fields.InputDataFields.
point_frame_index] = tf.expand_dims(tf.concat(frame_source_list,
axis=0),
axis=1)
inputs[standard_fields.InputDataFields.point_positions] = tf.concat(
point_positions_list, axis=0)
if view_indices_2d_inputs:
view_indices_2d_inputs[view_names[0]] = tf.concat(
view_indices_2d_list, axis=1)
# Validate inputs.
if standard_fields.InputDataFields.point_positions not in inputs:
raise ValueError('`inputs` must contain a point_positions')
if inputs[
standard_fields.InputDataFields.point_positions].shape.ndims != 2:
raise ValueError('points must be of rank 2.')
if inputs[standard_fields.InputDataFields.point_positions].shape[1] != 3:
raise ValueError('point should be 3 dimensional.')
# Remove normal nans.
if standard_fields.InputDataFields.point_normals in inputs:
inputs[standard_fields.InputDataFields.point_normals] = tf.where(
tf.math.is_nan(
inputs[standard_fields.InputDataFields.point_normals]),
tf.zeros_like(
inputs[standard_fields.InputDataFields.point_normals]),
inputs[standard_fields.InputDataFields.point_normals])
# Compute semantic labels if compute_semantic_labels_fn is not None
# An example is when the ground-truth contains 3d object boxes and not per
# point labels. This would be a function that infers point labels from boxes.
if compute_semantic_labels_fn is not None:
inputs[standard_fields.InputDataFields.
object_class_points] = compute_semantic_labels_fn(
inputs=frame0,
points_key=standard_fields.InputDataFields.point_positions)
if compute_motion_labels_fn is not None:
inputs[standard_fields.InputDataFields.
object_flow_points] = compute_motion_labels_fn(
scene=scene,
frame0=frame0,
frame1=frame1,
frame_start_index=frame_start_index,
points_key=standard_fields.InputDataFields.point_positions)
# Splitting inputs to {view_image_inputs,
# view_indices_2d_inputs,
# mesh_inputs,
# non_tensor_inputs}
mesh_keys = []
for key in [
standard_fields.InputDataFields.point_positions,
standard_fields.InputDataFields.point_colors,
standard_fields.InputDataFields.point_normals,
standard_fields.InputDataFields.point_intensities,
standard_fields.InputDataFields.point_elongations,
standard_fields.InputDataFields.object_class_points,
standard_fields.InputDataFields.point_spin_coordinates,
standard_fields.InputDataFields.object_flow_points,
standard_fields.InputDataFields.point_frame_index,
]:
if key is not None and key in inputs:
mesh_keys.append(key)
view_image_names = [('%s/features' % key) for key in view_names]
view_indices_2d_names = [('%s/indices_2d' % key) for key in view_names]
# Additional key collecting
for k, v in six.iteritems(inputs):
if k in view_image_names:
view_image_inputs[k] = v
elif k in view_indices_2d_names:
view_indices_2d_inputs[k] = v
elif k in mesh_keys:
if num_frame_to_load > 1:
pad_size = tf.shape(
inputs[standard_fields.InputDataFields.
point_positions])[0] - tf.shape(v)[0]
if k == standard_fields.InputDataFields.object_class_points:
pad_value = -1
else:
pad_value = 0
v = tf.pad(v, [[0, pad_size], [0, 0]],
constant_values=pad_value)
mesh_inputs[k] = v
else:
non_tensor_inputs[k] = v
# Remove points that are not in the lidar first return (optional)
if only_keep_first_return_lidar_points:
_remove_second_return_lidar_points(
mesh_inputs=mesh_inputs,
view_indices_2d_inputs=view_indices_2d_inputs)
# Randomly sample points
preprocessor_utils.randomly_sample_points(
mesh_inputs=mesh_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
target_num_points=num_points_to_randomly_sample)
# Add weights if it does not exist in inputs. The weight of the points with
# label in `ignore_labels` is set to 0. This helps the loss and metrics to
# ignore those labels.
use_weights = (
standard_fields.InputDataFields.object_class_points in mesh_inputs
or standard_fields.InputDataFields.object_flow_points in mesh_inputs)
if use_weights:
if num_frame_to_load > 1:
num_valid_points_frame0 = tf.shape(
frame0[standard_fields.InputDataFields.point_positions])[0]
num_additional_frame_points = tf.shape(
mesh_inputs[standard_fields.InputDataFields.
object_class_points])[0] - num_valid_points_frame0
weights = tf.concat([
tf.ones([num_valid_points_frame0, 1], tf.float32),
tf.zeros([num_additional_frame_points, 1], tf.float32)
],
axis=0)
else:
weights = tf.ones_like(mesh_inputs[
standard_fields.InputDataFields.object_class_points],
dtype=tf.float32)
if standard_fields.InputDataFields.object_class_points in mesh_inputs:
mesh_inputs[
standard_fields.InputDataFields.object_class_points] = tf.cast(
mesh_inputs[
standard_fields.InputDataFields.object_class_points],
dtype=tf.int32)
for ignore_label in ignore_labels:
weights *= tf.cast(tf.not_equal(
mesh_inputs[
standard_fields.InputDataFields.object_class_points],
ignore_label),
dtype=tf.float32)
mesh_inputs[
standard_fields.InputDataFields.point_loss_weights] = weights
mesh_inputs[standard_fields.InputDataFields.
object_class_points] += semantic_labels_offset
# We normalize the intensities and elongations to be in a smaller range.
if standard_fields.InputDataFields.point_intensities in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.
point_intensities] = change_intensity_range(
intensities=mesh_inputs[
standard_fields.InputDataFields.point_intensities])
if standard_fields.InputDataFields.point_elongations in mesh_inputs:
mesh_inputs[
standard_fields.InputDataFields.point_elongations] = (tf.cast(
mesh_inputs[standard_fields.InputDataFields.point_elongations],
dtype=tf.float32) * 2.0 / 255.0) - 1.0
# Random scale the points.
if min_scale_ratio is not None and max_scale_ratio is not None:
scale_ratio = tf.random.uniform([],
minval=min_scale_ratio,
maxval=max_scale_ratio,
dtype=tf.float32)
mesh_inputs[
standard_fields.InputDataFields.point_positions] *= scale_ratio
if standard_fields.InputDataFields.object_flow_points in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.
object_flow_points] *= scale_ratio
# Random crop the points.
randomly_crop_points(mesh_inputs=mesh_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
x_random_crop_size=x_random_crop_size,
y_random_crop_size=y_random_crop_size)
# If training, pick the best labeled image and points that project to it.
# In many datasets, only one image is labeled anyways.
if remove_unlabeled_images_and_points:
pick_labeled_image(mesh_inputs=mesh_inputs,
view_image_inputs=view_image_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
view_name=labeled_view_name)
# Process images.
preprocessor_utils.preprocess_images(
view_image_inputs=view_image_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
image_preprocess_fn_dic=image_preprocess_fn_dic,
is_training=is_training)
# Record the original points.
original_points = mesh_inputs[
standard_fields.InputDataFields.point_positions]
if standard_fields.InputDataFields.point_colors in mesh_inputs:
original_colors = mesh_inputs[
standard_fields.InputDataFields.point_colors]
if standard_fields.InputDataFields.point_normals in mesh_inputs:
original_normals = mesh_inputs[
standard_fields.InputDataFields.point_normals]
# Update feature visibility count.
if 'feature_visibility_count' in mesh_inputs:
mesh_inputs['feature_visibility_count'] = tf.maximum(
mesh_inputs['feature_visibility_count'], 1)
mesh_inputs['features'] /= tf.cast(
mesh_inputs['feature_visibility_count'], dtype=tf.float32)
# Subtract mean from points.
mean_points = tf.reduce_mean(
mesh_inputs[standard_fields.InputDataFields.point_positions], axis=0)
mesh_inputs[
standard_fields.InputDataFields.point_positions] -= tf.expand_dims(
mean_points, axis=0)
# Rotate points randomly.
if standard_fields.InputDataFields.point_normals in mesh_inputs:
normals = mesh_inputs[standard_fields.InputDataFields.point_normals]
else:
normals = None
if standard_fields.InputDataFields.object_flow_points in mesh_inputs:
motions = mesh_inputs[
standard_fields.InputDataFields.object_flow_points]
else:
motions = None
(mesh_inputs[standard_fields.InputDataFields.point_positions],
rotated_normals, rotated_motions) = rotate_randomly(
points=mesh_inputs[standard_fields.InputDataFields.point_positions],
normals=normals,
motions=motions,
x_min_degree_rotation=x_min_degree_rotation,
x_max_degree_rotation=x_max_degree_rotation,
y_min_degree_rotation=y_min_degree_rotation,
y_max_degree_rotation=y_max_degree_rotation,
z_min_degree_rotation=z_min_degree_rotation,
z_max_degree_rotation=z_max_degree_rotation)
# Random flipping in x and y directions.
(mesh_inputs[standard_fields.InputDataFields.point_positions],
flipped_normals,
flipped_motions) = flip_randomly_points_and_normals_motions(
points=mesh_inputs[standard_fields.InputDataFields.point_positions],
normals=rotated_normals,
motions=rotated_motions,
is_training=is_training)
if standard_fields.InputDataFields.point_normals in mesh_inputs:
mesh_inputs[
standard_fields.InputDataFields.point_normals] = flipped_normals
if standard_fields.InputDataFields.object_flow_points in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.
object_flow_points] = flipped_motions
# Normalize RGB to [-1.0, 1.0].
if standard_fields.InputDataFields.point_colors in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.point_colors] = tf.cast(
mesh_inputs[standard_fields.InputDataFields.point_colors],
dtype=tf.float32)
mesh_inputs[standard_fields.InputDataFields.point_colors] *= (2.0 /
255.0)
mesh_inputs[standard_fields.InputDataFields.point_colors] -= 1.0
# Add original points to mesh inputs.
mesh_inputs[standard_fields.InputDataFields.
point_positions_original] = original_points
if standard_fields.InputDataFields.point_colors in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.
point_colors_original] = original_colors
if standard_fields.InputDataFields.point_normals in mesh_inputs:
mesh_inputs[standard_fields.InputDataFields.
point_normals_original] = original_normals
# Pad or clip the point tensors.
pad_or_clip(mesh_inputs=mesh_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
pad_or_clip_size=points_pad_or_clip_size)
if num_frame_to_load > 1:
# Note: num_valid_points is the sum of 'num_points_per_fram' for now.
# num_points_per_frame is each frame's valid num of points.
# TODO(huangrui): if random sampling is called earlier, the count here
# is not guaranteed to be in order. need sorting.
if num_points_to_randomly_sample is not None:
raise ValueError(
'randomly sample is not compatible with padding multi frame point clouds yet!'
)
_, _, mesh_inputs[standard_fields.InputDataFields.
num_valid_points_per_frame] = tf.unique_with_counts(
tf.reshape(
mesh_inputs[standard_fields.InputDataFields.
point_frame_index], [-1]))
if points_pad_or_clip_size is not None:
padded_points = tf.where_v2(
tf.greater(
points_pad_or_clip_size, mesh_inputs[
standard_fields.InputDataFields.num_valid_points]),
points_pad_or_clip_size -
mesh_inputs[standard_fields.InputDataFields.num_valid_points],
0)
# Correct the potential unique count error from optionally padded 0s point
# frame index.
mesh_inputs[
standard_fields.InputDataFields.
num_valid_points_per_frame] -= tf.pad(
tf.expand_dims(padded_points, 0), [[
0,
tf.shape(mesh_inputs[standard_fields.InputDataFields.
num_valid_points_per_frame])[0] -
1
]])
# Putting back the dictionaries together
processed_inputs = mesh_inputs.copy()
processed_inputs.update(non_tensor_inputs)
for key in sorted(view_image_inputs):
processed_inputs[('%s/features' % key)] = view_image_inputs[key]
for key in sorted(view_indices_2d_inputs):
processed_inputs[('%s/indices_2d' % key)] = view_indices_2d_inputs[key]
# Create features that do not exist
if 'point_offsets' in point_feature_keys:
preprocessor_utils.add_point_offsets(
inputs=processed_inputs, voxel_grid_cell_size=voxel_grid_cell_size)
if 'point_offset_bins' in point_feature_keys:
preprocessor_utils.add_point_offset_bins(
inputs=processed_inputs,
voxel_grid_cell_size=voxel_grid_cell_size,
num_bins_x=num_offset_bins_x,
num_bins_y=num_offset_bins_y,
num_bins_z=num_offset_bins_z)
# Voxelize point features
preprocessor_utils.voxelize_point_features(
inputs=processed_inputs,
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size,
point_feature_keys=point_feature_keys,
point_to_voxel_segment_func=point_to_voxel_segment_func,
num_frame_to_load=num_frame_to_load)
# Voxelize point / image correspondence indices
preprocessor_utils.voxelize_point_to_view_correspondences(
inputs=processed_inputs,
view_indices_2d_inputs=view_indices_2d_inputs,
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size)
# Voxelizing the semantic labels
preprocessor_utils.voxelize_semantic_labels(
inputs=processed_inputs,
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size)
# Voxelizing the loss weights
preprocessor_utils.voxelize_property_tensor(
inputs=processed_inputs,
point_tensor_key=standard_fields.InputDataFields.point_loss_weights,
corresponding_voxel_tensor_key=standard_fields.InputDataFields.
voxel_loss_weights,
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size,
segment_func=tf.math.unsorted_segment_max)
# Voxelizing the object flow
if standard_fields.InputDataFields.object_flow_points in processed_inputs:
preprocessor_utils.voxelize_property_tensor(
inputs=processed_inputs,
point_tensor_key=standard_fields.InputDataFields.
object_flow_points,
corresponding_voxel_tensor_key='object_flow_voxels_max',
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size,
segment_func=tf.math.unsorted_segment_max)
preprocessor_utils.voxelize_property_tensor(
inputs=processed_inputs,
point_tensor_key=standard_fields.InputDataFields.
object_flow_points,
corresponding_voxel_tensor_key='object_flow_voxels_min',
voxels_pad_or_clip_size=voxels_pad_or_clip_size,
voxel_grid_cell_size=voxel_grid_cell_size,
segment_func=tf.math.unsorted_segment_min)
processed_inputs[standard_fields.InputDataFields.
object_flow_voxels] = processed_inputs[
'object_flow_voxels_max'] + processed_inputs[
'object_flow_voxels_min']
if num_frame_to_load > 1:
mesh_inputs[
standard_fields.InputDataFields.num_valid_points] = mesh_inputs[
standard_fields.InputDataFields.num_valid_points_per_frame][0]
# Filter preprocessed_inputs by output_keys if it is not None.
if output_keys is not None:
processed_inputs = {
k: v
for k, v in six.iteritems(processed_inputs) if k in output_keys
}
return processed_inputs
def compute_module_criticality(
objective_fn,
module_variables_init,
module_variables_final,
num_samples_per_iteration=10,
alpha_grid_size=10,
sigma_grid_size=10,
sigma_ratio=1.0,
loss_threshold_condition=relative_error_condition,
normalize_error=False,
):
"""Compute the criticality of a module parameterized by `module_variables`.
Args:
objective_fn: A callable that takes in an iterable of the module-specific
variables and produces the value of the objective function.
module_variables_init: A list of tf.Tensors; the variables of the module at
initialization.
module_variables_final: A list of tf.Tensors; the variables of the module at
convergence.
num_samples_per_iteration: Number of perturbations to sample each iteration.
alpha_grid_size: The number of values to test for alpha, the interpolation
coefficient.
sigma_grid_size: The number of values to test for sigma, the standard
deviation of the perturbation.
sigma_ratio: Positive scalar multiplier k for values of sigma, to enforce
that the tested values of sigma lie in [k * 1e-16, k]; the default is 1.0,
implying that the tested values of sigma lie in the interval [1e-16, 1].
loss_threshold_condition: A callable that takes in a reference objective
value and a candidate objective value and produces a thresholding
decision.
normalize_error: Whether to normalize the error that is minimized over in
the definition of criticality by the Frobenius norm of the distance
between initial and final parameters.
Returns:
A `collections.NamedTuple` that contains the results of the criticality
analysis.
"""
initial_objective_value = objective_fn(module_variables_init)
final_objective_value = objective_fn(module_variables_final)
# Test a 2D grid of alpha and sigma values.
float_zero = tf.cast(0, tf.float32)
alphas, sigmas = tf.meshgrid(
tf.linspace(float_zero, 1, alpha_grid_size + 1),
tf.linspace(float_zero + 1e-16, 1, sigma_grid_size + 1) * sigma_ratio,
)
alphas, sigmas = tf.reshape(alphas, [-1]), tf.reshape(sigmas, [-1])
def _evaluate_alpha_sigma(alpha_sigma):
alpha, sigma = alpha_sigma
return _interpolate_and_perturb(
alpha=alpha,
sigma=sigma,
params_init=module_variables_init,
params_final=module_variables_final,
objective_fn=objective_fn,
loss_threshold_condition=functools.partial(
loss_threshold_condition,
reference_error=final_objective_value),
normalize_error=normalize_error,
num_samples_per_iteration=num_samples_per_iteration,
)
(threshold_conditions, interpolated_and_perturbed_losses,
interpolated_and_perturbed_norms) = tf.map_fn(
_evaluate_alpha_sigma,
elems=(alphas, sigmas),
dtype=(tf.bool, tf.float32, tf.float32),
)
masked_interpolated_and_perturbed_norms = tf.where(
threshold_conditions, interpolated_and_perturbed_norms,
tf.ones_like(interpolated_and_perturbed_norms) * np.inf)
idx_min = tf.math.argmin(masked_interpolated_and_perturbed_norms)
(loss_final, norm_final, alpha_final,
sigma_final) = (interpolated_and_perturbed_losses[idx_min],
interpolated_and_perturbed_norms[idx_min],
alphas[idx_min], sigmas[idx_min])
return ModuleCriticalityAnalysis(
criticality_score=norm_final,
alpha=alpha_final,
sigma=sigma_final,
loss_value=loss_final,
num_samples_per_iteration=num_samples_per_iteration,
alpha_grid_size=alpha_grid_size,
sigma_grid_size=sigma_grid_size,
sigma_ratio=sigma_ratio,
initial_objective_value=initial_objective_value,
final_objective_value=final_objective_value,
)
def _safe_div(a, b): """Divides two numbers, returns 0 if denominator is (close to) 0.""" return tf.where(tf.less(tf.abs(b), 1e-10), 0.0, a / b)
def prepare_scannet_scene_dataset(inputs, valid_object_classes=None):
"""Maps the fields from loaded input to standard fields.
Args:
inputs: A dictionary of input tensors.
valid_object_classes: List of valid object classes. if None, it is ignored.
Returns:
A dictionary of input tensors with standard field names.
"""
prepared_inputs = {}
if 'mesh/vertices/positions' in inputs:
prepared_inputs[standard_fields.InputDataFields
.point_positions] = inputs['mesh/vertices/positions']
if 'mesh/vertices/normals' in inputs:
prepared_inputs[standard_fields.InputDataFields
.point_normals] = inputs['mesh/vertices/normals']
prepared_inputs[standard_fields.InputDataFields.point_normals] = tf.where(
tf.math.is_nan(
prepared_inputs[standard_fields.InputDataFields.point_normals]),
tf.zeros_like(
prepared_inputs[standard_fields.InputDataFields.point_normals]),
prepared_inputs[standard_fields.InputDataFields.point_normals])
if 'mesh/vertices/colors' in inputs:
prepared_inputs[standard_fields.InputDataFields
.point_colors] = inputs['mesh/vertices/colors'][:, 0:3]
prepared_inputs[standard_fields.InputDataFields.point_colors] = tf.cast(
prepared_inputs[standard_fields.InputDataFields.point_colors],
dtype=tf.float32)
prepared_inputs[standard_fields.InputDataFields.point_colors] *= (2.0 /
255.0)
prepared_inputs[standard_fields.InputDataFields.point_colors] -= 1.0
if 'scene_name' in inputs:
prepared_inputs[standard_fields.InputDataFields
.camera_image_name] = inputs['scene_name']
if 'mesh/vertices/semantic_labels' in inputs:
prepared_inputs[
standard_fields.InputDataFields
.object_class_points] = inputs['mesh/vertices/semantic_labels']
if 'mesh/vertices/instance_labels' in inputs:
prepared_inputs[
standard_fields.InputDataFields.object_instance_id_points] = tf.reshape(
inputs['mesh/vertices/instance_labels'], [-1])
if valid_object_classes is not None:
valid_objects_mask = tf.cast(
tf.zeros_like(
prepared_inputs[
standard_fields.InputDataFields.object_class_points],
dtype=tf.int32),
dtype=tf.bool)
for object_class in valid_object_classes:
valid_objects_mask = tf.logical_or(
valid_objects_mask,
tf.equal(
prepared_inputs[
standard_fields.InputDataFields.object_class_points],
object_class))
valid_objects_mask = tf.cast(
valid_objects_mask,
dtype=prepared_inputs[
standard_fields.InputDataFields.object_class_points].dtype)
prepared_inputs[standard_fields.InputDataFields
.object_class_points] *= valid_objects_mask
return prepared_inputs
def precision(y, y_hat, name='precision'):
tp = true_p(y, y_hat)
fp = false_p(y, y_hat)
return tf.where(
tf.greater(tp + fp, 0), tf.div(tp, tp + fp), 0, name)
def recall(y, y_hat, name='recall'):
tp = true_p(y, y_hat)
fn = false_p(y, y_hat)
return tf.where(
tf.greater(tp + fn, 0),
tf.div(tp, tp + fn), 0, name)
