def fn():
"""Loss function for when number of input and output boxes is positive."""
if is_balanced:
weights = loss_utils.get_balanced_loss_weights_multiclass(
labels=input_boxes_instance_id)
else:
weights = tf.ones([tf.shape(input_boxes_instance_id)[0], 1],
dtype=tf.float32)
gt_length = tf.reshape(input_boxes_length, [-1, 1])
gt_height = tf.reshape(input_boxes_height, [-1, 1])
gt_width = tf.reshape(input_boxes_width, [-1, 1])
predicted_length = tf.reshape(output_boxes_length, [-1, 1])
predicted_height = tf.reshape(output_boxes_height, [-1, 1])
predicted_width = tf.reshape(output_boxes_width, [-1, 1])
predicted_length /= gt_length
predicted_height /= gt_height
predicted_width /= gt_width
predicted_size = tf.concat(
[predicted_length, predicted_height, predicted_width], axis=1)
gt_size = tf.ones_like(predicted_size)
if loss_type == 'huber':
loss_fn = tf.keras.losses.Huber(
delta=delta, reduction=tf.keras.losses.Reduction.NONE)
elif loss_type == 'absolute_difference':
loss_fn = tf.keras.losses.MeanAbsoluteError(
reduction=tf.keras.losses.Reduction.NONE)
else:
raise ValueError(('Unknown loss type %s.' % loss_type))
size_losses = loss_fn(y_true=gt_size, y_pred=predicted_size)
return tf.reduce_mean(size_losses * tf.reshape(weights, [-1]))
def compute_pointcloud_weights_based_on_voxel_density(points, grid_cell_size):
"""Computes pointcloud weights based on voxel density.
Args:
points: A tf.float32 tensor of size [num_points, 3].
grid_cell_size: The size of the grid cells in x, y, z dimensions in the
voxel grid. It should be either a tf.float32 tensor, a numpy array or a
list of size [3].
Returns:
A tf.float32 tensor of size [num_points, 1] containing weights that are
inverse proportional to the denisty of the points in voxels.
"""
num_points = tf.shape(points)[0]
features = tf.ones([num_points, 1], dtype=tf.float32)
voxel_features, _, segment_ids, _ = (
pointcloud_to_sparse_voxel_grid_unbatched(
points=points,
features=features,
grid_cell_size=grid_cell_size,
segment_func=tf.math.unsorted_segment_sum))
num_voxels = tf.shape(voxel_features)[0]
point_features = sparse_voxel_grid_to_pointcloud(
voxel_features=tf.expand_dims(voxel_features, axis=0),
segment_ids=tf.expand_dims(segment_ids, axis=0),
num_valid_voxels=tf.expand_dims(num_voxels, axis=0),
num_valid_points=tf.expand_dims(num_points, axis=0))
inverse_point_densities = 1.0 / tf.squeeze(point_features, axis=0)
total_inverse_density = tf.reduce_sum(inverse_point_densities)
return (inverse_point_densities * tf.cast(num_points, dtype=tf.float32) /
total_inverse_density)
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 fn():
"""Loss function for when number of input and output boxes is positive."""
if is_balanced:
weights = loss_utils.get_balanced_loss_weights_multiclass(
labels=input_boxes_instance_id)
else:
weights = tf.ones([tf.shape(input_boxes_instance_id)[0], 1],
dtype=tf.float32)
normalized_box_size = 5.0
predicted_boxes_length = output_boxes_length
predicted_boxes_height = output_boxes_height
predicted_boxes_width = output_boxes_width
predicted_boxes_center = output_boxes_center
predicted_boxes_rotation_matrix = output_boxes_rotation_matrix
gt_boxes_length = input_boxes_length
gt_boxes_height = input_boxes_height
gt_boxes_width = input_boxes_width
gt_boxes_center = input_boxes_center
gt_boxes_rotation_matrix = input_boxes_rotation_matrix
if loss_type in ['normalized_huber', 'normalized_euclidean']:
predicted_boxes_length /= (gt_boxes_length / normalized_box_size)
predicted_boxes_height /= (gt_boxes_height / normalized_box_size)
predicted_boxes_width /= (gt_boxes_width / normalized_box_size)
gt_boxes_length = tf.ones_like(
gt_boxes_length, dtype=tf.float32) * normalized_box_size
gt_boxes_height = tf.ones_like(
gt_boxes_height, dtype=tf.float32) * normalized_box_size
gt_boxes_width = tf.ones_like(
gt_boxes_width, dtype=tf.float32) * normalized_box_size
gt_box_corners = box_utils.get_box_corners_3d(
boxes_length=gt_boxes_length,
boxes_height=gt_boxes_height,
boxes_width=gt_boxes_width,
boxes_rotation_matrix=gt_boxes_rotation_matrix,
boxes_center=gt_boxes_center)
predicted_box_corners = box_utils.get_box_corners_3d(
boxes_length=predicted_boxes_length,
boxes_height=predicted_boxes_height,
boxes_width=predicted_boxes_width,
boxes_rotation_matrix=predicted_boxes_rotation_matrix,
boxes_center=predicted_boxes_center)
corner_weights = tf.tile(weights, [1, 8])
if loss_type in ['huber', 'normalized_huber']:
loss_fn = tf.keras.losses.Huber(
delta=delta, reduction=tf.keras.losses.Reduction.NONE)
elif loss_type in [
'normalized_absolute_difference', 'absolute_difference'
]:
loss_fn = tf.keras.losses.MeanAbsoluteError(
reduction=tf.keras.losses.Reduction.NONE)
else:
raise ValueError(('Unknown loss type %s.' % loss_type))
box_corner_losses = loss_fn(y_true=tf.reshape(gt_box_corners, [-1, 3]),
y_pred=tf.reshape(predicted_box_corners,
[-1, 3]))
return tf.reduce_mean(box_corner_losses *
tf.reshape(corner_weights, [-1]))
def fn():
"""Loss function for when number of input and output boxes is positive."""
if is_balanced:
weights = loss_utils.get_balanced_loss_weights_multiclass(
labels=input_boxes_instance_id)
else:
weights = tf.ones([tf.shape(input_boxes_instance_id)[0], 1],
dtype=tf.float32)
gt_center = tf.reshape(input_boxes_center, [-1, 3])
predicted_center = tf.reshape(output_boxes_center, [-1, 3])
if loss_type == 'huber':
loss_fn = tf.keras.losses.Huber(
delta=delta, reduction=tf.keras.losses.Reduction.NONE)
elif loss_type == 'absolute_difference':
loss_fn = tf.keras.losses.MeanAbsoluteError(
reduction=tf.keras.losses.Reduction.NONE)
else:
raise ValueError(('Unknown loss type %s.' % loss_type))
center_losses = loss_fn(y_true=gt_center, y_pred=predicted_center)
return tf.reduce_mean(center_losses * tf.reshape(weights, [-1]))
def knn_graph_from_points_unbatched(points,
k,
distance_upper_bound,
mask=None):
"""Returns the distances and indices of the neighbors of each point.
Note that each point will have at least k neighbors unless the number of
points is less than k. In that case, the python function that is wrapped in
py_function will raise a value error.
Args:
points: A tf.float32 tensor of size [N, D] where D is the point dimensions.
k: Number of neighbors for each point.
distance_upper_bound: Only build the graph using points that are closer than
this distance.
mask: If not None, A tf.bool tensor of size [N]. If None, it is ignored.
If not None, knn will be applied to only points where the mask is True.
The points where the mask is False will have themselves as their
neighbors.
Returns:
distances: A tf.float32 tensor of size [N, k].
indices: A tf.int32 tensor of [N, k].
"""
def fn(np_points, np_mask):
return np_knn_graph_from_points_unbatched(
points=np_points,
k=k,
distance_upper_bound=distance_upper_bound,
mask=np_mask)
num_points = tf.shape(points)[0]
if mask is None:
mask = tf.cast(tf.ones([num_points], dtype=tf.int32), dtype=tf.bool)
else:
mask = tf.reshape(mask, [num_points])
distances, indices = tf.compat.v1.py_func(fn, [points, mask],
[tf.float32, tf.int32])
distances = tf.reshape(distances, [num_points, k])
indices = tf.reshape(indices, [num_points, k])
return distances, indices
def __init__(self,
observation_spec,
num_of_goals,
name="RandomCategoricalGoalGenerator"):
"""Create a RandomCategoricalGoalGenerator.
Args:
observation_spec (nested TensorSpec): representing the observations.
num_of_goals (int): total number of goals the agent can sample from
name (str): name of the algorithm
"""
goal_spec = tf.TensorSpec((num_of_goals, ))
train_state_spec = GoalState(goal=goal_spec)
super().__init__(observation_spec=observation_spec,
action_spec=tensor_spec.BoundedTensorSpec(
shape=(num_of_goals, ),
dtype=tf.float32,
minimum=0.,
maximum=1.),
train_state_spec=train_state_spec,
name=name)
self._num_of_goals = num_of_goals
self._p_goal = tf.ones(self._num_of_goals)
def linear_classifier(embeddings, num_classes, cosine_classifier,
cosine_logits_multiplier, use_weight_norm, weight_decay):
"""Forward pass through a linear classifier, or possibly a cosine classifier.
Args:
embeddings: A Tensor of size [batch size, embedding dim].
num_classes: An integer; the dimension of the classification.
cosine_classifier: A bool. If true, a cosine classifier is used, which does
not require a bias.
cosine_logits_multiplier: A float. Only used if cosine_classifier is True,
and multiplies the resulting logits.
use_weight_norm: A bool. Whether weight norm was used. If so, then if using
cosine classifier, normalize only the embeddings but not the weights.
weight_decay: A float; the scalar multiple on the L2 regularization of the
weight matrix.
Returns:
logits: A Tensor of size [batch size, num outputs].
"""
embedding_dims = embeddings.get_shape().as_list()[-1]
if use_weight_norm:
# A variable to keep track of whether the initialization has already
# happened.
data_dependent_init_done = tf.get_variable('data_dependent_init_done',
initializer=0,
dtype=tf.int32,
trainable=False)
w_fc = tf.get_variable('w_fc', [embedding_dims, num_classes],
initializer=tf.random_normal_initializer(
0, 0.05),
trainable=True)
# This init is temporary as it needs to be done in a data-dependent way.
# It will be overwritten during the first forward pass through this layer.
g = tf.get_variable('g',
dtype=tf.float32,
initializer=tf.ones([num_classes]),
trainable=True)
b_fc = None
if not cosine_classifier:
# Also initialize a bias.
b_fc = tf.get_variable('b_fc',
initializer=tf.zeros([num_classes]),
trainable=True)
def _do_data_dependent_init():
"""Returns ops for the data-dependent init of g and maybe b_fc."""
w_fc_normalized = tf.nn.l2_normalize(w_fc.read_value(), [0])
output_init = tf.matmul(embeddings, w_fc_normalized)
mean_init, var_init = tf.nn.moments(output_init, [0])
# Data-dependent init values.
g_init_value = 1. / tf.sqrt(var_init + 1e-10)
ops = [tf.assign(g, g_init_value)]
if not cosine_classifier:
# Also initialize a bias in a data-dependent way.
b_fc_init_value = -mean_init * g_init_value
ops.append(tf.assign(b_fc, b_fc_init_value))
# Mark that the data-dependent initialization is done to prevent it from
# happening again in the future.
ops.append(tf.assign(data_dependent_init_done, 1))
return tf.group(*ops)
# Possibly perform data-dependent init (if it hasn't been done already).
init_op = tf.cond(tf.equal(data_dependent_init_done, 0),
_do_data_dependent_init, tf.no_op)
with tf.control_dependencies([init_op]):
# Apply weight normalization.
w_fc *= g / tf.sqrt(tf.reduce_sum(tf.square(w_fc), [0]))
# Forward pass through the layer defined by w_fc and b_fc.
logits = linear_classifier_forward_pass(embeddings, w_fc, b_fc,
cosine_classifier,
cosine_logits_multiplier,
True)
else:
# No weight norm.
w_fc = functional_backbones.weight_variable(
[embedding_dims, num_classes], weight_decay=weight_decay)
b_fc = None
if not cosine_classifier:
# Also initialize a bias.
b_fc = functional_backbones.bias_variable([num_classes])
# Forward pass through the layer defined by w_fc and b_fc.
logits = linear_classifier_forward_pass(embeddings, w_fc, b_fc,
cosine_classifier,
cosine_logits_multiplier,
False)
return logits
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 fit_gaussian_mixture(embeddings,
responsibilities,
damping=1e-7,
full_covariance=False):
"""Fits a unimodal Gaussian distribution `embeddings`.
Args:
embeddings: A [batch_size, embedding_dim] tf.Tensor of embeddings.
responsibilities: The per-component responsibilities.
damping: The scale of the covariance damping coefficient.
full_covariance: Whether to use a full or diagonal covariance.
Returns:
Parameter estimates for a Gaussian mixture model.
"""
num, dim = tf.split(tf.shape(input=embeddings), num_or_size_splits=2)
num, dim = tf.squeeze(num), tf.squeeze(dim)
num_classes = responsibilities.shape[1]
mixing_proportion = tf.einsum('jk->k', responsibilities)
mixing_proportion /= tf.cast(num, dtype=tf.float32)
mixing_logits = tf.math.log(mixing_proportion)
sample_mean = tf.einsum('ij,ik->jk', responsibilities, embeddings)
sample_mean /= tf.reduce_sum(
input_tensor=responsibilities, axis=0)[:, tf.newaxis]
centered_embeddings = (
embeddings[:, tf.newaxis, :] - sample_mean[tf.newaxis, :, :])
if full_covariance:
sample_covariance = tf.einsum('ijk,ijl->ijkl', centered_embeddings,
centered_embeddings) # Outer product.
sample_covariance += damping * tf.eye(dim) # Positive definiteness.
weighted_covariance = tf.einsum('ij,ijkl->jkl', responsibilities,
sample_covariance)
weighted_covariance /= tf.reduce_sum(
input_tensor=responsibilities, axis=0)[:, tf.newaxis, tf.newaxis]
return (
_split_and_squeeze(sample_mean, num_splits=num_classes),
_split_and_squeeze(weighted_covariance, num_splits=num_classes),
[mixing_logits],
)
else:
avg_x_squared = (
tf.matmul(responsibilities, embeddings**2, transpose_a=True) /
tf.reduce_sum(input_tensor=responsibilities, axis=0)[:, tf.newaxis])
avg_means_squared = sample_mean**2
avg_x_means = (
sample_mean *
tf.matmul(responsibilities, embeddings, transpose_a=True) /
tf.reduce_sum(input_tensor=responsibilities, axis=0)[:, tf.newaxis])
sample_variances = (
avg_x_squared - 2 * avg_x_means + avg_means_squared +
damping * tf.ones(dim))
log_variances = tf.math.log(sample_variances)
return (
_split_and_squeeze(sample_mean, num_splits=num_classes),
_split_and_squeeze(log_variances, num_splits=num_classes),
[mixing_logits],
)
def _build_train_op(self, optimizer):
"""Build the TensorFlow graph used to learn the bisimulation metric.
Args:
optimizer: a tf.train optimizer.
Returns:
A TensorFlow op to minimize the bisimulation loss.
"""
self.online_network = tf.make_template('Online',
self._network_template)
self.target_network = tf.make_template('Target',
self._network_template)
self.s1_ph = tf.placeholder(tf.float64, (self.batch_size, 2),
name='s1_ph')
self.s2_ph = tf.placeholder(tf.float64, (self.batch_size, 2),
name='s2_ph')
self.s1_online_distances = self.online_network(
self._concat_states(self.s1_ph))
self.s1_target_distances = self.target_network(
self._concat_states(self.s1_ph))
self.s2_target_distances = self.target_network(
self._concat_states(self.s2_ph))
self.action_ph = tf.placeholder(tf.int32, (self.batch_size,))
self.rewards_ph = tf.placeholder(tf.float64, (self.batch_size,))
# We use an expanding horizon for computing the distances.
self.bisim_horizon_ph = tf.placeholder(tf.float64, ())
# bisimulation_target_1 = rew_diff + gamma * next_distance.
bisimulation_target_1 = tf.stop_gradient(self._build_bisimulation_target())
# bisimulation_target_2 = curr_distance.
bisimulation_target_2 = tf.stop_gradient(self.s1_target_distances)
# We slowly taper in the maximum according to the bisim horizon.
bisimulation_target = tf.maximum(
bisimulation_target_1, bisimulation_target_2 * self.bisim_horizon_ph)
# We zero-out diagonal entries, since those are estimating the distance
# between a state and itself, which we know to be 0.
diagonal_mask = 1.0 - tf.diag(tf.ones(self.batch_size, dtype=tf.float64))
diagonal_mask = tf.reshape(diagonal_mask, (self.batch_size**2, 1))
bisimulation_target *= diagonal_mask
bisimulation_estimate = self.s1_online_distances
# We start with a mask that includes everything.
loss_mask = tf.ones(tf.shape(bisimulation_estimate))
# We have to enforce that states being compared are done only using the same
# action.
indicators = self.action_ph
indicators = tf.cast(indicators, tf.float64)
# indicators will initially have shape [batch_size], we first tile it:
square_ids = tf.tile([indicators], [self.batch_size, 1])
# We subtract square_ids from its transpose:
square_ids = square_ids - tf.transpose(square_ids)
# At this point all zero-entries are the ones with equal IDs.
# Now we would like to convert the zeros in this matrix to 1s, and make
# everything else a 0. We can do this with the following operation:
loss_mask = 1 - tf.abs(tf.sign(square_ids))
# Now reshape to match the shapes of the estimate and target.
loss_mask = tf.reshape(loss_mask, (self.batch_size**2, 1))
larger_targets = bisimulation_target - bisimulation_estimate
larger_targets_count = tf.reduce_sum(
tf.cast(larger_targets > 0., tf.float64))
tf.summary.scalar('Learning/LargerTargets', larger_targets_count)
tf.summary.scalar('Learning/NumUpdates', tf.count_nonzero(loss_mask))
tf.summary.scalar('Learning/BisimHorizon', self.bisim_horizon_ph)
bisimulation_loss = tf.losses.mean_squared_error(
bisimulation_target,
bisimulation_estimate,
weights=loss_mask)
tf.summary.scalar('Learning/loss', bisimulation_loss)
# Plot average distance between sampled representations.
average_distance = tf.reduce_mean(bisimulation_estimate)
tf.summary.scalar('Approx/AverageDistance', average_distance)
return optimizer.minimize(bisimulation_loss)