# ---------------------------------------------------------------------------- # - Open3D: www.open3d.org - # ---------------------------------------------------------------------------- # Copyright (c) 2018-2024 www.open3d.org # SPDX-License-Identifier: MIT # ---------------------------------------------------------------------------- from ...python.ops import ops from .neighbor_search import FixedRadiusSearch, RadiusSearch import tensorflow as tf import numpy as np __all__ = ['ContinuousConv', 'SparseConv', 'SparseConvTranspose'] class ContinuousConv(tf.keras.layers.Layer): r"""Continuous Convolution. This convolution supports continuous input and output point positions. This layer implements the convolution defined in *B. Ummenhofer and V. Koltun, Lagrangian Fluid Simulation with Continuous Convolutions, ICLR 2020.* The convolution at position :math:`\mathbf x` is defined as .. math:: (f*g)(\mathbf x) = \frac{1}{\psi(\mathbf x)} \sum_{i \in \mathcal N(\mathbf x, R)} a(\mathbf x_i, \mathbf x)\; f_i\; g(\Lambda(\mathbf x_i - \mathbf x)). With :math:`f` as the input feature function and :math:`g` as the filter function. The input points are :math:`\mathbf x_i` and the input features are :math:`f_i`. The normalization :math:`\frac{1}{\psi(\mathbf x)}` can be turned on with the **normalize** parameter. The per neighbor value :math:`a(\mathbf x_i, \mathbf x)` can be used to implement window functions; see parameter **window_function**. The function :math:`\Lambda` for looking up filter values is defined by the parameters **coordinate_mapping** and **interpolation**. Example: This shows a minimal example of how to use the layer:: import tensorflow as tf import open3d.ml.tf as ml3d inp_positions = tf.random.normal([20,3]) inp_features = tf.random.normal([20,8]) out_positions = tf.random.normal([10,3]) conv = ml3d.layers.ContinuousConv(filters=16, kernel_size=[3,3,3]) out_features = conv(inp_features, inp_positions, out_positions, extents=2.0) Arguments: filters: The number of filters/output channels. kernel_size: The spatial resolution of the filter, e.g. [3,3,3]. activation: The activation function to use. None means no activation. use_bias: If True adds an additive bias vector. kernel_initializer: Initializer for the kernel weights. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer for the kernel weights. bias_regularizer: Regularizer for the bias vector. align_corners: If true then the voxel centers of the outer voxels of the filter array are mapped to the boundary of the filter shape. If false then the boundary of the filter array is mapped to the boundary of the filter shape. coordinate_mapping: The mapping that is applied to the input coordinates. One of 'ball_to_cube_radial', 'ball_to_cube_volume_preserving', 'identity'. * 'ball_to_cube_radial' uses radial stretching to map a sphere to a cube. * 'ball_to_cube_volume_preserving' is using a more expensive volume preserving mapping to map a sphere to a cube. * 'identity' no mapping is applied to the coordinates. interpolation: One of 'linear', 'linear_border', 'nearest_neighbor'. * 'linear' is trilinear interpolation with coordinate clamping. * 'linear_border' uses a zero border if outside the range. * 'nearest_neighbor' uses the nearest neighbor instead of interpolation. normalize: If true then the result is normalized either by the number of points (neighbors_importance is null) or by the sum of the respective values in neighbors_importance. radius_search_ignore_query_points: If true the points that coincide with the center of the search window will be ignored. This excludes the query point if 'queries' and 'points' are the same point cloud. radius_search_metric: Either L1, L2 or Linf. Default is L2 offset: A single 3D vector used in the filter coordinate computation. The shape is [3]. window_function: Optional radial window function to steer the importance of points based on their distance to the center. The input to the function is a 1D tensor of distances (squared distances if radius_search_metric is 'L2'). The output must be a tensor of the same shape. Example:: def window_fn(r_sqr): return tf.clip_by_value((1 - r_sqr)**3, 0, 1) use_dense_layer_for_center: If True a linear dense layer is used to process the input features for each point. The result is added to the result of the convolution before adding the bias. This option is useful when using even kernel sizes that have no center element and input and output point sets are the same and 'radius_search_ignore_query_points' has been set to True. dense_kernel_initializer: Initializer for the kernel weights of the linear layer used for the center if 'use_dense_layer_for_center' is True. dense_kernel_regularizer: Regularizer for the kernel weights of the linear layer used for the center if 'use_dense_layer_for_center' is True. in_channels: This keyword argument is for compatibility with PyTorch. It is not used and in_channels will be inferred at the first execution of the layer. """ def __init__(self, filters, kernel_size, activation=None, use_bias=True, kernel_initializer='uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, align_corners=True, coordinate_mapping='ball_to_cube_radial', interpolation='linear', normalize=True, radius_search_ignore_query_points=False, radius_search_metric='L2', offset=None, window_function=None, use_dense_layer_for_center=False, dense_kernel_initializer='glorot_uniform', dense_kernel_regularizer=None, in_channels=None, **kwargs): from tensorflow.keras import activations, initializers, regularizers self.filters = filters self.kernel_size = kernel_size self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.align_corners = align_corners self.coordinate_mapping = coordinate_mapping self.interpolation = interpolation self.normalize = normalize self.radius_search_ignore_query_points = radius_search_ignore_query_points self.radius_search_metric = radius_search_metric self.dense_kernel_initializer = initializers.get( dense_kernel_initializer) self.dense_kernel_regularizer = regularizers.get( dense_kernel_regularizer) if offset is None: self.offset = tf.zeros(shape=(3,)) else: self.offset = offset self.window_function = window_function self.fixed_radius_search = FixedRadiusSearch( metric=self.radius_search_metric, ignore_query_point=self.radius_search_ignore_query_points, return_distances=not self.window_function is None) self.radius_search = RadiusSearch( metric=self.radius_search_metric, ignore_query_point=self.radius_search_ignore_query_points, return_distances=not self.window_function is None, normalize_distances=not self.window_function is None) self.use_dense_layer_for_center = use_dense_layer_for_center if self.use_dense_layer_for_center: self.dense = tf.keras.layers.Dense( self.filters, kernel_initializer=dense_kernel_initializer, kernel_regularizer=dense_kernel_regularizer, use_bias=False) super().__init__(**kwargs) def build(self, inp_features_shape): self.in_channels = inp_features_shape[-1] kernel_shape = tf.TensorShape( (*self.kernel_size, self.in_channels, self.filters)) self.kernel = self.add_weight( name="kernel", shape=kernel_shape, initializer=self.kernel_initializer, regularizer=self.kernel_regularizer, trainable=self.trainable, ) if self.use_bias: bias_shape = tf.TensorShape((self.filters,)) self.bias = self.add_weight( name="bias", shape=bias_shape, initializer=self.bias_initializer, regularizer=self.bias_regularizer, trainable=self.trainable, ) super().build(inp_features_shape) def call(self, inp_features, inp_positions, out_positions, extents, inp_importance=None, fixed_radius_search_hash_table=None, user_neighbors_index=None, user_neighbors_row_splits=None, user_neighbors_importance=None): """This function computes the output features. Arguments: inp_features: A 2D tensor which stores a feature vector for each input point. *This argument must be given as a positional argument!* inp_positions: A 2D tensor with the 3D point positions of each input point. The coordinates for each point is a vector with format [x,y,z]. out_positions: A 2D tensor with the 3D point positions of each output point. The coordinates for each point is a vector with format [x,y,z]. extents: The extent defines the spatial size of the filter for each output point. For 'ball to cube' coordinate mappings the extent defines the bounding box of the ball. The shape of the tensor is either [1] or [num output points]. inp_importance: Optional scalar importance value for each input point. fixed_radius_search_hash_table: A precomputed hash table generated with build_spatial_hash_table(). This input can be used to explicitly force the reuse of a hash table in special cases and is usually not needed. Note that the hash table must have been generated with the same 'points' array. Note that this parameter is only used if 'extents' is a scalar. user_neighbors_index: This parameter together with 'user_neighbors_row_splits' and 'user_neighbors_importance' allows to override the automatic neighbor search. This is the list of neighbor indices for each output point. This is a nested list for which the start and end of each sublist is defined by 'user_neighbors_row_splits'. user_neighbors_row_splits: Defines the start and end of each neighbors list in 'user_neighbors_index'. user_neighbors_importance: Defines a scalar importance value for each element in 'user_neighbors_index'. Returns: A tensor of shape [num output points, filters] with the output features. """ offset = self.offset if inp_importance is None: inp_importance = tf.ones((0,), dtype=tf.float32) extents = tf.convert_to_tensor(extents) return_distances = not self.window_function is None if not user_neighbors_index is None and not user_neighbors_row_splits is None: if user_neighbors_importance is None: neighbors_importance = tf.ones((0,), dtype=tf.float32) else: neighbors_importance = user_neighbors_importance neighbors_index = user_neighbors_index neighbors_row_splits = user_neighbors_row_splits else: if extents.shape.rank == 0: radius = 0.5 * extents self.nns = self.fixed_radius_search( inp_positions, queries=out_positions, radius=radius, hash_table=fixed_radius_search_hash_table) if return_distances: if self.radius_search_metric == 'L2': neighbors_distance_normalized = self.nns.neighbors_distance / ( radius * radius) else: # L1 neighbors_distance_normalized = self.nns.neighbors_distance / radius elif extents.shape.rank == 1: radii = 0.5 * extents self.nns = self.radius_search(inp_positions, queries=out_positions, radii=radii) else: raise Exception("extents rank must be 0 or 1") if self.window_function is None: neighbors_importance = tf.ones((0,), dtype=tf.float32) else: neighbors_importance = self.window_function( neighbors_distance_normalized) neighbors_index = self.nns.neighbors_index neighbors_row_splits = self.nns.neighbors_row_splits # for stats and debugging num_pairs = tf.shape(neighbors_index)[0] self._avg_neighbors = tf.dtypes.cast( num_pairs, tf.float32) / tf.dtypes.cast( tf.shape(out_positions)[0], tf.float32) extents_rank2 = extents while extents_rank2.shape.rank < 2: extents_rank2 = tf.expand_dims(extents_rank2, axis=-1) self._conv_values = { 'filters': self.kernel, 'out_positions': out_positions, 'extents': extents_rank2, 'offset': offset, 'inp_positions': inp_positions, 'inp_features': inp_features, 'inp_importance': inp_importance, 'neighbors_index': neighbors_index, 'neighbors_row_splits': neighbors_row_splits, 'neighbors_importance': neighbors_importance, 'align_corners': self.align_corners, 'coordinate_mapping': self.coordinate_mapping, 'interpolation': self.interpolation, 'normalize': self.normalize, } out_features = ops.continuous_conv(**self._conv_values) self._conv_output = out_features if self.use_dense_layer_for_center: self._dense_output = self.dense(inp_features) out_features = out_features + self._dense_output if self.use_bias: out_features += self.bias out_features = self.activation(out_features) return out_features def compute_output_shape(self, inp_features_shape): return tf.TensorShape((None, self.filters)) class SparseConv(tf.keras.layers.Layer): """Sparse Convolution. This layer computes a convolution which is only evaluated at the specified output positions. The layer assumes that input and output points lie on a regular grid. Example: This shows a minimal example of how to use the layer:: import tensorflow as tf import open3d.ml.tf as ml3d # +0.5 to move the points to the voxel center inp_positions = tf.cast(tf.random.uniform([20,3], 0, 10, dtype=tf.int32), tf.float32)+0.5 inp_features = tf.random.normal([20,8]) out_positions = tf.cast(tf.random.uniform([20,3], 0, 10, dtype=tf.int32), tf.float32)+0.5 conv = ml3d.layers.SparseConv(filters=16, kernel_size=[3,3,3]) out_features = conv(inp_features, inp_positions, out_positions, voxel_size=1.0) Arguments: filters: The number of filters/output channels. kernel_size: The spatial resolution of the filter, e.g. [3,3,3]. activation: The activation function to use. None means no activation. use_bias: If True adds an additive bias vector. kernel_initializer: Initializer for the kernel weights. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer for the kernel weights. bias_regularizer: Regularizer for the bias vector. normalize: If true then the result is normalized by the number of input points. offset: A single 3D vector used in the filter coordinate computation. The shape is [3]. This can be used to control how the filters are centered. It will be set automatically for kernels with even sizes. in_channels: This keyword argument is for compatibility with PyTorch. It is not used and in_channels will be inferred at the first execution of the layer. """ def __init__(self, filters, kernel_size, activation=None, use_bias=True, kernel_initializer='uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, normalize=False, offset=None, in_channels=None, **kwargs): from tensorflow.keras import activations, initializers, regularizers self.filters = filters self.kernel_size = kernel_size self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.normalize = normalize if not (np.asarray(kernel_size) == kernel_size[0]).all(): raise Exception("Only cubic kernel sizes are supported.") if offset is None: if kernel_size[0] % 2: self.offset = tf.zeros(shape=(3,)) else: self.offset = tf.fill([3], -0.5) else: self.offset = offset self.fixed_radius_search = FixedRadiusSearch(metric='Linf', ignore_query_point=False, return_distances=False) super().__init__(**kwargs) def build(self, inp_features_shape): self.in_channels = inp_features_shape[-1] kernel_shape = tf.TensorShape( (*self.kernel_size, self.in_channels, self.filters)) self.kernel = self.add_weight( name="kernel", shape=kernel_shape, initializer=self.kernel_initializer, regularizer=self.kernel_regularizer, trainable=self.trainable, ) if self.use_bias: bias_shape = tf.TensorShape((self.filters,)) self.bias = self.add_weight( name="bias", shape=bias_shape, initializer=self.bias_initializer, regularizer=self.bias_regularizer, trainable=self.trainable, ) super().build(inp_features_shape) def call(self, inp_features, inp_positions, out_positions, voxel_size, inp_importance=None, fixed_radius_search_hash_table=None): """This function computes the output features. Arguments: inp_features: A 2D tensor which stores a feature vector for each input point. *This argument must be given as a positional argument!* inp_positions: A 2D tensor with the 3D point positions of each input point. The coordinates for each point is a vector with format [x,y,z]. out_positions: A 2D tensor with the 3D point positions of each output point. The coordinates for each point is a vector with format [x,y,z]. voxel_size: A scalar float that defines the edge length of a voxel. inp_importance: Optional scalar importance value for each input point. fixed_radius_search_hash_table: A precomputed hash table generated with build_spatial_hash_table(). This input can be used to explicitly force the reuse of a hash table in special cases and is usually not needed. Note that the hash table must have been generated with the same 'points' array. Note that this parameter is only used if 'extents' is a scalar. Returns: A tensor of shape [num output points, filters] with the output features. """ offset = self.offset voxel_size = tf.convert_to_tensor(voxel_size, dtype=inp_positions.dtype) if voxel_size.shape.rank != 0: raise Exception("voxel_size must be a scalar") if inp_importance is None: inp_importance = tf.ones((0,), dtype=tf.float32) if isinstance(inp_features, tf.RaggedTensor): if not (isinstance(inp_positions, tf.RaggedTensor) and isinstance(out_positions, tf.RaggedTensor)): raise Exception( "All of inp_positions, inp_features and out_positions must be tf.Tensor, or tf.RaggedTensor" ) hash_table_size_factor = 1 / 64 self.nns = self.fixed_radius_search( inp_positions, queries=out_positions - offset * voxel_size, radius=self.kernel_size[0] * voxel_size * 0.51, hash_table_size_factor=hash_table_size_factor, hash_table=fixed_radius_search_hash_table) out_positions_split = None if isinstance(inp_positions, tf.RaggedTensor): inp_positions = inp_positions.values inp_features = inp_features.values out_positions_split = out_positions.row_splits out_positions = out_positions.values # for stats and debugging num_pairs = tf.shape(self.nns.neighbors_index)[0] self._avg_neighbors = num_pairs / tf.shape(out_positions)[0] extents_rank2 = tf.fill([1, 1], voxel_size * self.kernel_size[0]) self._conv_values = { 'filters': self.kernel, 'out_positions': out_positions, 'extents': extents_rank2, 'offset': offset, 'inp_positions': inp_positions, 'inp_features': inp_features, 'inp_importance': inp_importance, 'neighbors_index': self.nns.neighbors_index, 'neighbors_importance': tf.ones((0,), dtype=tf.float32), 'neighbors_row_splits': self.nns.neighbors_row_splits, 'align_corners': False, 'coordinate_mapping': 'identity', 'interpolation': 'nearest_neighbor', 'normalize': self.normalize, } out_features = ops.continuous_conv(**self._conv_values) self._conv_output = out_features if self.use_bias: out_features += self.bias out_features = self.activation(out_features) if out_positions_split is not None: out_features = tf.RaggedTensor.from_row_splits( values=out_features, row_splits=out_positions_split) return out_features def compute_output_shape(self, inp_features_shape): return tf.TensorShape((None, self.filters)) class SparseConvTranspose(tf.keras.layers.Layer): """Sparse Transposed Convolution. This layer computes a transposed convolution which is only evaluated at the specified output positions. Example: This shows a minimal example of how to use the layer:: import tensorflow as tf import open3d.ml.tf as ml3d # +0.5 to move the points to the voxel center inp_positions = tf.cast(tf.random.uniform([20,3], 0, 10, dtype=tf.int32), tf.float32)+0.5 inp_features = tf.random.normal([20,8]) out_positions = tf.cast(tf.random.uniform([20,3], 0, 10, dtype=tf.int32), tf.float32)+0.5 conv = ml3d.layers.SparseConvTranspose(filters=16, kernel_size=[3,3,3]) out_features = conv(inp_features, inp_positions, out_positions, voxel_size=1.0) Arguments: filters: The number of filters/output channels. kernel_size: The spatial resolution of the filter, e.g. [3,3,3]. activation: The activation function to use. None means no activation. use_bias: If True adds an additive bias vector. kernel_initializer: Initializer for the kernel weights. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer for the kernel weights. bias_regularizer: Regularizer for the bias vector. normalize: If true then the input features will be normalized with the number of output points. offset: A single 3D vector used in the filter coordinate computation. The shape is [3]. This can be used to control how the filters are centered. It will be set automatically for kernels with even sizes. in_channels: This keyword argument is for compatibility with PyTorch. It is not used and in_channels will be inferred at the first execution of the layer. """ def __init__(self, filters, kernel_size, activation=None, use_bias=True, kernel_initializer='uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, normalize=False, offset=None, in_channels=None, **kwargs): from tensorflow.keras import activations, initializers, regularizers self.filters = filters self.kernel_size = kernel_size self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.normalize = normalize if not (np.asarray(kernel_size) == kernel_size[0]).all(): raise Exception("Only cubic kernel sizes are supported.") if offset is None: if kernel_size[0] % 2: self.offset = tf.zeros(shape=(3,)) else: self.offset = tf.fill([3], -0.5) else: self.offset = offset self.fixed_radius_search = FixedRadiusSearch(metric='Linf', ignore_query_point=False, return_distances=False) super().__init__(**kwargs) def build(self, inp_features_shape): self.in_channels = inp_features_shape[-1] kernel_shape = tf.TensorShape( (*self.kernel_size, self.in_channels, self.filters)) self.kernel = self.add_weight( name="kernel", shape=kernel_shape, initializer=self.kernel_initializer, regularizer=self.kernel_regularizer, trainable=self.trainable, ) if self.use_bias: bias_shape = tf.TensorShape((self.filters,)) self.bias = self.add_weight( name="bias", shape=bias_shape, initializer=self.bias_initializer, regularizer=self.bias_regularizer, trainable=self.trainable, ) super().build(inp_features_shape) def call(self, inp_features, inp_positions, out_positions, voxel_size, out_importance=None, fixed_radius_search_hash_table=None): """This function computes the output features. Arguments: inp_features: A 2D tensor which stores a feature vector for each input point. *This argument must be given as a positional argument!* inp_positions: A 2D tensor with the 3D point positions of each input point. The coordinates for each point is a vector with format [x,y,z]. out_positions: A 2D tensor with the 3D point positions of each output point. The coordinates for each point is a vector with format [x,y,z]. voxel_size: A scalar float that defines the edge length of a voxel. out_importance: Optional scalar importance value for each output point. fixed_radius_search_hash_table: A precomputed hash table generated with build_spatial_hash_table(). This input can be used to explicitly force the reuse of a hash table in special cases and is usually not needed. Note that the hash table must have been generated with the same 'points' array. Note that this parameter is only used if 'extents' is a scalar. Returns: A tensor of shape [num output points, filters] with the output features. """ offset = self.offset voxel_size = tf.convert_to_tensor(voxel_size, dtype=inp_positions.dtype) if voxel_size.shape.rank != 0: raise Exception("voxel_size must be a scalar") if out_importance is None: out_importance = tf.ones((0,), dtype=tf.float32) empty_vec = tf.ones((0,), dtype=tf.float32) if isinstance(inp_features, tf.RaggedTensor): if not (isinstance(inp_positions, tf.RaggedTensor) and isinstance(out_positions, tf.RaggedTensor)): raise Exception( "All of inp_positions, inp_features and out_positions must be tf.Tensor, or tf.RaggedTensor" ) hash_table_size_factor = 1 / 64 self.nns_inp = self.fixed_radius_search( out_positions, queries=inp_positions - offset * voxel_size, radius=self.kernel_size[0] * voxel_size * 0.51, hash_table_size_factor=hash_table_size_factor, hash_table=fixed_radius_search_hash_table) out_positions_split = None if isinstance(inp_positions, tf.RaggedTensor): inp_positions = inp_positions.values inp_features = inp_features.values out_positions_split = out_positions.row_splits out_positions = out_positions.values num_out = tf.shape(out_positions, out_type=tf.int64)[0] neighbors_index, neighbors_row_splits, _ = ops.invert_neighbors_list( num_out, self.nns_inp.neighbors_index, self.nns_inp.neighbors_row_splits, empty_vec) # for stats and debugging num_pairs = tf.shape(neighbors_index)[0] self._avg_neighbors = num_pairs / tf.shape(out_positions)[0] extents_rank2 = tf.fill([1, 1], voxel_size * self.kernel_size[0]) self._conv_values = { 'filters': self.kernel, 'out_positions': out_positions, 'extents': extents_rank2, 'offset': offset, 'inp_positions': inp_positions, 'inp_features': inp_features, 'out_importance': out_importance, 'inp_neighbors_index': self.nns_inp.neighbors_index, 'inp_neighbors_importance_sum': empty_vec, 'inp_neighbors_row_splits': self.nns_inp.neighbors_row_splits, 'neighbors_index': neighbors_index, 'neighbors_importance': empty_vec, 'neighbors_row_splits': neighbors_row_splits, 'align_corners': False, 'coordinate_mapping': 'identity', 'interpolation': 'nearest_neighbor', 'normalize': self.normalize, } out_features = ops.continuous_conv_transpose(**self._conv_values) self._conv_output = out_features if self.use_bias: out_features += self.bias out_features = self.activation(out_features) if out_positions_split is not None: out_features = tf.RaggedTensor.from_row_splits( values=out_features, row_splits=out_positions_split) return out_features def compute_output_shape(self, inp_features_shape): return tf.TensorShape((None, self.filters))