open3d.ml.torch.ops.sparse_conv

open3d.ml.torch.ops.sparse_conv(filters, inp_features, inp_importance, neighbors_index, neighbors_kernel_index, neighbors_importance, neighbors_row_splits, normalize=False, max_temp_mem_MB=64)

General sparse convolution.

This op computes the features for the forward pass. This example shows how to use this op:

import tensorflow as tf
import open3d.ml.tf as ml3d

# This filter has 3 "spatial" elements with 8 input and 16 output channels
filters = tf.random.normal([3,8,16])

inp_features = tf.random.normal([5,8])


out_features = ml3d.ops.sparse_conv(
    filters,
    inp_features=inp_features,
    inp_importance=[],
    neighbors_index=[0,1,2, 1,2,3, 2,3,4],
    # neighbors_kernel_index defines which of the "spatial"
    # elements of the filter to use
    neighbors_kernel_index=tf.convert_to_tensor([0,1,2, 0,1,2, 0,1,2], dtype=tf.uint8),
    neighbors_importance=[],
    neighbors_row_splits=[0,3,6,9]
)

# or with pytorch
import torch
import open3d.ml.torch as ml3d

# This filter has 3 "spatial" elements with 8 input and 16 output channels
filters = torch.randn([3,8,16])

inp_features = torch.randn([5,8])


out_features = ml3d.ops.sparse_conv(
    filters,
    inp_features=inp_features,
    inp_importance=torch.FloatTensor([]),
    neighbors_index=torch.IntTensor([0,1,2, 1,2,3, 2,3,4]),
    # neighbors_kernel_index defines which of the "spatial"
    # elements of the filter to use
    neighbors_kernel_index=torch.ByteTensor([0,1,2, 0,1,2, 0,1,2]),
    neighbors_importance=torch.FloatTensor([]),
    neighbors_row_splits=torch.LongTensor([0,3,6,9])
)

normalize:

If True the output feature values will be normalized using the sum for ‘neighbors_importance’ for each output point.

max_temp_mem_MB:

Defines the maximum temporary memory in megabytes to be used for the GPU implementation. More memory means fewer kernel invocations. Note that the a minimum amount of temp memory will always be allocated even if this variable is set to 0.

filters:

The filter parameters. The shape of the filter is [depth, height, width, in_ch, out_ch]. The dimensions ‘depth’, ‘height’, ‘width’ define the spatial resolution of the filter. The spatial size of the filter is defined by the parameter ‘extents’.

inp_features:

A 2D tensor which stores a feature vector for each input point.

inp_importance:

An optional scalar importance for each input point. The features of each point will be multiplied with the corresponding value. The shape is [num input points]. Use a zero length Tensor to disable.

neighbors_index:

The neighbors_index stores a list of indices of neighbors for each output point as nested lists. The start and end of each list can be computed using ‘neighbors_row_splits’.

neighbors_kernel_index:

Defines which kernel element to use for each neighbor. This array has the same length as neighbors_index.

neighbors_importance:

Tensor of the same shape as ‘neighbors_index’ with a scalar value that is used to scale the features of each neighbor. Use a zero length Tensor to weigh each neighbor with 1.

neighbors_row_splits:

The exclusive prefix sum of the neighbor count for the output points including the total neighbor count as the last element. The size of this array is the number of output points + 1.

output_type: The type for the output.

out_features:

A Tensor with the output feature vectors for each output point.