SplineConv implementation for Paddle

This module implements the SplineConv operators from

Matthias Fey, Jan Eric Lenssen, Frank Weichert, Heinrich Müller: SplineCNN: Fast Geometric Deep Learning with Continuous B-Spline Kernels (CVPR 2018).

It is in early development, and may have problems. Feel free to open an issue if you find one.

Requirements

It only needs paddle. It is tested on paddle >= 2.1.0, <= 2.2.0rc1, but should work for any recent paddle versions.

For development -- since we run tests against torch-spline-conv, you will need that.

Installation

pip install paddle-spline-conv

Usage

Here are some basic usage descriptions. See docstring in code for more detailed descriptions, types and shapes of parameters.

Currently only degree-1 splines are supported. But the basic operators have been ready, and adding more shouldn't be very hard. You are welcome to contribute for higher degree splines!

import paddle_spline_conv

# Stacked SplineConv layers implemented in SConv
g = paddle_spline_conv.nn.GraphData(x, edge_index, edge_attr)
# Input n_features and output n_features
sconv = paddle_spline_conv.nn.SConv(10, 40)
sconv(g)

# Standalone SplineConv layer
paddle_spline_conv.nn.SplineConv(
    in_channels: int,
    out_channels: int,
    dim: int,
    kernel_size: int,
    is_open_spline: bool = True,
    degree: int = 1,
    aggr: str = 'mean',
    root_weight: bool = True,
    bias: bool = True
)

# Standalone SplineConv functional API
paddle_spline_conv.functional.spline_conv(
    x: paddle.Tensor, edge_index: paddle.Tensor,
    pseudo: paddle.Tensor, weight: paddle.Tensor,
    kernel_size: paddle.Tensor, is_open_spline: paddle.Tensor,
    degree: int = 1, aggr: str = 'mean',
    root_weight: Optional[paddle.Tensor] = None,
    bias: Optional[paddle.Tensor] = None
)

# SplineConv-specific operators
paddle_spline_conv.ops.spline_basis
paddle_spline_conv.ops.spline_weighting
paddle_spline_conv.ops.basis_kernel_1d