propago

Graph learning primitives built on candle tensors.

This repo focuses on a small set of reusable building blocks (layers + small loops), not a full training framework.

Quickstart

[dependencies]
propago = "0.1"
candle-core = "0.9"
candle-nn = "0.9"

Hyperbolic distance on the Poincaré ball (Tensor-native):

use candle_core::{Device, Tensor};
use propago::hyperbolic::CandlePoincareBall;

let dev = &Device::Cpu;
let x = Tensor::from_vec(vec![0.10f32, 0.00, 0.00], (1, 3), dev)?;
let y = Tensor::from_vec(vec![0.00f32, 0.10, 0.00], (1, 3), dev)?;

let ball = CandlePoincareBall::new(1.0);
let d = ball.distance(&x, &y)?.to_vec2::<f32>()?[0][0];
assert!(d >= 0.0);

# Ok::<(), candle_core::Error>(())

Status / scope

• Keep interfaces small so higher-level graph stacks can integrate without tight coupling.
• HGCNConv uses a Tensor-native Poincaré ball implementation (CandlePoincareBall) so it can run on Candle backends (CPU/GPU).

API surface

• propago::GCNConv: simple graph convolution (linear projection + adjacency matmul).
• propago::HGCNConv: hyperbolic graph convolution on the Poincaré ball.

Notes:

• Many ops assume inputs are shaped [batch, d] (row-major feature vectors).

Backends

• Candle (default): --features backend-candle (enabled by default).
• Burn (opt-in): --features backend-burn exposes Burn-tensor Poincaré ops.
• MLX (opt-in): --features backend-mlx builds mlx-rs and requires cmake + Xcode MetalToolchain; tests force CPU for determinism.

License

MIT OR Apache-2.0