DGCNN¶

warpconvnet.models.DGCNN implements the dynamic-graph CNN of Wang et al., Dynamic Graph CNN for Learning on Point Clouds, TOG 2019.

Each layer rebuilds a kNN graph in feature space and applies a PointConv edge function. Concatenated multi-scale features feed a global pool + classifier head.

Components¶

• DGCNNEncoder — stack of PointConvBlock stages with skip concatenation; outputs a feature Geometry.
• DGCNN — encoder + global pool + linear classifier head.

Signature¶

class DGCNN(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        encoder_channels: Union[int, List[int]] = (64, 64, 128, 256),
        classifier_channels: List[int] = [512, 256],
        knn_k: int = 20,
    ): ...

Usage¶

import torch
from warpconvnet.geometry.types.points import Points
from warpconvnet.models import DGCNN, DGCNNEncoder

device = "cuda"
pc = Points(
    [torch.rand(2048, 3) for _ in range(4)],
    [torch.rand(2048, 3) for _ in range(4)],
).to(device)

cls = DGCNN(in_channels=3, out_channels=40).to(device)
logits = cls(pc)              # (B, 40)

# Or use just the encoder for per-point features:
encoder = DGCNNEncoder(in_channels=3).to(device)
feat = encoder(pc)            # Points with concatenated multi-scale features

Reference¶

• Wang, Sun, Liu, Sarma, Bronstein, Solomon. Dynamic Graph CNN for Learning on Point Clouds. ACM TOG, 2019.