MinkUNet¶

warpconvnet.models.MinkUNet* implements the Minkowski U-Net family from Choy, Gwak, Savarese, 4D Spatio-Temporal ConvNets, CVPR 2019.

A symmetric encoder–decoder over voxelized inputs:

• Encoder: 4 strided SparseConv3d stages with BasicBlock / BottleneckBlock residual blocks
• Decoder: matching transposed-conv stages with skip connections from the encoder
• Output: per-voxel logits

Variants¶

Class	Block	Layers per stage	Params
MinkUNet18	BasicBlock	(2, 2, 2, 2, 2, 2, 2, 2)	~22 M
MinkUNet34	BasicBlock	(2, 3, 4, 6, 2, 2, 2, 2)	~38 M
MinkUNet50	BottleneckBlock	(2, 3, 4, 6, 2, 2, 2, 2)	~52 M
MinkUNet101	BottleneckBlock	(2, 3, 4, 23, 2, 2, 2, 2)	~83 M

PointMinkUNet18 / PointMinkUNet34 wrap the voxel network with PointToVoxel so the input/output types are Points.

Signature¶

class MinkUNetBase(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        planes: Tuple[int, ...] = (32, 64, 128, 256, 256, 128, 96, 96),
        layers: Tuple[int, ...] = (2, 2, 2, 2, 2, 2, 2, 2),
        init_dim: int = 32,
        block_type: type = BasicBlock,
    ): ...

Usage¶

Voxels in / voxels out¶

from warpconvnet.geometry.types.voxels import Voxels
from warpconvnet.models import MinkUNet34

model = MinkUNet34(in_channels=3, out_channels=20).cuda()
voxels: Voxels = ...                 # already voxelized
logits = model(voxels)               # Voxels with (N, 20) features

Points in / points out¶

from warpconvnet.geometry.types.points import Points
from warpconvnet.models import PointMinkUNet18

model = PointMinkUNet18(in_channels=3, out_channels=20).cuda()
pc: Points = ...
out = model(pc)                       # Points with (N, 20) features

Hydra¶

model:
  _target_: warpconvnet.models.MinkUNet34
  in_channels: 3
  out_channels: 20

Reference¶

• Choy, Gwak, Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. CVPR 2019.