Hi3Dv1.0-mlx / src /sparc3d /structures /point_modules.py
devin-lai's picture
Publish Hi3D MLX inference runtime, verified release docs and visual model card
5b702b4 verified
Raw History Blame Contribute Delete
2.4 kB
"""Module base and sequential container for layers that operate on a ``Point``."""
# Adapted from Pointcept / Point Transformer V3 under the MIT License.
# Modified for SPARC3D to drop the Lightning hyperparameters mixin and to
# dispatch sparse-convolution children through ``ops.sparse_backend``.
from __future__ import annotations
from typing import Optional
from torch import nn
from sparc3d.ops import sparse_backend
from sparc3d.structures.point import Point
class PointModule(nn.Module):
"""Marker base class for modules whose ``forward`` takes and returns a ``Point``."""
class PointSequential(PointModule):
"""
Sequential container that dispatches on each child's kind.
A ``PointModule`` child receives the whole ``Point``; a sparse convolution consumes
and replaces ``sparse_conv_feat``; any other module is applied to ``feat``.
"""
def __init__(self, *args: nn.Module):
"""
Args:
*args: Child modules, named by their index to match the checkpoint
key layout.
"""
super().__init__()
for idx, module in enumerate(args):
self.add_module(str(idx), module)
def add(self, module: nn.Module, name: Optional[str] = None) -> None:
"""Append ``module``, naming it by position when ``name`` is ``None``."""
if name is None:
name = str(len(self._modules))
if name in self._modules:
raise KeyError("name exists")
self.add_module(name, module)
def forward(self, input: Point) -> Point:
"""Run each child in order, dispatching on how it consumes the point.
A :class:`PointModule` receives the whole point; a sparse convolution receives
``sparse_conv_feat`` and its output is mirrored back onto ``feat``; anything
else receives ``feat`` alone.
"""
for module in self._modules.values():
if isinstance(module, PointModule):
input = module(input)
elif sparse_backend.is_sparse_conv(module):
assert isinstance(input, Point)
input.sparse_conv_feat = module(input.sparse_conv_feat)
input.feat = sparse_backend.get_feats(input.sparse_conv_feat)
else:
assert isinstance(input, Point)
input.update_feat(module(input.feat))
return input