[Enhance] Support more Torchsparse blocks (#2532)

* enhance

* enhance

* fix __init__

* fix minku backbone

mmdet3d/models/backbones/minkunet_backbone.py

@@ -5,8 +5,7 @@ from mmengine.model import BaseModule
 from mmengine.registry import MODELS
 from torch import Tensor, nn
 
-from mmdet3d.models.layers import (TorchSparseConvModule,
-                                   TorchSparseResidualBlock)
+from mmdet3d.models.layers import TorchSparseBasicBlock, TorchSparseConvModule
 from mmdet3d.models.layers.torchsparse import IS_TORCHSPARSE_AVAILABLE
 from mmdet3d.utils import OptMultiConfig
 
@@ -64,11 +63,11 @@ class MinkUNetBackbone(BaseModule):
                         encoder_channels[i],
                         kernel_size=2,
                         stride=2),
-                    TorchSparseResidualBlock(
+                    TorchSparseBasicBlock(
                         encoder_channels[i],
                         encoder_channels[i + 1],
                         kernel_size=3),
-                    TorchSparseResidualBlock(
+                    TorchSparseBasicBlock(
                         encoder_channels[i + 1],
                         encoder_channels[i + 1],
                         kernel_size=3)))
@@ -82,11 +81,11 @@ class MinkUNetBackbone(BaseModule):
                         stride=2,
                         transposed=True),
                     nn.Sequential(
-                        TorchSparseResidualBlock(
+                        TorchSparseBasicBlock(
                             decoder_channels[i + 1] + encoder_channels[-2 - i],
                             decoder_channels[i + 1],
                             kernel_size=3),
-                        TorchSparseResidualBlock(
+                        TorchSparseBasicBlock(
                             decoder_channels[i + 1],
                             decoder_channels[i + 1],
                             kernel_size=3))

mmdet3d/models/layers/__init__.py

@@ -14,7 +14,8 @@ from .pointnet_modules import (PAConvCUDASAModule, PAConvCUDASAModuleMSG,
                                build_sa_module)
 from .sparse_block import (SparseBasicBlock, SparseBottleneck,
                            make_sparse_convmodule)
-from .torchsparse_block import TorchSparseConvModule, TorchSparseResidualBlock
+from .torchsparse_block import (TorchSparseBasicBlock, TorchSparseBottleneck,
+                                TorchSparseConvModule)
 from .transformer import GroupFree3DMHA
 from .vote_module import VoteModule
 
@@ -28,5 +29,5 @@ __all__ = [
     'nms_normal_bev', 'build_sa_module', 'PointSAModuleMSG', 'PointSAModule',
     'PointFPModule', 'PAConvSAModule', 'PAConvSAModuleMSG',
     'PAConvCUDASAModule', 'PAConvCUDASAModuleMSG', 'TorchSparseConvModule',
-    'TorchSparseResidualBlock'
+    'TorchSparseBasicBlock', 'TorchSparseBottleneck'
 ]

mmdet3d/models/layers/torchsparse/torchsparse_wrapper.py

 # Copyright (c) OpenMMLab. All rights reserved.
+import torch.nn as nn
 from mmengine.registry import MODELS
 
 
@@ -7,12 +8,21 @@ def register_torchsparse() -> bool:
     try:
         from torchsparse.nn import (BatchNorm, Conv3d, GroupNorm, LeakyReLU,
                                     ReLU)
+        from torchsparse.nn.utils import fapply
+        from torchsparse.tensor import SparseTensor
     except ImportError:
         return False
     else:
+
+        class SyncBatchNorm(nn.SyncBatchNorm):
+
+            def forward(self, input: SparseTensor) -> SparseTensor:
+                return fapply(input, super().forward)
+
         MODELS._register_module(Conv3d, 'TorchSparseConv3d')
-        MODELS._register_module(BatchNorm, 'TorchSparseBatchNorm')
-        MODELS._register_module(GroupNorm, 'TorchSparseGroupNorm')
+        MODELS._register_module(BatchNorm, 'TorchSparseBN')
+        MODELS._register_module(SyncBatchNorm, 'TorchSparseSyncBN')
+        MODELS._register_module(GroupNorm, 'TorchSparseGN')
         MODELS._register_module(ReLU, 'TorchSparseReLU')
         MODELS._register_module(LeakyReLU, 'TorchSparseLeakyReLU')
         return True

mmdet3d/models/layers/torchsparse_block.py

 # Copyright (c) OpenMMLab. All rights reserved.
 from typing import Sequence, Union
 
+from mmcv.cnn import build_activation_layer, build_norm_layer
 from mmengine.model import BaseModule
 from torch import nn
 
-from mmdet3d.utils import OptConfigType
+from mmdet3d.utils import ConfigType, OptConfigType
 from .torchsparse import IS_TORCHSPARSE_AVAILABLE
 
 if IS_TORCHSPARSE_AVAILABLE:
@@ -23,37 +24,46 @@ class TorchSparseConvModule(BaseModule):
         kernel_size (int or Tuple[int]): Kernel_size of block.
         stride (int or Tuple[int]): Stride of the first block. Defaults to 1.
         dilation (int): Dilation of block. Defaults to 1.
+        bias (bool): Whether use bias in conv. Defaults to False.
         transposed (bool): Whether use transposed convolution operator.
             Defaults to False.
+        norm_cfg (:obj:`ConfigDict` or dict): The config of normalization.
         init_cfg (:obj:`ConfigDict` or dict, optional): Initialization config.
             Defaults to None.
     """
 
-    def __init__(
-        self,
-        in_channels: int,
-        out_channels: int,
-        kernel_size: Union[int, Sequence[int]],
-        stride: Union[int, Sequence[int]] = 1,
-        dilation: int = 1,
-        bias: bool = False,
-        transposed: bool = False,
-        init_cfg: OptConfigType = None,
-    ) -> None:
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 kernel_size: Union[int, Sequence[int]],
+                 stride: Union[int, Sequence[int]] = 1,
+                 dilation: int = 1,
+                 bias: bool = False,
+                 transposed: bool = False,
+                 norm_cfg: ConfigType = dict(type='TorchSparseBN'),
+                 act_cfg: ConfigType = dict(
+                     type='TorchSparseReLU', inplace=True),
+                 init_cfg: OptConfigType = None,
+                 **kwargs) -> None:
         super().__init__(init_cfg)
-        self.net = nn.Sequential(
+        layers = [
             spnn.Conv3d(in_channels, out_channels, kernel_size, stride,
-                        dilation, bias, transposed),
-            spnn.BatchNorm(out_channels),
-            spnn.ReLU(inplace=True),
-        )
+                        dilation, bias, transposed)
+        ]
+        if norm_cfg is not None:
+            _, norm = build_norm_layer(norm_cfg, out_channels)
+            layers.append(norm)
+        if act_cfg is not None:
+            activation = build_activation_layer(act_cfg)
+            layers.append(activation)
+        self.net = nn.Sequential(*layers)
 
     def forward(self, x: SparseTensor) -> SparseTensor:
         out = self.net(x)
         return out
 
 
-class TorchSparseResidualBlock(BaseModule):
+class TorchSparseBasicBlock(BaseModule):
     """Torchsparse residual basic block for MinkUNet.
 
     Args:
@@ -62,38 +72,114 @@ class TorchSparseResidualBlock(BaseModule):
         kernel_size (int or Tuple[int]): Kernel_size of block.
         stride (int or Tuple[int]): Stride of the first block. Defaults to 1.
         dilation (int): Dilation of block. Defaults to 1.
+        bias (bool): Whether use bias in conv. Defaults to False.
+        norm_cfg (:obj:`ConfigDict` or dict): The config of normalization.
         init_cfg (:obj:`ConfigDict` or dict, optional): Initialization config.
             Defaults to None.
     """
 
-    def __init__(
-        self,
-        in_channels: int,
-        out_channels: int,
-        kernel_size: Union[int, Sequence[int]],
-        stride: Union[int, Sequence[int]] = 1,
-        dilation: int = 1,
-        bias: bool = False,
-        init_cfg: OptConfigType = None,
-    ) -> None:
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 kernel_size: Union[int, Sequence[int]] = 3,
+                 stride: Union[int, Sequence[int]] = 1,
+                 dilation: int = 1,
+                 bias: bool = False,
+                 norm_cfg: ConfigType = dict(type='TorchSparseBN'),
+                 init_cfg: OptConfigType = None,
+                 **kwargs) -> None:
         super().__init__(init_cfg)
+        _, norm1 = build_norm_layer(norm_cfg, out_channels)
+        _, norm2 = build_norm_layer(norm_cfg, out_channels)
+
         self.net = nn.Sequential(
             spnn.Conv3d(in_channels, out_channels, kernel_size, stride,
-                        dilation, bias),
-            spnn.BatchNorm(out_channels),
-            spnn.ReLU(inplace=True),
+                        dilation, bias), norm1, spnn.ReLU(inplace=True),
+            spnn.Conv3d(
+                out_channels,
+                out_channels,
+                kernel_size,
+                stride=1,
+                dilation=dilation,
+                bias=bias), norm2)
+
+        if in_channels == out_channels and stride == 1:
+            self.downsample = nn.Identity()
+        else:
+            _, norm3 = build_norm_layer(norm_cfg, out_channels)
+            self.downsample = nn.Sequential(
+                spnn.Conv3d(
+                    in_channels,
+                    out_channels,
+                    kernel_size=1,
+                    stride=stride,
+                    dilation=dilation,
+                    bias=bias), norm3)
+
+        self.relu = spnn.ReLU(inplace=True)
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        out = self.relu(self.net(x) + self.downsample(x))
+        return out
+
+
+class TorchSparseBottleneck(BaseModule):
+    """Torchsparse residual basic block for MinkUNet.
+
+    Args:
+        in_channels (int): In channels of block.
+        out_channels (int): Out channels of block.
+        kernel_size (int or Tuple[int]): Kernel_size of block.
+        stride (int or Tuple[int]): Stride of the second block. Defaults to 1.
+        dilation (int): Dilation of block. Defaults to 1.
+        bias (bool): Whether use bias in conv. Defaults to False.
+        norm_cfg (:obj:`ConfigDict` or dict): The config of normalization.
+        init_cfg (:obj:`ConfigDict` or dict, optional): Initialization config.
+            Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 kernel_size: Union[int, Sequence[int]] = 3,
+                 stride: Union[int, Sequence[int]] = 1,
+                 dilation: int = 1,
+                 bias: bool = False,
+                 norm_cfg: ConfigType = dict(type='TorchSparseBN'),
+                 init_cfg: OptConfigType = None,
+                 **kwargs) -> None:
+        super().__init__(init_cfg)
+        _, norm1 = build_norm_layer(norm_cfg, out_channels)
+        _, norm2 = build_norm_layer(norm_cfg, out_channels)
+        _, norm3 = build_norm_layer(norm_cfg, out_channels)
+
+        self.net = nn.Sequential(
+            spnn.Conv3d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                stride=1,
+                dilation=dilation,
+                bias=bias), norm1, spnn.ReLU(inplace=True),
             spnn.Conv3d(
                 out_channels,
                 out_channels,
                 kernel_size,
+                stride,
+                dilation=dilation,
+                bias=bias), norm2, spnn.ReLU(inplace=True),
+            spnn.Conv3d(
+                out_channels,
+                out_channels,
+                kernel_size=1,
                 stride=1,
                 dilation=dilation,
-                bias=bias),
-            spnn.BatchNorm(out_channels),
-        )
+                bias=bias), norm3)
+
         if in_channels == out_channels and stride == 1:
             self.downsample = nn.Identity()
         else:
+            _, norm4 = build_norm_layer(norm_cfg, out_channels)
             self.downsample = nn.Sequential(
                 spnn.Conv3d(
                     in_channels,
@@ -101,9 +187,7 @@ class TorchSparseResidualBlock(BaseModule):
                     kernel_size=1,
                     stride=stride,
                     dilation=dilation,
-                    bias=bias),
-                spnn.BatchNorm(out_channels),
-            )
+                    bias=bias), norm4)
 
         self.relu = spnn.ReLU(inplace=True)
 

tests/test_models/test_layers/test_torchsparse/test_torchsparse_module.py

@@ -7,8 +7,9 @@ from mmdet3d.models.layers.torchsparse import IS_TORCHSPARSE_AVAILABLE
 if IS_TORCHSPARSE_AVAILABLE:
     from torchsparse import SparseTensor
 
-    from mmdet3d.models.layers.torchsparse_block import (
-        TorchSparseConvModule, TorchSparseResidualBlock)
+    from mmdet3d.models.layers.torchsparse_block import (TorchSparseBasicBlock,
+                                                         TorchSparseBottleneck,
+                                                         TorchSparseConvModule)
 else:
     pytest.skip('test requires Torchsparse', allow_module_level=True)
 
@@ -53,8 +54,11 @@ def test_TorchsparseResidualBlock():
     # test
     input_sp_tensor = SparseTensor(voxel_features, coordinates)
 
-    sparse_block0 = TorchSparseResidualBlock(4, 16, kernel_size=3).cuda()
+    sparse_block0 = TorchSparseBasicBlock(4, 16, kernel_size=3).cuda()
+    sparse_block1 = TorchSparseBottleneck(4, 16, kernel_size=3).cuda()
 
     # test forward
-    out_features = sparse_block0(input_sp_tensor)
-    assert out_features.F.shape == torch.Size([4, 16])
+    out_features0 = sparse_block0(input_sp_tensor)
+    out_features1 = sparse_block1(input_sp_tensor)
+    assert out_features0.F.shape == torch.Size([4, 16])
+    assert out_features1.F.shape == torch.Size([4, 16])