[Fix] Fix cylinder3d (#2377)

* mv cylinder blocks from sparse blocks to backbone

* update

configs/cylinder3d/cylinder3d_4xb2_3x_semantickitti.py → configs/cylinder3d/cylinder3d_4xb4_3x_semantickitti.py

@@ -28,6 +28,8 @@ param_scheduler = [
         gamma=0.1)
 ]
 
+train_dataloader = dict(batch_size=4, )
+
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically
 #       or not by default.

mmdet3d/models/backbones/cylinder3d.py

@@ -5,19 +5,389 @@ Please refer to `Cylinder3D github page
 <https://github.com/xinge008/Cylinder3D>`_ for details
 """
 
-from typing import List
+from typing import List, Optional
 
 import numpy as np
 import torch
-from mmcv.ops import SparseConvTensor
+from mmcv.cnn import build_activation_layer, build_norm_layer
+from mmcv.ops import (SparseConv3d, SparseConvTensor, SparseInverseConv3d,
+                      SubMConv3d)
 from mmengine.model import BaseModule
 
-from mmdet3d.models.layers.sparse_block import (AsymmeDownBlock, AsymmeUpBlock,
-                                                AsymmResBlock, DDCMBlock)
 from mmdet3d.registry import MODELS
 from mmdet3d.utils import ConfigType
 
 
+class AsymmResBlock(BaseModule):
+    """Asymmetrical Residual Block.
+
+    Args:
+        in_channels (int): Input channels of the block.
+        out_channels (int): Output channels of the block.
+        norm_cfg (:obj:`ConfigDict` or dict): Config dict for
+            normalization layer.
+        act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
+            Defaults to dict(type='LeakyReLU').
+        indice_key (str, optional): Name of indice tables. Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 norm_cfg: ConfigType,
+                 act_cfg: ConfigType = dict(type='LeakyReLU'),
+                 indice_key: Optional[str] = None):
+        super().__init__()
+
+        self.conv0_0 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(1, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act0_0 = build_activation_layer(act_cfg)
+        self.bn0_0 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv0_1 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(3, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act0_1 = build_activation_layer(act_cfg)
+        self.bn0_1 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv1_0 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(3, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act1_0 = build_activation_layer(act_cfg)
+        self.bn1_0 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv1_1 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(1, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act1_1 = build_activation_layer(act_cfg)
+        self.bn1_1 = build_norm_layer(norm_cfg, out_channels)[1]
+
+    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
+        """Forward pass."""
+        shortcut = self.conv0_0(x)
+
+        shortcut.features = self.act0_0(shortcut.features)
+        shortcut.features = self.bn0_0(shortcut.features)
+
+        shortcut = self.conv0_1(shortcut)
+        shortcut.features = self.act0_1(shortcut.features)
+        shortcut.features = self.bn0_1(shortcut.features)
+
+        res = self.conv1_0(x)
+        res.features = self.act1_0(res.features)
+        res.features = self.bn1_0(res.features)
+
+        res = self.conv1_1(res)
+        res.features = self.act1_1(res.features)
+        res.features = self.bn1_1(res.features)
+
+        res.features = res.features + shortcut.features
+
+        return res
+
+
+class AsymmeDownBlock(BaseModule):
+    """Asymmetrical DownSample Block.
+
+    Args:
+       in_channels (int): Input channels of the block.
+       out_channels (int): Output channels of the block.
+       norm_cfg (:obj:`ConfigDict` or dict): Config dict for
+            normalization layer.
+       act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
+            Defaults to dict(type='LeakyReLU').
+       pooling (bool): Whether pooling features at the end of
+           block. Defaults: True.
+       height_pooling (bool): Whether pooling features at
+           the height dimension. Defaults: False.
+       indice_key (str, optional): Name of indice tables. Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 norm_cfg: ConfigType,
+                 act_cfg: ConfigType = dict(type='LeakyReLU'),
+                 pooling: bool = True,
+                 height_pooling: bool = False,
+                 indice_key: Optional[str] = None):
+        super().__init__()
+        self.pooling = pooling
+
+        self.conv0_0 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(3, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act0_0 = build_activation_layer(act_cfg)
+        self.bn0_0 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv0_1 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(1, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act0_1 = build_activation_layer(act_cfg)
+        self.bn0_1 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv1_0 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(1, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act1_0 = build_activation_layer(act_cfg)
+        self.bn1_0 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv1_1 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(3, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'bef')
+        self.act1_1 = build_activation_layer(act_cfg)
+        self.bn1_1 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        if pooling:
+            if height_pooling:
+                self.pool = SparseConv3d(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    indice_key=indice_key,
+                    bias=False)
+            else:
+                self.pool = SparseConv3d(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    stride=(2, 2, 1),
+                    padding=1,
+                    indice_key=indice_key,
+                    bias=False)
+
+    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
+        """Forward pass."""
+        shortcut = self.conv0_0(x)
+        shortcut.features = self.act0_0(shortcut.features)
+        shortcut.features = self.bn0_0(shortcut.features)
+
+        shortcut = self.conv0_1(shortcut)
+        shortcut.features = self.act0_1(shortcut.features)
+        shortcut.features = self.bn0_1(shortcut.features)
+
+        res = self.conv1_0(x)
+        res.features = self.act1_0(res.features)
+        res.features = self.bn1_0(res.features)
+
+        res = self.conv1_1(res)
+        res.features = self.act1_1(res.features)
+        res.features = self.bn1_1(res.features)
+
+        res.features = res.features + shortcut.features
+
+        if self.pooling:
+            pooled_res = self.pool(res)
+            return pooled_res, res
+        else:
+            return res
+
+
+class AsymmeUpBlock(BaseModule):
+    """Asymmetrical UpSample Block.
+
+    Args:
+        in_channels (int): Input channels of the block.
+        out_channels (int): Output channels of the block.
+        norm_cfg (:obj:`ConfigDict` or dict): Config dict for
+                normalization layer.
+        act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
+                Defaults to dict(type='LeakyReLU').
+        indice_key (str, optional): Name of indice tables. Defaults to None.
+        up_key (str, optional): Name of indice tables used in
+            SparseInverseConv3d. Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 norm_cfg: ConfigType,
+                 act_cfg: ConfigType = dict(type='LeakyReLU'),
+                 indice_key: Optional[str] = None,
+                 up_key: Optional[str] = None):
+        super().__init__()
+
+        self.trans_conv = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(3, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key + 'new_up')
+        self.trans_act = build_activation_layer(act_cfg)
+        self.trans_bn = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv1 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(1, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.act1 = build_activation_layer(act_cfg)
+        self.bn1 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv2 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(3, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.act2 = build_activation_layer(act_cfg)
+        self.bn2 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.conv3 = SubMConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=(3, 3, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.act3 = build_activation_layer(act_cfg)
+        self.bn3 = build_norm_layer(norm_cfg, out_channels)[1]
+
+        self.up_subm = SparseInverseConv3d(
+            out_channels,
+            out_channels,
+            kernel_size=3,
+            indice_key=up_key,
+            bias=False)
+
+    def forward(self, x: SparseConvTensor,
+                skip: SparseConvTensor) -> SparseConvTensor:
+        """Forward pass."""
+        x_trans = self.trans_conv(x)
+        x_trans.features = self.trans_act(x_trans.features)
+        x_trans.features = self.trans_bn(x_trans.features)
+
+        # upsample
+        up = self.up_subm(x_trans)
+
+        up.features = up.features + skip.features
+
+        up = self.conv1(up)
+        up.features = self.act1(up.features)
+        up.features = self.bn1(up.features)
+
+        up = self.conv2(up)
+        up.features = self.act2(up.features)
+        up.features = self.bn2(up.features)
+
+        up = self.conv3(up)
+        up.features = self.act3(up.features)
+        up.features = self.bn3(up.features)
+
+        return up
+
+
+class DDCMBlock(BaseModule):
+    """Dimension-Decomposition based Context Modeling.
+
+    Args:
+        in_channels (int): Input channels of the block.
+        out_channels (int): Output channels of the block.
+        norm_cfg (:obj:`ConfigDict` or dict): Config dict for
+            normalization layer.
+        act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
+            Defaults to dict(type='Sigmoid').
+        indice_key (str, optional): Name of indice tables. Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 out_channels: int,
+                 norm_cfg: ConfigType,
+                 act_cfg: ConfigType = dict(type='Sigmoid'),
+                 indice_key: Optional[str] = None):
+        super().__init__()
+
+        self.conv1 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(3, 1, 1),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.bn1 = build_norm_layer(norm_cfg, out_channels)[1]
+        self.act1 = build_activation_layer(act_cfg)
+
+        self.conv2 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(1, 3, 1),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.bn2 = build_norm_layer(norm_cfg, out_channels)[1]
+        self.act2 = build_activation_layer(act_cfg)
+
+        self.conv3 = SubMConv3d(
+            in_channels,
+            out_channels,
+            kernel_size=(1, 1, 3),
+            padding=1,
+            bias=False,
+            indice_key=indice_key)
+        self.bn3 = build_norm_layer(norm_cfg, out_channels)[1]
+        self.act3 = build_activation_layer(act_cfg)
+
+    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
+        """Forward pass."""
+        shortcut = self.conv1(x)
+        shortcut.features = self.bn1(shortcut.features)
+        shortcut.features = self.act1(shortcut.features)
+
+        shortcut2 = self.conv2(x)
+        shortcut2.features = self.bn2(shortcut2.features)
+        shortcut2.features = self.act2(shortcut2.features)
+
+        shortcut3 = self.conv3(x)
+        shortcut3.features = self.bn3(shortcut3.features)
+        shortcut3.features = self.act3(shortcut3.features)
+        shortcut.features = shortcut.features + \
+            shortcut2.features + shortcut3.features
+
+        shortcut.features = shortcut.features * x.features
+
+        return shortcut
+
+
 @MODELS.register_module()
 class Asymm3DSpconv(BaseModule):
     """Asymmetrical 3D convolution networks.

mmdet3d/models/layers/sparse_block.py

 # Copyright (c) OpenMMLab. All rights reserved.
-from typing import Optional, Tuple, Union
+from typing import Tuple, Union
 
-from mmcv.cnn import build_activation_layer, build_conv_layer, build_norm_layer
+from mmcv.cnn import build_conv_layer, build_norm_layer
 from mmdet.models.backbones.resnet import BasicBlock, Bottleneck
 from torch import nn
 
-from mmdet3d.utils import ConfigType, OptConfigType
+from mmdet3d.utils import OptConfigType
 from .spconv import IS_SPCONV2_AVAILABLE
 
 if IS_SPCONV2_AVAILABLE:
     from spconv.pytorch import SparseConvTensor, SparseModule, SparseSequential
 else:
-    from mmcv.ops import (SparseConvTensor, SparseModule, SparseSequential,
-                          SparseConv3d, SparseInverseConv3d, SubMConv3d)
-
-from mmengine.model import BaseModule
+    from mmcv.ops import SparseConvTensor, SparseModule, SparseSequential
 
 
 def replace_feature(out: SparseConvTensor,
@@ -210,374 +207,3 @@ def make_sparse_convmodule(
 
     layers = SparseSequential(*layers)
     return layers
-
-
-# The following module only supports spconv_v1
-class AsymmResBlock(BaseModule):
-    """Asymmetrical Residual Block.
-
-    Args:
-        in_channels (int): Input channels of the block.
-        out_channels (int): Output channels of the block.
-        norm_cfg (:obj:`ConfigDict` or dict): Config dict for
-            normalization layer.
-       act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
-            Defaults to dict(type='LeakyReLU').
-        indice_key (str, optional): Name of indice tables. Defaults to None.
-    """
-
-    def __init__(self,
-                 in_channels: int,
-                 out_channels: int,
-                 norm_cfg: ConfigType,
-                 act_cfg: ConfigType = dict(type='LeakyReLU'),
-                 indice_key: Optional[str] = None):
-        super().__init__()
-
-        self.conv0_0 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(1, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act0_0 = build_activation_layer(act_cfg)
-        self.bn0_0 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv0_1 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(3, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act0_1 = build_activation_layer(act_cfg)
-        self.bn0_1 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv1_0 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(3, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act1_0 = build_activation_layer(act_cfg)
-        self.bn1_0 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv1_1 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(1, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act1_1 = build_activation_layer(act_cfg)
-        self.bn1_1 = build_norm_layer(norm_cfg, out_channels)[1]
-
-    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
-        """Forward pass."""
-        shortcut = self.conv0_0(x)
-
-        shortcut.features = self.act0_0(shortcut.features)
-        shortcut.features = self.bn0_0(shortcut.features)
-
-        shortcut = self.conv0_1(shortcut)
-        shortcut.features = self.act0_1(shortcut.features)
-        shortcut.features = self.bn0_1(shortcut.features)
-
-        res = self.conv1_0(x)
-        res.features = self.act1_0(res.features)
-        res.features = self.bn1_0(res.features)
-
-        res = self.conv1_1(res)
-        res.features = self.act1_1(res.features)
-        res.features = self.bn1_1(res.features)
-
-        res.features = res.features + shortcut.features
-
-        return res
-
-
-class AsymmeDownBlock(BaseModule):
-    """Asymmetrical DownSample Block.
-
-    Args:
-       in_channels (int): Input channels of the block.
-       out_channels (int): Output channels of the block.
-       norm_cfg (:obj:`ConfigDict` or dict): Config dict for
-            normalization layer.
-       act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
-            Defaults to dict(type='LeakyReLU').
-       pooling (bool): Whether pooling features at the end of
-           block. Defaults: True.
-       height_pooling (bool): Whether pooling features at
-           the height dimension. Defaults: False.
-       indice_key (str, optional): Name of indice tables. Defaults to None.
-    """
-
-    def __init__(self,
-                 in_channels: int,
-                 out_channels: int,
-                 norm_cfg: ConfigType,
-                 act_cfg: ConfigType = dict(type='LeakyReLU'),
-                 pooling: bool = True,
-                 height_pooling: bool = False,
-                 indice_key: Optional[str] = None):
-        super().__init__()
-        self.pooling = pooling
-
-        self.conv0_0 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(3, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act0_0 = build_activation_layer(act_cfg)
-        self.bn0_0 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv0_1 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(1, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act0_1 = build_activation_layer(act_cfg)
-        self.bn0_1 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv1_0 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(1, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act1_0 = build_activation_layer(act_cfg)
-        self.bn1_0 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv1_1 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(3, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'bef')
-        self.act1_1 = build_activation_layer(act_cfg)
-        self.bn1_1 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        if pooling:
-            if height_pooling:
-                self.pool = SparseConv3d(
-                    out_channels,
-                    out_channels,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    indice_key=indice_key,
-                    bias=False)
-            else:
-                self.pool = SparseConv3d(
-                    out_channels,
-                    out_channels,
-                    kernel_size=3,
-                    stride=(2, 2, 1),
-                    padding=1,
-                    indice_key=indice_key,
-                    bias=False)
-
-    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
-        """Forward pass."""
-        shortcut = self.conv0_0(x)
-        shortcut.features = self.act0_0(shortcut.features)
-        shortcut.features = self.bn0_0(shortcut.features)
-
-        shortcut = self.conv0_1(shortcut)
-        shortcut.features = self.act0_1(shortcut.features)
-        shortcut.features = self.bn0_1(shortcut.features)
-
-        res = self.conv1_0(x)
-        res.features = self.act1_0(res.features)
-        res.features = self.bn1_0(res.features)
-
-        res = self.conv1_1(res)
-        res.features = self.act1_1(res.features)
-        res.features = self.bn1_1(res.features)
-
-        res.features = res.features + shortcut.features
-
-        if self.pooling:
-            pooled_res = self.pool(res)
-            return pooled_res, res
-        else:
-            return res
-
-
-class AsymmeUpBlock(BaseModule):
-    """Asymmetrical UpSample Block.
-
-    Args:
-       in_channels (int): Input channels of the block.
-       out_channels (int): Output channels of the block.
-       norm_cfg (:obj:`ConfigDict` or dict): Config dict for
-            normalization layer.
-       act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
-            Defaults to dict(type='LeakyReLU').
-       indice_key (str, optional): Name of indice tables. Defaults to None.
-       up_key (str, optional): Name of indice tables used in
-           SparseInverseConv3d. Defaults to None.
-    """
-
-    def __init__(self,
-                 in_channels: int,
-                 out_channels: int,
-                 norm_cfg: ConfigType,
-                 act_cfg: ConfigType = dict(type='LeakyReLU'),
-                 indice_key: Optional[str] = None,
-                 up_key: Optional[str] = None):
-        super().__init__()
-
-        self.trans_conv = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(3, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key + 'new_up')
-        self.trans_act = build_activation_layer(act_cfg)
-        self.trans_bn = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv1 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(1, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.act1 = build_activation_layer(act_cfg)
-        self.bn1 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv2 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(3, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.act2 = build_activation_layer(act_cfg)
-        self.bn2 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.conv3 = SubMConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=(3, 3, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.act3 = build_activation_layer(act_cfg)
-        self.bn3 = build_norm_layer(norm_cfg, out_channels)[1]
-
-        self.up_subm = SparseInverseConv3d(
-            out_channels,
-            out_channels,
-            kernel_size=3,
-            indice_key=up_key,
-            bias=False)
-
-    def forward(self, x: SparseConvTensor,
-                skip: SparseConvTensor) -> SparseConvTensor:
-        """Forward pass."""
-        x_trans = self.trans_conv(x)
-        x_trans.features = self.trans_act(x_trans.features)
-        x_trans.features = self.trans_bn(x_trans.features)
-
-        # upsample
-        up = self.up_subm(x_trans)
-
-        up.features = up.features + skip.features
-
-        up = self.conv1(up)
-        up.features = self.act1(up.features)
-        up.features = self.bn1(up.features)
-
-        up = self.conv2(up)
-        up.features = self.act2(up.features)
-        up.features = self.bn2(up.features)
-
-        up = self.conv3(up)
-        up.features = self.act3(up.features)
-        up.features = self.bn3(up.features)
-
-        return up
-
-
-class DDCMBlock(BaseModule):
-    """Dimension-Decomposition based Context Modeling.
-
-    Args:
-        in_channels (int): Input channels of the block.
-        out_channels (int): Output channels of the block.
-        norm_cfg (:obj:`ConfigDict` or dict): Config dict for
-            normalization layer.
-        act_cfg (:obj:`ConfigDict` or dict): Config dict of activation layers.
-            Defaults to dict(type='Sigmoid').
-        indice_key (str, optional): Name of indice tables. Defaults to None.
-    """
-
-    def __init__(self,
-                 in_channels: int,
-                 out_channels: int,
-                 norm_cfg: ConfigType,
-                 act_cfg: ConfigType = dict(type='Sigmoid'),
-                 indice_key: Optional[str] = None):
-        super().__init__()
-
-        self.conv1 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(3, 1, 1),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.bn1 = build_norm_layer(norm_cfg, out_channels)[1]
-        self.act1 = build_activation_layer(act_cfg)
-
-        self.conv2 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(1, 3, 1),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.bn2 = build_norm_layer(norm_cfg, out_channels)[1]
-        self.act2 = build_activation_layer(act_cfg)
-
-        self.conv3 = SubMConv3d(
-            in_channels,
-            out_channels,
-            kernel_size=(1, 1, 3),
-            padding=1,
-            bias=False,
-            indice_key=indice_key)
-        self.bn3 = build_norm_layer(norm_cfg, out_channels)[1]
-        self.act3 = build_activation_layer(act_cfg)
-
-    def forward(self, x: SparseConvTensor) -> SparseConvTensor:
-        """Forward pass."""
-        shortcut = self.conv1(x)
-        shortcut.features = self.bn1(shortcut.features)
-        shortcut.features = self.act1(shortcut.features)
-
-        shortcut2 = self.conv2(x)
-        shortcut2.features = self.bn2(shortcut2.features)
-        shortcut2.features = self.act2(shortcut2.features)
-
-        shortcut3 = self.conv3(x)
-        shortcut3.features = self.bn3(shortcut3.features)
-        shortcut3.features = self.act3(shortcut3.features)
-        shortcut.features = shortcut.features + \
-            shortcut2.features + shortcut3.features
-
-        shortcut.features = shortcut.features * x.features
-
-        return shortcut

tests/test_models/test_segmentors/test_cylinder3d.py

@@ -17,7 +17,7 @@ class TestCylinder3D(unittest.TestCase):
         DefaultScope.get_instance('test_cylinder3d', scope_name='mmdet3d')
         setup_seed(0)
         cylinder3d_cfg = get_detector_cfg(
-            'cylinder3d/cylinder3d_4xb2_3x_semantickitti.py')
+            'cylinder3d/cylinder3d_4xb4_3x_semantickitti.py')
         cylinder3d_cfg.decode_head['ignore_index'] = 1
         model = MODELS.build(cylinder3d_cfg)
         num_gt_instance = 3