[Refactor]: Unified parameter initialization

.github/workflows/build.yml

@@ -91,8 +91,8 @@ jobs:
       - name: Install mmdet3d dependencies
         run: |
           pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu101/${{matrix.torch_version}}/index.html
-          pip install mmdet==2.11.0
-          pip install mmsegmentation==0.14.0
+          pip install mmdet==2.14.0
+          pip install mmsegmentation==0.14.1
           pip install -r requirements.txt
       - name: Build and install
         run: |

configs/_base_/models/fcos3d.py

@@ -16,7 +16,6 @@
         out_channels=256,
         start_level=1,
         add_extra_convs=True,
-        extra_convs_on_inputs=False,  # use P5
         num_outs=5,
         relu_before_extra_convs=True),
     bbox_head=dict(

configs/_base_/models/imvotenet_image.py

@@ -99,8 +99,8 @@
             nms_across_levels=False,
             nms_pre=1000,
             nms_post=1000,
-            max_num=1000,
-            nms_thr=0.7,
+            max_per_img=1000,
+            nms=dict(type='nms', iou_threshold=0.7),
             min_bbox_size=0),
         img_rcnn=dict(
             score_thr=0.05,

configs/imvoxelnet/imvoxelnet_kitti-3d-car.py

@@ -1,6 +1,5 @@
 model = dict(
     type='ImVoxelNet',
-    pretrained='torchvision://resnet50',
     backbone=dict(
         type='ResNet',
         depth=50,
@@ -9,6 +8,7 @@
         frozen_stages=1,
         norm_cfg=dict(type='BN', requires_grad=False),
         norm_eval=True,
+        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50'),
         style='pytorch'),
     neck=dict(
         type='FPN',

docs/compatibility.md

@@ -4,6 +4,15 @@ This document provides detailed descriptions of the BC-breaking changes in MMDet
 
 ## MMDetection3D 0.15.0
 
+### Unified parameter initialization
+
+To unify the parameter initialization in OpenMMLab projects, MMCV supports `BaseModule` that accepts `init_cfg` to allow the modules' parameters initialized in a flexible and unified manner. Now the users need to explicitly call `model.init_weights()` in the training script to initialize the model (as in [here](https://github.com/open-mmlab/mmdetection3d/blob/master/tools/train.py#L183), previously this was handled by the detector. Please refer to PR #622 for details.
+
+### BackgroundPointsFilter
+
+We modified the dataset aumentation function `BackgroundPointsFilter`(in [here](https://github.com/open-mmlab/mmdetection3d/blob/mmdet3d/datasets/pipelines/transforms_3d.py#L1101)). In previous version of MMdetection3D, `BackgroundPointsFilter` changes the gt_bboxes_3d's bottom center to the gravity center. In MMDetection3D 0.15.0,
+`BackgroundPointsFilter` will not change it. Please refer to PR #609 for details.
+
 ### Enhance `IndoorPatchPointSample` transform
 
 We enhance the pipeline function `IndoorPatchPointSample` used in point cloud segmentation task by adding more choices for patch selection. Also, we plan to remove the unused parameter `sample_rate` in the future. Please modify the code as well as the config files accordingly if you use this transform.

docs/getting_started.md

@@ -12,8 +12,8 @@ The required versions of MMCV, MMDetection and MMSegmentation for different vers
 
 | MMDetection3D version | MMDetection version | MMSegmentation version |    MMCV version     |
 |:-------------------:|:-------------------:|:-------------------:|:-------------------:|
-| master              | mmdet>=2.10.0, <=2.11.0| mmseg==0.14.0 | mmcv-full>=1.3.1, <=1.4|
-| 0.14.0              | mmdet>=2.10.0, <=2.11.0| mmseg==0.14.0 | mmcv-full>=1.3.1, <=1.4|
+| master              | mmdet>=2.12.0          | mmseg>=0.14.1 | mmcv-full>=1.3.2, <=1.4|
+| 0.14.0              | mmdet>=2.10.0, <=2.11.0| mmseg>=0.13.0 | mmcv-full>=1.3.1, <=1.4|
 | 0.13.0              | mmdet>=2.10.0, <=2.11.0| Not required  | mmcv-full>=1.2.4, <=1.4|
 | 0.12.0              | mmdet>=2.5.0, <=2.11.0 | Not required  | mmcv-full>=1.2.4, <=1.4|
 | 0.11.0              | mmdet>=2.5.0, <=2.11.0 | Not required  | mmcv-full>=1.2.4, <=1.4|

docs/tutorials/customize_models.md

@@ -33,9 +33,6 @@ class HardVFE(nn.Module):
 
     def forward(self, x):  # should return a tuple
         pass
-
-    def init_weights(self, pretrained=None):
-        pass
 ```
 
 #### 2. Import the module
@@ -83,16 +80,13 @@ from ..builder import BACKBONES
 
 
 @BACKBONES.register_module()
-class SECOND(nn.Module):
+class SECOND(BaseModule):
 
     def __init__(self, arg1, arg2):
         pass
 
     def forward(self, x):  # should return a tuple
         pass
-
-    def init_weights(self, pretrained=None):
-        pass
 ```
 
 #### 2. Import the module
@@ -135,7 +129,7 @@ Create a new file `mmdet3d/models/necks/second_fpn.py`.
 from ..builder import NECKS
 
 @NECKS.register
-class SECONDFPN(nn.Module):
+class SECONDFPN(BaseModule):
 
     def __init__(self,
                  in_channels=[128, 128, 256],
@@ -144,7 +138,8 @@ class SECONDFPN(nn.Module):
                  norm_cfg=dict(type='BN', eps=1e-3, momentum=0.01),
                  upsample_cfg=dict(type='deconv', bias=False),
                  conv_cfg=dict(type='Conv2d', bias=False),
-                 use_conv_for_no_stride=False):
+                 use_conv_for_no_stride=False,
+                 init_cfg=None):
         pass
 
     def forward(self, X):
@@ -198,7 +193,7 @@ from mmdet.models.builder import HEADS
 from .bbox_head import BBoxHead
 
 @HEADS.register_module()
-class PartA2BboxHead(nn.Module):
+class PartA2BboxHead(BaseModule):
     """PartA2 RoI head."""
 
     def __init__(self,
@@ -224,11 +219,9 @@ class PartA2BboxHead(nn.Module):
                      type='CrossEntropyLoss',
                      use_sigmoid=True,
                      reduction='none',
-                     loss_weight=1.0)):
-        super(PartA2BboxHead, self).__init__()
-
-    def init_weights(self):
-        # conv layers are already initialized by ConvModule
+                     loss_weight=1.0),
+                 init_cfg=None):
+        super(PartA2BboxHead, self).__init__(init_cfg=init_cfg)
 
     def forward(self, seg_feats, part_feats):
 
@@ -242,15 +235,16 @@ from torch import nn as nn
 
 
 @HEADS.register_module()
-class Base3DRoIHead(nn.Module, metaclass=ABCMeta):
+class Base3DRoIHead(BaseModule, metaclass=ABCMeta):
     """Base class for 3d RoIHeads."""
 
     def __init__(self,
                  bbox_head=None,
                  mask_roi_extractor=None,
                  mask_head=None,
                  train_cfg=None,
-                 test_cfg=None):
+                 test_cfg=None,
+                 init_cfg=None):
 
     @property
     def with_bbox(self):
@@ -333,9 +327,13 @@ class PartAggregationROIHead(Base3DRoIHead):
                  part_roi_extractor=None,
                  bbox_head=None,
                  train_cfg=None,
-                 test_cfg=None):
+                 test_cfg=None,
+                 init_cfg=None):
         super(PartAggregationROIHead, self).__init__(
-            bbox_head=bbox_head, train_cfg=train_cfg, test_cfg=test_cfg)
+            bbox_head=bbox_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            init_cfg=init_cfg)
         self.num_classes = num_classes
         assert semantic_head is not None
         self.semantic_head = build_head(semantic_head)

mmdet3d/__init__.py

@@ -17,7 +17,7 @@ def digit_version(version_str):
     return digit_version
 
 
-mmcv_minimum_version = '1.3.1'
+mmcv_minimum_version = '1.3.8'
 mmcv_maximum_version = '1.4.0'
 mmcv_version = digit_version(mmcv.__version__)
 
@@ -27,17 +27,17 @@ def digit_version(version_str):
     f'MMCV=={mmcv.__version__} is used but incompatible. ' \
     f'Please install mmcv>={mmcv_minimum_version}, <={mmcv_maximum_version}.'
 
-mmdet_minimum_version = '2.10.0'
-mmdet_maximum_version = '2.11.0'
+mmdet_minimum_version = '2.14.0'
+mmdet_maximum_version = '3.0.0'
 mmdet_version = digit_version(mmdet.__version__)
 assert (mmdet_version >= digit_version(mmdet_minimum_version)
         and mmdet_version <= digit_version(mmdet_maximum_version)), \
     f'MMDET=={mmdet.__version__} is used but incompatible. ' \
     f'Please install mmdet>={mmdet_minimum_version}, ' \
     f'<={mmdet_maximum_version}.'
 
-mmseg_minimum_version = '0.14.0'
-mmseg_maximum_version = '0.14.0'
+mmseg_minimum_version = '0.14.1'
+mmseg_maximum_version = '1.0.0'
 mmseg_version = digit_version(mmseg.__version__)
 assert (mmseg_version >= digit_version(mmseg_minimum_version)
         and mmseg_version <= digit_version(mmseg_maximum_version)), \

mmdet3d/models/backbones/base_pointnet.py

@@ -1,23 +1,20 @@
+import warnings
 from abc import ABCMeta
-from mmcv.runner import load_checkpoint
-from torch import nn as nn
+from mmcv.runner import BaseModule
 
 
-class BasePointNet(nn.Module, metaclass=ABCMeta):
+class BasePointNet(BaseModule, metaclass=ABCMeta):
     """Base class for PointNet."""
 
-    def __init__(self):
-        super(BasePointNet, self).__init__()
+    def __init__(self, init_cfg=None, pretrained=None):
+        super(BasePointNet, self).__init__(init_cfg)
         self.fp16_enabled = False
-
-    def init_weights(self, pretrained=None):
-        """Initialize the weights of PointNet backbone."""
-        # Do not initialize the conv layers
-        # to follow the original implementation
+        assert not (init_cfg and pretrained), \
+            'init_cfg and pretrained cannot be setting at the same time'
         if isinstance(pretrained, str):
-            from mmdet3d.utils import get_root_logger
-            logger = get_root_logger()
-            load_checkpoint(self, pretrained, strict=False, logger=logger)
+            warnings.warn('DeprecationWarning: pretrained is a deprecated, '
+                          'please use "init_cfg" instead')
+            self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)
 
     @staticmethod
     def _split_point_feats(points):

mmdet3d/models/backbones/multi_backbone.py

@@ -1,14 +1,15 @@
 import copy
 import torch
+import warnings
 from mmcv.cnn import ConvModule
-from mmcv.runner import auto_fp16, load_checkpoint
+from mmcv.runner import BaseModule, auto_fp16
 from torch import nn as nn
 
 from mmdet.models import BACKBONES, build_backbone
 
 
 @BACKBONES.register_module()
-class MultiBackbone(nn.Module):
+class MultiBackbone(BaseModule):
     """MultiBackbone with different configs.
 
     Args:
@@ -31,8 +32,10 @@ def __init__(self,
                  norm_cfg=dict(type='BN1d', eps=1e-5, momentum=0.01),
                  act_cfg=dict(type='ReLU'),
                  suffixes=('net0', 'net1'),
+                 init_cfg=None,
+                 pretrained=None,
                  **kwargs):
-        super().__init__()
+        super().__init__(init_cfg=init_cfg)
         assert isinstance(backbones, dict) or isinstance(backbones, list)
         if isinstance(backbones, dict):
             backbones_list = []
@@ -77,14 +80,12 @@ def __init__(self,
                     bias=True,
                     inplace=True))
 
-    def init_weights(self, pretrained=None):
-        """Initialize the weights of PointNet++ backbone."""
-        # Do not initialize the conv layers
-        # to follow the original implementation
+        assert not (init_cfg and pretrained), \
+            'init_cfg and pretrained cannot be setting at the same time'
         if isinstance(pretrained, str):
-            from mmdet3d.utils import get_root_logger
-            logger = get_root_logger()
-            load_checkpoint(self, pretrained, strict=False, logger=logger)
+            warnings.warn('DeprecationWarning: pretrained is a deprecated, '
+                          'please use "init_cfg" instead')
+            self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)
 
     @auto_fp16()
     def forward(self, points):

mmdet3d/models/backbones/nostem_regnet.py

@@ -57,8 +57,8 @@ class NoStemRegNet(RegNet):
         (1, 1008, 1, 1)
     """
 
-    def __init__(self, arch, **kwargs):
-        super(NoStemRegNet, self).__init__(arch, **kwargs)
+    def __init__(self, arch, init_cfg=None, **kwargs):
+        super(NoStemRegNet, self).__init__(arch, init_cfg=init_cfg, **kwargs)
 
     def _make_stem_layer(self, in_channels, base_channels):
         """Override the original function that do not initialize a stem layer

mmdet3d/models/backbones/pointnet2_sa_msg.py

@@ -56,8 +56,9 @@ def __init__(self,
                      type='PointSAModuleMSG',
                      pool_mod='max',
                      use_xyz=True,
-                     normalize_xyz=False)):
-        super().__init__()
+                     normalize_xyz=False),
+                 init_cfg=None):
+        super().__init__(init_cfg=init_cfg)
         self.num_sa = len(sa_channels)
         self.out_indices = out_indices
         assert max(out_indices) < self.num_sa

mmdet3d/models/backbones/pointnet2_sa_ssg.py

@@ -43,8 +43,9 @@ def __init__(self,
                      type='PointSAModule',
                      pool_mod='max',
                      use_xyz=True,
-                     normalize_xyz=True)):
-        super().__init__()
+                     normalize_xyz=True),
+                 init_cfg=None):
+        super().__init__(init_cfg=init_cfg)
         self.num_sa = len(sa_channels)
         self.num_fp = len(fp_channels)
 

mmdet3d/models/backbones/second.py

@@ -1,12 +1,13 @@
+import warnings
 from mmcv.cnn import build_conv_layer, build_norm_layer
-from mmcv.runner import load_checkpoint
+from mmcv.runner import BaseModule
 from torch import nn as nn
 
 from mmdet.models import BACKBONES
 
 
 @BACKBONES.register_module()
-class SECOND(nn.Module):
+class SECOND(BaseModule):
     """Backbone network for SECOND/PointPillars/PartA2/MVXNet.
 
     Args:
@@ -24,8 +25,10 @@ def __init__(self,
                  layer_nums=[3, 5, 5],
                  layer_strides=[2, 2, 2],
                  norm_cfg=dict(type='BN', eps=1e-3, momentum=0.01),
-                 conv_cfg=dict(type='Conv2d', bias=False)):
-        super(SECOND, self).__init__()
+                 conv_cfg=dict(type='Conv2d', bias=False),
+                 init_cfg=None,
+                 pretrained=None):
+        super(SECOND, self).__init__(init_cfg=init_cfg)
         assert len(layer_strides) == len(layer_nums)
         assert len(out_channels) == len(layer_nums)
 
@@ -61,14 +64,14 @@ def __init__(self,
 
         self.blocks = nn.ModuleList(blocks)
 
-    def init_weights(self, pretrained=None):
-        """Initialize weights of the 2D backbone."""
-        # Do not initialize the conv layers
-        # to follow the original implementation
+        assert not (init_cfg and pretrained), \
+            'init_cfg and pretrained cannot be setting at the same time'
         if isinstance(pretrained, str):
-            from mmdet3d.utils import get_root_logger
-            logger = get_root_logger()
-            load_checkpoint(self, pretrained, strict=False, logger=logger)
+            warnings.warn('DeprecationWarning: pretrained is a deprecated, '
+                          'please use "init_cfg" instead')
+            self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)
+        else:
+            self.init_cfg = dict(type='Kaiming', layer='Conv2d')
 
     def forward(self, x):
         """Forward function.