Add init_cfg for dense heads.

mmrotate/datasets/dota.py

@@ -202,7 +202,6 @@ def evaluate(self,
                 scale_ranges=scale_ranges,
                 iou_thr=iou_thr,
                 dataset=self.CLASSES,
-                version=self.version,
                 logger=logger,
                 nproc=nproc)
             eval_results['mAP'] = mean_ap

mmrotate/models/dense_heads/kfiou_odm_refine_head.py

@@ -1,6 +1,6 @@
 # Copyright (c) SJTU. All rights reserved.
 import torch.nn as nn
-from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.cnn import ConvModule
 from mmcv.runner import force_fp32
 
 from ..builder import ROTATED_HEADS
@@ -28,6 +28,7 @@ class KFIoUODMRefineHead(KFIoURRetinaHead):
         loss_bbox (dict): Config of localization loss.
         train_cfg (dict): Training config of anchor head.
         test_cfg (dict): Testing config of anchor head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
     """  # noqa: W605
 
     def __init__(self,
@@ -39,6 +40,15 @@ def __init__(self,
                  anchor_generator=dict(
                      type='PseudoAnchorGenerator',
                      strides=[8, 16, 32, 64, 128]),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='odm_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
         self.bboxes_as_anchors = None
         self.stacked_convs = stacked_convs
@@ -49,6 +59,7 @@ def __init__(self,
             in_channels,
             stacked_convs=2,
             anchor_generator=anchor_generator,
+            init_cfg=init_cfg,
             **kwargs)
 
     def _init_layers(self):
@@ -91,18 +102,6 @@ def _init_layers(self):
         self.odm_reg = nn.Conv2d(
             self.feat_channels, self.num_anchors * 5, 3, padding=1)
 
-    def init_weights(self):
-        """Initialize weights of the head."""
-
-        normal_init(self.or_conv, std=0.01)
-        for m in self.cls_convs:
-            normal_init(m.conv, std=0.01)
-        for m in self.reg_convs:
-            normal_init(m.conv, std=0.01)
-        bias_cls = bias_init_with_prob(0.01)
-        normal_init(self.odm_cls, std=0.01, bias=bias_cls)
-        normal_init(self.odm_reg, std=0.01)
-
     def forward_single(self, x):
         """Forward feature of a single scale level.
 

mmrotate/models/dense_heads/kfiou_rotate_retina_head.py

@@ -23,6 +23,7 @@ class KFIoURRetinaHead(RotatedRetinaHead):
         loss_bbox (dict): Config of localization loss.
         train_cfg (dict): Training config of anchor head.
         test_cfg (dict): Testing config of anchor head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
     """  # noqa: W605
 
     def __init__(self,
@@ -37,6 +38,15 @@ def __init__(self,
                      scales_per_octave=3,
                      ratios=[0.5, 1.0, 2.0],
                      strides=[8, 16, 32, 64, 128]),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='retina_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
         self.bboxes_as_anchors = None
         super(KFIoURRetinaHead, self).__init__(
@@ -46,6 +56,7 @@ def __init__(self,
             conv_cfg=conv_cfg,
             norm_cfg=norm_cfg,
             anchor_generator=anchor_generator,
+            init_cfg=init_cfg,
             **kwargs)
 
     def loss_single(self, cls_score, bbox_pred, anchors, labels, label_weights,

mmrotate/models/dense_heads/kfiou_rotate_retina_refine_head.py

@@ -26,6 +26,7 @@ class KFIoURRetinaRefineHead(KFIoURRetinaHead):
         loss_bbox (dict): Config of localization loss.
         train_cfg (dict): Training config of anchor head.
         test_cfg (dict): Testing config of anchor head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
     """  # noqa: W605
 
     def __init__(self,
@@ -41,6 +42,15 @@ def __init__(self,
                      type='DeltaXYWHABBoxCoder',
                      target_means=(.0, .0, .0, .0, .0),
                      target_stds=(1.0, 1.0, 1.0, 1.0, 1.0)),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='retina_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
 
         self.bboxes_as_anchors = None
@@ -52,6 +62,7 @@ def __init__(self,
             norm_cfg=norm_cfg,
             anchor_generator=anchor_generator,
             bbox_coder=bbox_coder,
+            init_cfg=init_cfg,
             **kwargs)
 
     @force_fp32(apply_to=('cls_scores', 'bbox_preds'))

mmrotate/models/dense_heads/odm_refine_head.py

@@ -1,6 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import torch.nn as nn
-from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.cnn import ConvModule
 from mmcv.runner import force_fp32
 
 from ..builder import ROTATED_HEADS
@@ -28,6 +28,7 @@ class ODMRefineHead(RotatedRetinaHead):
         loss_bbox (dict): Config of localization loss.
         train_cfg (dict): Training config of anchor head.
         test_cfg (dict): Testing config of anchor head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
     """  # noqa: W605
 
     def __init__(self,
@@ -39,6 +40,15 @@ def __init__(self,
                  anchor_generator=dict(
                      type='PseudoAnchorGenerator',
                      strides=[8, 16, 32, 64, 128]),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='odm_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
         self.bboxes_as_anchors = None
         self.stacked_convs = stacked_convs
@@ -49,6 +59,7 @@ def __init__(self,
             in_channels,
             stacked_convs=2,
             anchor_generator=anchor_generator,
+            init_cfg=init_cfg,
             **kwargs)
 
     def _init_layers(self):
@@ -91,18 +102,6 @@ def _init_layers(self):
         self.odm_reg = nn.Conv2d(
             self.feat_channels, self.num_anchors * 5, 3, padding=1)
 
-    def init_weights(self):
-        """Initialize weights of the head."""
-
-        normal_init(self.or_conv, std=0.01)
-        for m in self.cls_convs:
-            normal_init(m.conv, std=0.01)
-        for m in self.reg_convs:
-            normal_init(m.conv, std=0.01)
-        bias_cls = bias_init_with_prob(0.01)
-        normal_init(self.odm_cls, std=0.01, bias=bias_cls)
-        normal_init(self.odm_reg, std=0.01)
-
     def forward_single(self, x):
         """Forward feature of a single scale level.
 

mmrotate/models/dense_heads/rotated_anchor_head.py

@@ -3,7 +3,6 @@
 
 import torch
 import torch.nn as nn
-from mmcv.cnn import normal_init
 from mmcv.runner import force_fp32
 from mmdet.core import images_to_levels, multi_apply, unmap
 from mmdet.models.dense_heads.base_dense_head import BaseDenseHead
@@ -33,6 +32,7 @@ class RotatedAnchorHead(BaseDenseHead):
         loss_bbox (dict): Config of localization loss.
         train_cfg (dict): Training config of anchor head.
         test_cfg (dict): Testing config of anchor head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
     """  # noqa: W605
 
     def __init__(self,
@@ -59,8 +59,9 @@ def __init__(self,
                      loss_weight=1.0),
                  loss_bbox=dict(type='L1Loss', loss_weight=1.0),
                  train_cfg=None,
-                 test_cfg=None):
-        super(RotatedAnchorHead, self).__init__()
+                 test_cfg=None,
+                 init_cfg=dict(type='Normal', layer='Conv2d', std=0.01)):
+        super(RotatedAnchorHead, self).__init__(init_cfg)
         self.in_channels = in_channels
         self.num_classes = num_classes
         self.feat_channels = feat_channels
@@ -105,11 +106,6 @@ def _init_layers(self):
                                   self.num_anchors * self.cls_out_channels, 1)
         self.conv_reg = nn.Conv2d(self.in_channels, self.num_anchors * 5, 1)
 
-    def init_weights(self):
-        """Initialize weights of the head."""
-        normal_init(self.conv_cls, std=0.01)
-        normal_init(self.conv_reg, std=0.01)
-
     def forward_single(self, x):
         """Forward feature of a single scale level.
 

mmrotate/models/dense_heads/rotated_retina_head.py

@@ -1,6 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import torch.nn as nn
-from mmcv.cnn import ConvModule, bias_init_with_prob, normal_init
+from mmcv.cnn import ConvModule
 from mmcv.runner import force_fp32
 
 from ..builder import ROTATED_HEADS
@@ -39,6 +39,15 @@ def __init__(self,
                      scales_per_octave=3,
                      ratios=[0.5, 1.0, 2.0],
                      strides=[8, 16, 32, 64, 128]),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='retina_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
         self.stacked_convs = stacked_convs
         self.conv_cfg = conv_cfg
@@ -47,6 +56,7 @@ def __init__(self,
             num_classes,
             in_channels,
             anchor_generator=anchor_generator,
+            init_cfg=init_cfg,
             **kwargs)
 
     def _init_layers(self):
@@ -82,16 +92,6 @@ def _init_layers(self):
         self.retina_reg = nn.Conv2d(
             self.feat_channels, self.num_anchors * 5, 3, padding=1)
 
-    def init_weights(self):
-        """Initialize weights of the head."""
-        for m in self.cls_convs:
-            normal_init(m.conv, std=0.01)
-        for m in self.reg_convs:
-            normal_init(m.conv, std=0.01)
-        bias_cls = bias_init_with_prob(0.01)
-        normal_init(self.retina_cls, std=0.01, bias=bias_cls)
-        normal_init(self.retina_reg, std=0.01)
-
     def forward_single(self, x):
         """Forward feature of a single scale level.
 

mmrotate/models/dense_heads/rotated_retina_refine_head.py

@@ -41,6 +41,15 @@ def __init__(self,
                      type='DeltaXYWHABBoxCoder',
                      target_means=(.0, .0, .0, .0, .0),
                      target_stds=(1.0, 1.0, 1.0, 1.0, 1.0)),
+                 init_cfg=dict(
+                     type='Normal',
+                     layer='Conv2d',
+                     std=0.01,
+                     override=dict(
+                         type='Normal',
+                         name='retina_cls',
+                         std=0.01,
+                         bias_prob=0.01)),
                  **kwargs):
 
         self.bboxes_as_anchors = None
@@ -52,6 +61,7 @@ def __init__(self,
             norm_cfg=norm_cfg,
             anchor_generator=anchor_generator,
             bbox_coder=bbox_coder,
+            init_cfg=init_cfg,
             **kwargs)
 
     @force_fp32(apply_to=('cls_scores', 'bbox_preds'))

mmrotate/models/detectors/r3det.py

@@ -1,5 +1,7 @@
 # Copyright (c) SJTU. All rights reserved.
-import torch.nn as nn
+import warnings
+
+from mmcv.runner import ModuleList
 
 from mmrotate.core import rbbox2result
 from ..builder import ROTATED_DETECTORS, build_backbone, build_head, build_neck
@@ -20,10 +22,13 @@ def __init__(self,
                  refine_heads=None,
                  train_cfg=None,
                  test_cfg=None,
-                 pretrained=None):
-        super(R3Det, self).__init__()
-
-        backbone.pretrained = pretrained
+                 pretrained=None,
+                 init_cfg=None):
+        super(R3Det, self).__init__(init_cfg)
+        if pretrained:
+            warnings.warn('DeprecationWarning: pretrained is deprecated, '
+                          'please use "init_cfg" instead')
+            backbone.pretrained = pretrained
         self.backbone = build_backbone(backbone)
         self.num_refine_stages = num_refine_stages
         if neck is not None:
@@ -32,19 +37,15 @@ def __init__(self,
             bbox_head.update(train_cfg=train_cfg['s0'])
         bbox_head.update(test_cfg=test_cfg)
         self.bbox_head = build_head(bbox_head)
-        self.bbox_head.init_weights()
-        self.feat_refine_module = nn.ModuleList()
-        self.refine_head = nn.ModuleList()
+        self.feat_refine_module = ModuleList()
+        self.refine_head = ModuleList()
         for i, (frm_cfg,
                 refine_head) in enumerate(zip(frm_cfgs, refine_heads)):
             self.feat_refine_module.append(FeatureRefineModule(**frm_cfg))
             if train_cfg is not None:
                 refine_head.update(train_cfg=train_cfg['sr'][i])
             refine_head.update(test_cfg=test_cfg)
             self.refine_head.append(build_head(refine_head))
-        for i in range(self.num_refine_stages):
-            self.feat_refine_module[i].init_weights()
-            self.refine_head[i].init_weights()
         self.train_cfg = train_cfg
         self.test_cfg = test_cfg
 

mmrotate/models/detectors/rotated_reppoints.py

@@ -16,4 +16,3 @@ def __init__(self,
                  pretrained=None):
         super(RotatedRepPoints, self).__init__(backbone, neck, bbox_head,
                                                train_cfg, test_cfg, pretrained)
-        self.bbox_head.init_weights()

mmrotate/models/detectors/rotated_retinanet.py

@@ -16,7 +16,8 @@ def __init__(self,
                  bbox_head,
                  train_cfg=None,
                  test_cfg=None,
-                 pretrained=None):
-        super(RotatedRetinaNet, self).__init__(backbone, neck, bbox_head,
-                                               train_cfg, test_cfg, pretrained)
-        self.bbox_head.init_weights()
+                 pretrained=None,
+                 init_cfg=None):
+        super(RotatedRetinaNet,
+              self).__init__(backbone, neck, bbox_head, train_cfg, test_cfg,
+                             pretrained, init_cfg)

mmrotate/models/detectors/s2anet.py

@@ -28,7 +28,6 @@ def __init__(self,
             fam_head.update(train_cfg=train_cfg['fam_cfg'])
         fam_head.update(test_cfg=test_cfg)
         self.fam_head = build_head(fam_head)
-        self.fam_head.init_weights()
 
         self.align_conv_type = align_cfgs['type']
         self.align_conv_size = align_cfgs['kernel_size']
@@ -44,7 +43,6 @@ def __init__(self,
             odm_head.update(train_cfg=train_cfg['odm_cfg'])
         odm_head.update(test_cfg=test_cfg)
         self.odm_head = build_head(odm_head)
-        self.odm_head.init_weights()
 
         self.train_cfg = train_cfg
         self.test_cfg = test_cfg

requirements/build.txt

@@ -1,3 +1,3 @@
-# These must be installed before building mmdetection
+# These must be installed before building mmrotate
 cython
 numpy