add group normalization head to cascade mask rcnn (#3237)

* add cascade gn

* add cascade mask gn config

* update configs

PaddleCV/PaddleDetection/configs/cascade_mask_rcnn_r50_fpn_1x.yml

@@ -86,14 +86,14 @@ MaskAssigner:
   resolution: 28
 
 CascadeBBoxHead:
-  head: FC6FC7Head
+  head: CascadeTwoFCHead 
   nms:
     keep_top_k: 100
     nms_threshold: 0.5
     score_threshold: 0.05
 
-FC6FC7Head:
-  num_chan: 1024
+CascadeTwoFCHead:
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

PaddleCV/PaddleDetection/configs/cascade_rcnn_r50_fpn_1x.yml

@@ -77,14 +77,14 @@ CascadeBBoxAssigner:
   fg_fraction: 0.25
 
 CascadeBBoxHead:
-  head: FC6FC7Head
+  head: CascadeTwoFCHead
   nms:
     keep_top_k: 100
     nms_threshold: 0.5
     score_threshold: 0.05
 
-FC6FC7Head:
-  num_chan: 1024
+CascadeTwoFCHead:
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

PaddleCV/PaddleDetection/configs/dcn/cascade_rcnn_dcn_r101_vd_fpn_1x.yml

@@ -79,14 +79,14 @@ CascadeBBoxAssigner:
   fg_fraction: 0.25
 
 CascadeBBoxHead:
-  head: FC6FC7Head
+  head: CascadeTwoFCHead
   nms:
     keep_top_k: 100
     nms_threshold: 0.5
     score_threshold: 0.05
 
-FC6FC7Head:
-  num_chan: 1024
+CascadeTwoFCHead:
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

PaddleCV/PaddleDetection/configs/dcn/cascade_rcnn_dcn_r50_fpn_1x.yml

@@ -79,14 +79,14 @@ CascadeBBoxAssigner:
   fg_fraction: 0.25
 
 CascadeBBoxHead:
-  head: FC6FC7Head
+  head: CascadeTwoFCHead
   nms:
     keep_top_k: 100
     nms_threshold: 0.5
     score_threshold: 0.05
 
-FC6FC7Head:
-  num_chan: 1024
+CascadeTwoFCHead:
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

PaddleCV/PaddleDetection/configs/dcn/cascade_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml

@@ -81,14 +81,14 @@ CascadeBBoxAssigner:
   fg_fraction: 0.25
 
 CascadeBBoxHead:
-  head: FC6FC7Head
+  head: CascadeTwoFCHead
   nms:
     keep_top_k: 100
     nms_threshold: 0.5
     score_threshold: 0.05
 
-FC6FC7Head:
-  num_chan: 1024
+CascadeTwoFCHead:
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

PaddleCV/PaddleDetection/configs/gn/cascade_mask_rcnn_r50_fpn_gn_2x.yml

+architecture: CascadeMaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+max_iters: 180000
+snapshot_iter: 10000
+use_gpu: true
+log_smooth_window: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar
+weights: output/cascade_mask_rcnn_r50_fpn_gn_2x/model_final/
+metric: COCO
+num_classes: 81
+
+CascadeMaskRCNN:
+  backbone: ResNet
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: CascadeBBoxHead
+  bbox_assigner: CascadeBBoxAssigner
+  mask_head: MaskHead
+  mask_assigner: MaskAssigner
+
+ResNet:
+  depth: 50
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  norm_type: affine_channel
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+  norm_type: gn
+
+FPNRPNHead:
+  anchor_generator:
+    aspect_ratios: [0.5, 1.0, 2.0]
+    variance: [1.0, 1.0, 1.0, 1.0]
+  anchor_start_size: 32
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  rpn_target_assign:
+    rpn_batch_size_per_im: 256
+    rpn_fg_fraction: 0.5
+    rpn_negative_overlap: 0.3
+    rpn_positive_overlap: 0.7
+    rpn_straddle_thresh: 0.0
+  train_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    pre_nms_top_n: 2000
+    post_nms_top_n: 2000
+  test_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    pre_nms_top_n: 1000
+    post_nms_top_n: 1000
+
+FPNRoIAlign:
+  canconical_level: 4
+  canonical_size: 224
+  max_level: 5
+  min_level: 2
+  sampling_ratio: 2
+  box_resolution: 7
+  mask_resolution: 14
+
+MaskHead:
+  dilation: 1
+  conv_dim: 256
+  num_convs: 4
+  resolution: 28
+  norm_type: gn
+
+CascadeBBoxAssigner:
+  batch_size_per_im: 512
+  bbox_reg_weights: [10, 20, 30]
+  bg_thresh_hi: [0.5, 0.6, 0.7]
+  bg_thresh_lo: [0.0, 0.0, 0.0]
+  fg_fraction: 0.25
+  fg_thresh: [0.5, 0.6, 0.7]
+
+MaskAssigner:
+  resolution: 28
+
+CascadeBBoxHead:
+  head: CascadeXConvNormHead 
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+CascadeXConvNormHead:
+  norm_type: gn
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [120000, 160000]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2
+
+MaskRCNNTrainFeed:
+  batch_size: 2
+  dataset:
+    dataset_dir: dataset/coco
+    annotation: annotations/instances_train2017.json
+    image_dir: train2017
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+
+MaskRCNNEvalFeed:
+  batch_size: 1
+  dataset:
+    dataset_dir: dataset/coco
+    annotation: annotations/instances_val2017.json
+    image_dir: val2017
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+
+MaskRCNNTestFeed:
+  batch_size: 1
+  dataset:
+    annotation: dataset/coco/annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

PaddleCV/PaddleDetection/ppdet/modeling/ops.py

@@ -35,6 +35,7 @@ def ConvNorm(input,
              norm_type='affine_channel',
              norm_groups=32,
              dilation=1,
+             lr_scale=1,
              freeze_norm=False,
              act=None,
              norm_name=None,
@@ -51,18 +52,20 @@ def ConvNorm(input,
         groups=groups,
         act=None,
         param_attr=ParamAttr(
-            name=name + "_weights", initializer=initializer),
+            name=name + "_weights",
+            initializer=initializer,
+            learning_rate=lr_scale),
         bias_attr=False,
         name=name + '.conv2d.output.1')
 
     norm_lr = 0. if freeze_norm else 1.
     pattr = ParamAttr(
         name=norm_name + '_scale',
-        learning_rate=norm_lr,
+        learning_rate=norm_lr * lr_scale,
         regularizer=L2Decay(norm_decay))
     battr = ParamAttr(
         name=norm_name + '_offset',
-        learning_rate=norm_lr,
+        learning_rate=norm_lr * lr_scale,
         regularizer=L2Decay(norm_decay))
 
     if norm_type in ['bn', 'sync_bn']:

PaddleCV/PaddleDetection/ppdet/modeling/roi_heads/cascade_head.py

@@ -19,8 +19,10 @@ import paddle.fluid as fluid
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Normal, Xavier
 from paddle.fluid.regularizer import L2Decay
+from paddle.fluid.initializer import MSRA
 
 from ppdet.modeling.ops import MultiClassNMS
+from ppdet.modeling.ops import ConvNorm
 from ppdet.core.workspace import register
 
 __all__ = ['CascadeBBoxHead']
@@ -50,7 +52,7 @@ class CascadeBBoxHead(object):
     def get_output(self,
                    roi_feat,
                    cls_agnostic_bbox_reg=2,
-                   wb_scalar=2.0,
+                   wb_scalar=1.0,
                    name=''):
         """
         Get bbox head output.
@@ -77,7 +79,7 @@ class CascadeBBoxHead(object):
                                         learning_rate=wb_scalar),
                                     bias_attr=ParamAttr(
                                         name='cls_score%s_b' % name,
-                                        learning_rate=wb_scalar,
+                                        learning_rate=wb_scalar * 2,
                                         regularizer=L2Decay(0.)))
         bbox_pred = fluid.layers.fc(input=head_feat,
                                     size=4 * cls_agnostic_bbox_reg,
@@ -90,7 +92,7 @@ class CascadeBBoxHead(object):
                                         learning_rate=wb_scalar),
                                     bias_attr=ParamAttr(
                                         name='bbox_pred%s_b' % name,
-                                        learning_rate=wb_scalar,
+                                        learning_rate=wb_scalar * 2,
                                         regularizer=L2Decay(0.)))
         return cls_score, bbox_pred
 
@@ -177,7 +179,7 @@ class CascadeBBoxHead(object):
         for i in range(repreat_num):
             # cls score
             if i < 2:
-                cls_score = self._head_share(
+                cls_score, _ = self.get_output(
                     roi_feat_list[-1],  # roi_feat_3
                     name='_' + str(i + 1) if i > 0 else '')
             else:
@@ -216,66 +218,82 @@ class CascadeBBoxHead(object):
         pred_result = self.nms(bboxes=box_out, scores=boxes_cls_prob_mean)
         return {"bbox": pred_result}
 
-    def _head_share(self, roi_feat, wb_scalar=2.0, name=''):
-        # FC6 FC7
-        fan = roi_feat.shape[1] * roi_feat.shape[2] * roi_feat.shape[3]
-        fc6 = fluid.layers.fc(input=roi_feat,
-                              size=self.head.num_chan,
+
+@register
+class CascadeXConvNormHead(object):
+    """
+    RCNN head with serveral convolution layers
+
+    Args:
+        conv_num (int): num of convolution layers for the rcnn head
+        conv_dim (int): num of filters for the conv layers
+        mlp_dim (int): num of filters for the fc layers
+    """
+    __shared__ = ['norm_type', 'freeze_norm']
+
+    def __init__(self,
+                 num_conv=4,
+                 conv_dim=256,
+                 mlp_dim=1024,
+                 norm_type=None,
+                 freeze_norm=False):
+        super(CascadeXConvNormHead, self).__init__()
+        self.conv_dim = conv_dim
+        self.mlp_dim = mlp_dim
+        self.num_conv = num_conv
+        self.norm_type = norm_type
+        self.freeze_norm = freeze_norm
+
+    def __call__(self, roi_feat, wb_scalar=1.0, name=''):
+        conv = roi_feat
+        fan = self.conv_dim * 3 * 3
+        initializer = MSRA(uniform=False, fan_in=fan)
+        for i in range(self.num_conv):
+            name = 'bbox_head_conv' + str(i)
+            conv = ConvNorm(
+                conv,
+                self.conv_dim,
+                3,
+                act='relu',
+                initializer=initializer,
+                norm_type=self.norm_type,
+                freeze_norm=self.freeze_norm,
+                lr_scale=wb_scalar,
+                name=name,
+                norm_name=name)
+        fan = conv.shape[1] * conv.shape[2] * conv.shape[3]
+        head_heat = fluid.layers.fc(input=conv,
+                                    size=self.mlp_dim,
                                     act='relu',
                                     name='fc6' + name,
                                     param_attr=ParamAttr(
                                         name='fc6%s_w' % name,
                                         initializer=Xavier(fan_out=fan),
-                                  learning_rate=wb_scalar, ),
+                                        learning_rate=wb_scalar),
                                     bias_attr=ParamAttr(
                                         name='fc6%s_b' % name,
-                                  learning_rate=2.0,
-                                  regularizer=L2Decay(0.)))
-        fc7 = fluid.layers.fc(input=fc6,
-                              size=self.head.num_chan,
-                              act='relu',
-                              name='fc7' + name,
-                              param_attr=ParamAttr(
-                                  name='fc7%s_w' % name,
-                                  initializer=Xavier(),
-                                  learning_rate=wb_scalar, ),
-                              bias_attr=ParamAttr(
-                                  name='fc7%s_b' % name,
-                                  learning_rate=2.0,
-                                  regularizer=L2Decay(0.)))
-        cls_score = fluid.layers.fc(input=fc7,
-                                    size=self.num_classes,
-                                    act=None,
-                                    name='cls_score' + name,
-                                    param_attr=ParamAttr(
-                                        name='cls_score%s_w' % name,
-                                        initializer=Normal(
-                                            loc=0.0, scale=0.01),
-                                        learning_rate=wb_scalar, ),
-                                    bias_attr=ParamAttr(
-                                        name='cls_score%s_b' % name,
-                                        learning_rate=2.0,
-                                        regularizer=L2Decay(0.)))
-        return cls_score
+                                        regularizer=L2Decay(0.),
+                                        learning_rate=wb_scalar * 2))
+        return head_heat
 
 
 @register
-class FC6FC7Head(object):
+class CascadeTwoFCHead(object):
     """
-    Cascade RCNN head with two Fully Connected layers
+    RCNN head with serveral convolution layers
 
     Args:
-        num_chan (int): num of filters for the fc layers
+        mlp_dim (int): num of filters for the fc layers
     """
 
-    def __init__(self, num_chan):
-        super(FC6FC7Head, self).__init__()
-        self.num_chan = num_chan
+    def __init__(self, mlp_dim):
+        super(CascadeTwoFCHead, self).__init__()
+        self.mlp_dim = mlp_dim
 
     def __call__(self, roi_feat, wb_scalar=1.0, name=''):
         fan = roi_feat.shape[1] * roi_feat.shape[2] * roi_feat.shape[3]
         fc6 = fluid.layers.fc(input=roi_feat,
-                              size=self.num_chan,
+                              size=self.mlp_dim,
                               act='relu',
                               name='fc6' + name,
                               param_attr=ParamAttr(
@@ -284,10 +302,10 @@ class FC6FC7Head(object):
                                   learning_rate=wb_scalar),
                               bias_attr=ParamAttr(
                                   name='fc6%s_b' % name,
-                                  learning_rate=wb_scalar,
+                                  learning_rate=wb_scalar * 2,
                                   regularizer=L2Decay(0.)))
         head_feat = fluid.layers.fc(input=fc6,
-                                    size=self.num_chan,
+                                    size=self.mlp_dim,
                                     act='relu',
                                     name='fc7' + name,
                                     param_attr=ParamAttr(
@@ -296,6 +314,6 @@ class FC6FC7Head(object):
                                         learning_rate=wb_scalar),
                                     bias_attr=ParamAttr(
                                         name='fc7%s_b' % name,
-                                        learning_rate=wb_scalar,
+                                        learning_rate=wb_scalar * 2,
                                         regularizer=L2Decay(0.)))
         return head_feat