Add group norm (#3140)

* add a global norm_type flag

* add group norm on fpn

* add gn on box head

* add faster rcnn fpn 50 gn config

* add mask branch norm

* update configs

configs/dcn/faster_rcnn_dcn_r101_vd_fpn_1x.yml

@@ -86,7 +86,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/dcn/faster_rcnn_dcn_r50_fpn_1x.yml

@@ -85,7 +85,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/dcn/faster_rcnn_dcn_r50_vd_fpn_2x.yml

@@ -86,7 +86,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/dcn/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml

@@ -88,7 +88,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/dcn/mask_rcnn_dcn_r101_vd_fpn_1x.yml

@@ -71,7 +71,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -94,7 +94,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/dcn/mask_rcnn_dcn_r50_fpn_1x.yml

@@ -70,7 +70,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -93,7 +93,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/dcn/mask_rcnn_dcn_r50_vd_fpn_2x.yml

@@ -71,7 +71,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -94,7 +94,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/dcn/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml

@@ -73,7 +73,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -96,7 +96,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_r101_fpn_1x.yml

@@ -83,7 +83,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_r101_fpn_2x.yml

@@ -83,7 +83,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_r101_vd_fpn_1x.yml

@@ -84,7 +84,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_r101_vd_fpn_2x.yml

@@ -84,7 +84,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_r50_fpn_1x.yml

@@ -84,7 +84,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/faster_rcnn_r50_fpn_2x.yml

@@ -84,7 +84,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/faster_rcnn_r50_vd_fpn_2x.yml

@@ -84,7 +84,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.02

configs/faster_rcnn_se154_vd_fpn_s1x.yml

@@ -86,7 +86,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_x101_vd_64x4d_fpn_1x.yml

@@ -86,7 +86,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/faster_rcnn_x101_vd_64x4d_fpn_2x.yml

@@ -86,7 +86,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/gn/faster_rcnn_r50_fpn_gn_2x.yml

+architecture: FasterRCNN
+train_feed: FasterRCNNTrainFeed
+eval_feed: FasterRCNNEvalFeed
+test_feed: FasterRCNNTestFeed
+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
+metric: COCO
+weights: output/faster_rcnn_r50_fpn_gn/model_final
+num_classes: 81
+
+FasterRCNN:
+  backbone: ResNet
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: BBoxHead
+  bbox_assigner: BBoxAssigner
+
+ResNet:
+  depth: 50
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  norm_type: affine_channel
+
+FPN:
+  min_level: 2
+  max_level: 6
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+  norm_type: gn
+
+FPNRPNHead:
+  anchor_generator:
+    anchor_sizes: [32, 64, 128, 256, 512]
+    aspect_ratios: [0.5, 1.0, 2.0]
+    stride: [16.0, 16.0]
+    variance: [1.0, 1.0, 1.0, 1.0]
+  anchor_start_size: 32
+  min_level: 2
+  max_level: 6
+  num_chan: 256
+  rpn_target_assign:
+    rpn_batch_size_per_im: 256
+    rpn_fg_fraction: 0.5
+    rpn_positive_overlap: 0.7
+    rpn_negative_overlap: 0.3
+    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
+  min_level: 2
+  max_level: 5
+  box_resolution: 7
+  sampling_ratio: 2
+
+BBoxAssigner:
+  batch_size_per_im: 512
+  bbox_reg_weights: [0.1, 0.1, 0.2, 0.2]
+  bg_thresh_lo: 0.0
+  bg_thresh_hi: 0.5
+  fg_fraction: 0.25
+  fg_thresh: 0.5
+
+BBoxHead:
+  head: XConvNormHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+XConvNormHead:
+  norm_type: gn
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [120000, 160000]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2
+
+FasterRCNNTrainFeed:
+  batch_size: 2
+  dataset:
+    dataset_dir: dataset/coco
+    annotation: annotations/instances_train2017.json
+    image_dir: train2017
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  drop_last: false
+  num_workers: 16
+
+FasterRCNNEvalFeed:
+  batch_size: 1
+  dataset:
+    dataset_dir: dataset/coco
+    annotation: annotations/instances_val2017.json
+    image_dir: val2017
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+
+FasterRCNNTestFeed:
+  batch_size: 1
+  dataset:
+    annotation: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  drop_last: false
+  num_workers: 2

configs/gn/mask_rcnn_r50_fpn_gn_2x.yml

+architecture: MaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+max_iters: 360000
+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/mask_rcnn_r50_fpn_gn_2x/model_final/
+metric: COCO
+num_classes: 81
+
+MaskRCNN:
+  backbone: ResNet
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: BBoxHead
+  bbox_assigner: BBoxAssigner
+
+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
+
+BBoxAssigner:
+  batch_size_per_im: 512
+  bbox_reg_weights: [0.1, 0.1, 0.2, 0.2]
+  bg_thresh_hi: 0.5
+  bg_thresh_lo: 0.0
+  fg_fraction: 0.25
+  fg_thresh: 0.5
+
+MaskAssigner:
+  resolution: 28
+
+BBoxHead:
+  head: XConvNormHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+XConvNormHead:
+  norm_type: gn
+
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [240000, 320000]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2
+
+MaskRCNNTrainFeed:
+  batch_size: 1
+  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

configs/mask_rcnn_r101_fpn_1x.yml

@@ -68,7 +68,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -91,7 +91,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_r101_vd_fpn_1x.yml

@@ -69,7 +69,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -92,7 +92,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_r50_1x.yml

@@ -70,7 +70,7 @@ BBoxHead:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   resolution: 14
 
 BBoxAssigner:

configs/mask_rcnn_r50_2x.yml

@@ -71,7 +71,7 @@ BBoxHead:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   resolution: 14
 
 BBoxAssigner:

configs/mask_rcnn_r50_fpn_1x.yml

@@ -68,7 +68,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -91,7 +91,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_r50_fpn_2x.yml

@@ -68,7 +68,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -91,7 +91,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_r50_vd_fpn_2x.yml

@@ -69,7 +69,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -92,7 +92,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_se154_vd_fpn_s1x.yml

@@ -71,7 +71,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -94,7 +94,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_x101_vd_64x4d_fpn_1x.yml

@@ -71,7 +71,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -94,7 +94,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

configs/mask_rcnn_x101_vd_64x4d_fpn_2x.yml

@@ -71,7 +71,7 @@ FPNRoIAlign:
 
 MaskHead:
   dilation: 1
-  num_chan_reduced: 256
+  conv_dim: 256
   num_convs: 4
   resolution: 28
 
@@ -94,7 +94,7 @@ BBoxHead:
     score_threshold: 0.05
 
 TwoFCHead:
-  num_chan: 1024
+  mlp_dim: 1024
 
 LearningRate:
   base_lr: 0.01

ppdet/data/source/coco_loader.py

@@ -67,8 +67,8 @@ def load(anno_path, sample_num=-1, with_background=True):
     for img_id in img_ids:
         img_anno = coco.loadImgs(img_id)[0]
         im_fname = img_anno['file_name']
-        im_w = img_anno['width']
-        im_h = img_anno['height']
+        im_w = float(img_anno['width'])
+        im_h = float(img_anno['height'])
 
         ins_anno_ids = coco.getAnnIds(imgIds=img_id, iscrowd=False)
         instances = coco.loadAnns(ins_anno_ids)
@@ -85,8 +85,8 @@ def load(anno_path, sample_num=-1, with_background=True):
                 bboxes.append(inst)
             else:
                 logger.warn(
-                    'Found an invalid bbox in annotations: im_id: {}, area: {} x: {}, y: {}, h: {}, w: {}.'.
-                    format(img_id, float(inst['area']), x, y, box_w, box_h))
+                    'Found an invalid bbox in annotations: im_id: {}, area: {} x1: {}, y1: {}, x2: {}, y2: {}.'.
+                    format(img_id, float(inst['area']), x1, y1, x2, y2))
         num_bbox = len(bboxes)
 
         gt_bbox = np.zeros((num_bbox, 4), dtype=np.float32)

ppdet/modeling/backbones/darknet.py

@@ -36,6 +36,7 @@ class DarkNet(object):
         norm_type (str): normalization type, 'bn' and 'sync_bn' are supported
         norm_decay (float): weight decay for normalization layer weights
     """
+    __shared__ = ['norm_type']
 
     def __init__(self, depth=53, norm_type='bn', norm_decay=0.):
         assert depth in [53], "unsupported depth value"

ppdet/modeling/backbones/fpn.py

@@ -42,6 +42,7 @@ class FPN(object):
         has_extra_convs (bool): whether has extral convolutions in higher levels
         norm_type (str|None): normalization type, 'bn'/'sync_bn'/'affine_channel'
     """
+    __shared__ = ['norm_type', 'freeze_norm']
 
     def __init__(self,
                  num_chan=256,
@@ -49,7 +50,9 @@ class FPN(object):
                  max_level=6,
                  spatial_scale=[1. / 32., 1. / 16., 1. / 8., 1. / 4.],
                  has_extra_convs=False,
-                 norm_type=None):
+                 norm_type=None,
+                 freeze_norm=False):
+        self.freeze_norm = freeze_norm
         self.num_chan = num_chan
         self.min_level = min_level
         self.max_level = max_level
@@ -69,8 +72,9 @@ class FPN(object):
                 1,
                 initializer=initializer,
                 norm_type=self.norm_type,
+                freeze_norm=self.freeze_norm,
                 name=lateral_name,
-                bn_name=lateral_name)
+                norm_name=lateral_name)
         else:
             lateral = fluid.layers.conv2d(
                 body_input,
@@ -120,8 +124,9 @@ class FPN(object):
                 1,
                 initializer=initializer,
                 norm_type=self.norm_type,
+                freeze_norm=self.freeze_norm,
                 name=fpn_inner_name,
-                bn_name=fpn_inner_name)
+                norm_name=fpn_inner_name)
         else:
             self.fpn_inner_output[0] = fluid.layers.conv2d(
                 body_input,
@@ -155,8 +160,9 @@ class FPN(object):
                     3,
                     initializer=initializer,
                     norm_type=self.norm_type,
+                    freeze_norm=self.freeze_norm,
                     name=fpn_name,
-                    bn_name=fpn_name)
+                    norm_name=fpn_name)
             else:
                 fpn_output = fluid.layers.conv2d(
                     self.fpn_inner_output[i],

ppdet/modeling/backbones/mobilenet.py

@@ -37,6 +37,7 @@ class MobileNet(object):
         with_extra_blocks (bool): if extra blocks should be added
         extra_block_filters (list): number of filter for each extra block
     """
+    __shared__ = ['norm_type']
 
     def __init__(self,
                  norm_type='bn',

ppdet/modeling/backbones/resnet.py

@@ -47,6 +47,7 @@ class ResNet(object):
         feature_maps (list): index of stages whose feature maps are returned
         dcn_v2_stages (list): index of stages who select deformable conv v2
     """
+    __shared__ = ['norm_type', 'freeze_norm']
 
     def __init__(self,
                  depth=50,

ppdet/modeling/ops.py

@@ -16,7 +16,6 @@ from numbers import Integral
 
 from paddle import fluid
 from paddle.fluid.param_attr import ParamAttr
-from paddle.fluid.initializer import MSRA
 from paddle.fluid.regularizer import L2Decay
 from ppdet.core.workspace import register, serializable
 
@@ -34,9 +33,11 @@ def ConvNorm(input,
              groups=1,
              norm_decay=0.,
              norm_type='affine_channel',
+             norm_groups=32,
+             dilation=1,
              freeze_norm=False,
              act=None,
-             bn_name=None,
+             norm_name=None,
              initializer=None,
              name=None):
     fan = num_filters
@@ -45,7 +46,8 @@ def ConvNorm(input,
         num_filters=num_filters,
         filter_size=filter_size,
         stride=stride,
-        padding=(filter_size - 1) // 2,
+        padding=((filter_size - 1) // 2) * dilation,
+        dilation=dilation,
         groups=groups,
         act=None,
         param_attr=ParamAttr(
@@ -55,11 +57,11 @@ def ConvNorm(input,
 
     norm_lr = 0. if freeze_norm else 1.
     pattr = ParamAttr(
-        name=bn_name + '_scale',
+        name=norm_name + '_scale',
         learning_rate=norm_lr,
         regularizer=L2Decay(norm_decay))
     battr = ParamAttr(
-        name=bn_name + '_offset',
+        name=norm_name + '_offset',
         learning_rate=norm_lr,
         regularizer=L2Decay(norm_decay))
 
@@ -68,14 +70,24 @@ def ConvNorm(input,
         out = fluid.layers.batch_norm(
             input=conv,
             act=act,
-            name=bn_name + '.output.1',
+            name=norm_name + '.output.1',
             param_attr=pattr,
             bias_attr=battr,
-            moving_mean_name=bn_name + '_mean',
-            moving_variance_name=bn_name + '_variance',
+            moving_mean_name=norm_name + '_mean',
+            moving_variance_name=norm_name + '_variance',
             use_global_stats=global_stats)
         scale = fluid.framework._get_var(pattr.name)
         bias = fluid.framework._get_var(battr.name)
+    elif norm_type == 'gn':
+        out = fluid.layers.group_norm(
+            input=conv,
+            act=act,
+            name=norm_name + '.output.1',
+            groups=norm_groups,
+            param_attr=pattr,
+            bias_attr=battr)
+        scale = fluid.framework._get_var(pattr.name)
+        bias = fluid.framework._get_var(battr.name)
     elif norm_type == 'affine_channel':
         scale = fluid.layers.create_parameter(
             shape=[conv.shape[1]],

ppdet/modeling/roi_heads/bbox_head.py

@@ -22,11 +22,13 @@ from paddle import 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, serializable
 
-__all__ = ['BBoxHead', 'TwoFCHead']
+__all__ = ['BBoxHead', 'TwoFCHead', 'XConvNormHead']
 
 
 @register
@@ -47,23 +49,79 @@ class BoxCoder(object):
         self.axis = axis
 
 
+@register
+class XConvNormHead(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(XConvNormHead, 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):
+        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,
+                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)),
+                                    bias_attr=ParamAttr(
+                                        name='fc6%s_b' % name,
+                                        learning_rate=2,
+                                        regularizer=L2Decay(0.)))
+        return head_heat
+
+
 @register
 class TwoFCHead(object):
     """
     RCNN head with two Fully Connected 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=1024):
+    def __init__(self, mlp_dim=1024):
         super(TwoFCHead, self).__init__()
-        self.num_chan = num_chan
+        self.mlp_dim = mlp_dim
 
     def __call__(self, roi_feat):
         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',
                               param_attr=ParamAttr(
@@ -74,7 +132,7 @@ class TwoFCHead(object):
                                   learning_rate=2.,
                                   regularizer=L2Decay(0.)))
         head_feat = fluid.layers.fc(input=fc6,
-                                    size=self.num_chan,
+                                    size=self.mlp_dim,
                                     act='relu',
                                     name='fc7',
                                     param_attr=ParamAttr(
@@ -143,7 +201,8 @@ class BBoxHead(object):
         """
         head_feat = self.get_head_feat(roi_feat)
         # when ResNetC5 output a single feature map
-        if not isinstance(self.head, TwoFCHead):
+        if not isinstance(self.head, TwoFCHead) and not isinstance(
+                self.head, XConvNormHead):
             head_feat = fluid.layers.pool2d(
                 head_feat, pool_type='avg', global_pooling=True)
         cls_score = fluid.layers.fc(input=head_feat,

ppdet/modeling/roi_heads/mask_head.py

@@ -22,6 +22,7 @@ from paddle.fluid.initializer import MSRA
 from paddle.fluid.regularizer import L2Decay
 
 from ppdet.core.workspace import register
+from ppdet.modeling.ops import ConvNorm
 
 __all__ = ['MaskHead']
 
@@ -31,8 +32,8 @@ class MaskHead(object):
     """
     RCNN mask head
     Args:
-        num_convs (int): num of convolutions, 4 for FPN, 0 otherwise
-        num_chan_reduced (int): num of channels after first convolution
+        num_convs (int): num of convolutions, 4 for FPN, 1 otherwise
+        conv_dim (int): num of channels after first convolution
         resolution (int): size of the output mask
         dilation (int): dilation rate
         num_classes (int): number of output classes
@@ -42,42 +43,59 @@ class MaskHead(object):
 
     def __init__(self,
                  num_convs=0,
-                 num_chan_reduced=256,
+                 conv_dim=256,
                  resolution=14,
                  dilation=1,
-                 num_classes=81):
+                 num_classes=81,
+                 norm_type=None):
         super(MaskHead, self).__init__()
         self.num_convs = num_convs
-        self.num_chan_reduced = num_chan_reduced
+        self.conv_dim = conv_dim
         self.resolution = resolution
         self.dilation = dilation
         self.num_classes = num_classes
+        self.norm_type = norm_type
 
-    def _mask_conv_head(self, roi_feat, num_convs):
-        for i in range(num_convs):
-            layer_name = "mask_inter_feat_" + str(i + 1)
-            fan = self.num_chan_reduced * 3 * 3
-            roi_feat = fluid.layers.conv2d(
-                input=roi_feat,
-                num_filters=self.num_chan_reduced,
-                filter_size=3,
-                padding=1 * self.dilation,
-                act='relu',
-                stride=1,
-                dilation=self.dilation,
-                name=layer_name,
-                param_attr=ParamAttr(
-                    name=layer_name + '_w',
-                    initializer=MSRA(
-                        uniform=False, fan_in=fan)),
-                bias_attr=ParamAttr(
-                    name=layer_name + '_b',
-                    learning_rate=2.,
-                    regularizer=L2Decay(0.)))
+    def _mask_conv_head(self, roi_feat, num_convs, norm_type):
+        if norm_type == 'gn':
+            for i in range(num_convs):
+                layer_name = "mask_inter_feat_" + str(i + 1)
+                fan = self.conv_dim * 3 * 3
+                initializer = MSRA(uniform=False, fan_in=fan)
+                roi_feat = ConvNorm(
+                    roi_feat,
+                    self.conv_dim,
+                    3,
+                    act='relu',
+                    dilation=self.dilation,
+                    initializer=initializer,
+                    norm_type=self.norm_type,
+                    name=layer_name,
+                    norm_name=layer_name)
+        else:
+            for i in range(num_convs):
+                layer_name = "mask_inter_feat_" + str(i + 1)
+                fan = self.conv_dim * 3 * 3
+                initializer = MSRA(uniform=False, fan_in=fan)
+                roi_feat = fluid.layers.conv2d(
+                    input=roi_feat,
+                    num_filters=self.conv_dim,
+                    filter_size=3,
+                    padding=1 * self.dilation,
+                    act='relu',
+                    stride=1,
+                    dilation=self.dilation,
+                    name=layer_name,
+                    param_attr=ParamAttr(
+                        name=layer_name + '_w', initializer=initializer),
+                    bias_attr=ParamAttr(
+                        name=layer_name + '_b',
+                        learning_rate=2.,
+                        regularizer=L2Decay(0.)))
         fan = roi_feat.shape[1] * 2 * 2
         feat = fluid.layers.conv2d_transpose(
             input=roi_feat,
-            num_filters=self.num_chan_reduced,
+            num_filters=self.conv_dim,
             filter_size=2,
             stride=2,
             act='relu',
@@ -92,7 +110,8 @@ class MaskHead(object):
     def _get_output(self, roi_feat):
         class_num = self.num_classes
         # configure the conv number for FPN if necessary
-        head_feat = self._mask_conv_head(roi_feat, self.num_convs)
+        head_feat = self._mask_conv_head(roi_feat, self.num_convs,
+                                         self.norm_type)
         fan = class_num
         mask_logits = fluid.layers.conv2d(
             input=head_feat,