Merge branch 'master' into dev

README.md

@@ -35,11 +35,19 @@ Name | inf. time | box AP | download
 
 Name | inf. time | box AP | download
 --- |:---:|:---:|:---:
-[FCOS_RT_DLA_34_4x_shtw](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_shared_towers.yaml) | 52 FPS | 39.1 | [model](https://cloudstor.aarnet.edu.au/plus/s/4vc3XwQezyhNvnB/download)
-[FCOS_RT_DLA_34_4x](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn.yaml) | 46 FPS | 40.3 | [model](https://cloudstor.aarnet.edu.au/plus/s/zNPNyTkizaOOsUQ/download)
-[FCOS_RT_R_50_4x](configs/FCOS-Detection/FCOS_RT/MS_R_50_4x_syncbn.yaml) | 38 FPS | 40.2 | [model](https://cloudstor.aarnet.edu.au/plus/s/TlnlXUr6lNNSyoZ/download)
+[FCOS_RT_MS_DLA_34_4x_shtw](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_shared_towers.yaml) | 52 FPS | 39.1 | [model](https://cloudstor.aarnet.edu.au/plus/s/4vc3XwQezyhNvnB/download)
+[FCOS_RT_MS_DLA_34_4x](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn.yaml) | 46 FPS | 40.3 | [model](https://cloudstor.aarnet.edu.au/plus/s/zNPNyTkizaOOsUQ/download)
+[FCOS_RT_MS_R_50_4x](configs/FCOS-Detection/FCOS_RT/MS_R_50_4x_syncbn.yaml) | 38 FPS | 40.2 | [model](https://cloudstor.aarnet.edu.au/plus/s/TlnlXUr6lNNSyoZ/download)
 
-*Inference time is measured on a NVIDIA 1080Ti with batch size 1.*
+If you prefer BN in FCOS heads, please try the following models.
+
+Name | inf. time | box AP | download
+--- |:---:|:---:|:---:
+[FCOS_RT_MS_DLA_34_4x_shtw_bn](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_shared_towers_bn_head.yaml) | 52 FPS | 38.9 | [model](https://cloudstor.aarnet.edu.au/plus/s/rdmHHSs4oCg7l7U/download)
+[FCOS_RT_MS_DLA_34_4x_bn](configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_bn_head.yaml) | 48 FPS | 39.4 | [model](https://cloudstor.aarnet.edu.au/plus/s/T5httPVo1VndbD4/download)
+[FCOS_RT_MS_R_50_4x_bn](configs/FCOS-Detection/FCOS_RT/MS_R_50_4x_syncbn_bn_head.yaml) | 40 FPS | 39.3 | [model](https://cloudstor.aarnet.edu.au/plus/s/dHNUNs0YxVhZAmg/download)
+
+*Inference time is measured on a NVIDIA 1080Ti with batch size 1. Real-time models use shorter side 512 for inference.*
 
 ### COCO Instance Segmentation Baselines with [BlendMask](https://arxiv.org/abs/2001.00309)
 

adet/modeling/fcos/fcos.py

@@ -4,7 +4,7 @@
 from torch import nn
 from torch.nn import functional as F
 
-from detectron2.layers import ShapeSpec
+from detectron2.layers import ShapeSpec, NaiveSyncBatchNorm
 from detectron2.modeling.proposal_generator.build import PROPOSAL_GENERATOR_REGISTRY
 
 from adet.layers import DFConv2d, IOULoss, NaiveGroupNorm
@@ -25,6 +25,19 @@ def forward(self, input):
         return input * self.scale
 
 
+class ModuleListDial(nn.ModuleList):
+    def __init__(self, modules=None):
+        super(ModuleListDial, self).__init__(modules)
+        self.cur_position = 0
+
+    def forward(self, x):
+        result = self[self.cur_position](x)
+        self.cur_position += 1
+        if self.cur_position >= len(self):
+            self.cur_position = 0
+        return result
+
+
 @PROPOSAL_GENERATOR_REGISTRY.register()
 class FCOS(nn.Module):
     """
@@ -184,6 +197,7 @@ def __init__(self, cfg, input_shape: List[ShapeSpec]):
                         "share": (cfg.MODEL.FCOS.NUM_SHARE_CONVS,
                                   False)}
         norm = None if cfg.MODEL.FCOS.NORM == "none" else cfg.MODEL.FCOS.NORM
+        self.num_levels = len(input_shape)
 
         in_channels = [s.channels for s in input_shape]
         assert len(set(in_channels)) == 1, "Each level must have the same channel!"
@@ -206,6 +220,14 @@ def __init__(self, cfg, input_shape: List[ShapeSpec]):
                     tower.append(nn.GroupNorm(32, in_channels))
                 elif norm == "NaiveGN":
                     tower.append(NaiveGroupNorm(32, in_channels))
+                elif norm == "BN":
+                    tower.append(ModuleListDial([
+                        nn.BatchNorm2d(in_channels) for _ in range(self.num_levels)
+                    ]))
+                elif norm == "SyncBN":
+                    tower.append(ModuleListDial([
+                        NaiveSyncBatchNorm(in_channels) for _ in range(self.num_levels)
+                    ]))
                 tower.append(nn.ReLU())
             self.add_module('{}_tower'.format(head),
                             nn.Sequential(*tower))
@@ -225,7 +247,7 @@ def __init__(self, cfg, input_shape: List[ShapeSpec]):
         )
 
         if cfg.MODEL.FCOS.USE_SCALE:
-            self.scales = nn.ModuleList([Scale(init_value=1.0) for _ in self.fpn_strides])
+            self.scales = nn.ModuleList([Scale(init_value=1.0) for _ in range(self.num_levels)])
         else:
             self.scales = None
 

configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_bn_head.yaml

@@ -0,0 +1,26 @@
+_BASE_: "../Base-FCOS.yaml"
+INPUT:
+  MIN_SIZE_TRAIN: (256, 288, 320, 352, 384, 416, 448, 480, 512, 544, 576, 608)
+  MAX_SIZE_TRAIN: 900
+  MAX_SIZE_TEST: 736
+  MIN_SIZE_TEST: 512
+MODEL:
+  BACKBONE:
+    NAME: "build_fcos_dla_fpn_backbone"
+    FREEZE_AT: -1
+  WEIGHTS: "http://dl.yf.io/dla/models/imagenet/dla34-ba72cf86.pth"
+  DLA:
+    CONV_BODY: "DLA34"
+    NORM: "SyncBN"
+  FPN:
+    IN_FEATURES: ["level3", "level4", "level5"]
+  FCOS:
+    TOP_LEVELS: 0
+    SIZES_OF_INTEREST: [64, 128]
+    FPN_STRIDES: [8, 16, 32]
+    IN_FEATURES: ['p3', 'p4', 'p5']
+    NORM: "SyncBN"
+SOLVER:
+  STEPS: (300000, 340000)
+  MAX_ITER: 360000
+OUTPUT_DIR: "output/fcos/FCOS_RT_MS_DLA_34_4x_syncbn_bn_head"

configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_shared_towers.yaml

@@ -25,4 +25,4 @@ MODEL:
 SOLVER:
   STEPS: (300000, 340000)
   MAX_ITER: 360000
-OUTPUT_DIR: "output/fcos/MS_DLA_34_4x_syncbn_shared_towers"
+OUTPUT_DIR: "output/fcos/FCOS_RT_MS_DLA_34_4x_syncbn_shared_towers"

configs/FCOS-Detection/FCOS_RT/MS_DLA_34_4x_syncbn_shared_towers_bn_head.yaml

@@ -0,0 +1,29 @@
+_BASE_: "../Base-FCOS.yaml"
+INPUT:
+  MIN_SIZE_TRAIN: (256, 288, 320, 352, 384, 416, 448, 480, 512, 544, 576, 608)
+  MAX_SIZE_TRAIN: 900
+  MAX_SIZE_TEST: 736
+  MIN_SIZE_TEST: 512
+MODEL:
+  BACKBONE:
+    NAME: "build_fcos_dla_fpn_backbone"
+    FREEZE_AT: -1
+  WEIGHTS: "http://dl.yf.io/dla/models/imagenet/dla34-ba72cf86.pth"
+  DLA:
+    CONV_BODY: "DLA34"
+    NORM: "SyncBN"
+  FPN:
+    IN_FEATURES: ["level3", "level4", "level5"]
+  FCOS:
+    TOP_LEVELS: 0
+    SIZES_OF_INTEREST: [64, 128]
+    FPN_STRIDES: [8, 16, 32]
+    IN_FEATURES: ['p3', 'p4', 'p5']
+    NUM_SHARE_CONVS: 4
+    NUM_BOX_CONVS: 0
+    NUM_CLS_CONVS: 0
+    NORM: "SyncBN"
+SOLVER:
+  STEPS: (300000, 340000)
+  MAX_ITER: 360000
+OUTPUT_DIR: "output/fcos/FCOS_RT_MS_DLA_34_4x_syncbn_shared_towers_bn_head"

configs/FCOS-Detection/FCOS_RT/MS_R_50_4x_syncbn.yaml

@@ -17,4 +17,4 @@ MODEL:
 SOLVER:
   STEPS: (300000, 340000)
   MAX_ITER: 360000
-OUTPUT_DIR: "output/fcos/FCOS_RT_MS_R_50_4x"
+OUTPUT_DIR: "output/fcos/FCOS_RT_MS_R_50_4x_syncbn"

configs/FCOS-Detection/FCOS_RT/MS_R_50_4x_syncbn_bn_head.yaml

@@ -0,0 +1,21 @@
+_BASE_: "../Base-FCOS.yaml"
+INPUT:
+  MIN_SIZE_TRAIN: (256, 288, 320, 352, 384, 416, 448, 480, 512, 544, 576, 608)
+  MAX_SIZE_TRAIN: 900
+  MAX_SIZE_TEST: 736
+  MIN_SIZE_TEST: 512
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  RESNETS:
+    DEPTH: 50
+    NORM: "SyncBN"
+  FCOS:
+    TOP_LEVELS: 0
+    SIZES_OF_INTEREST: [64, 128]
+    FPN_STRIDES: [8, 16, 32]
+    IN_FEATURES: ['p3', 'p4', 'p5']
+    NORM: "SyncBN"
+SOLVER:
+  STEPS: (300000, 340000)
+  MAX_ITER: 360000
+OUTPUT_DIR: "output/fcos/FCOS_RT_MS_R_50_4x_syncbn_bn_head"