Refactor (training cfgs): change default training settings for better… (open-mmlab#2410)

* Refactor (training cfgs): change default training settings for better performance

* Change RoIAlign sample num & caffe img norm

* Refactor (LR): change 1/1000 to 0.001

Author: ZwwWayne

configs/_base_/default_runtime.py

@@ -7,7 +7,7 @@
         # dict(type='TensorboardLoggerHook')
     ])
 # yapf:enable
-dist_params = dict(backend='nccl')
+dist_params = dict(backend='nccl', port=29510)
 log_level = 'INFO'
 load_from = None
 resume_from = None

configs/_base_/models/cascade_mask_rcnn_r50_fpn.py

@@ -34,7 +34,7 @@
         stage_loss_weights=[1, 0.5, 0.25],
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=[
@@ -85,7 +85,7 @@
         ],
         mask_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         mask_head=dict(

configs/_base_/models/cascade_rcnn_r50_fpn.py

@@ -34,7 +34,7 @@
         stage_loss_weights=[1, 0.5, 0.25],
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=[

configs/_base_/models/fast_rcnn_r50_fpn.py

@@ -20,7 +20,7 @@
         type='StandardRoIHead',
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=dict(
@@ -34,7 +34,7 @@
             reg_class_agnostic=False,
             loss_cls=dict(
                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
-            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))))
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
 # model training and testing settings
 train_cfg = dict(
     rcnn=dict(

configs/_base_/models/faster_rcnn_r50_caffe_c4.py

@@ -25,7 +25,7 @@
         target_stds=[1.0, 1.0, 1.0, 1.0],
         loss_cls=dict(
             type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
-        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
     roi_head=dict(
         type='StandardRoIHead',
         shared_head=dict(
@@ -39,7 +39,7 @@
             norm_eval=True),
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=1024,
             featmap_strides=[16]),
         bbox_head=dict(
@@ -53,7 +53,7 @@
             reg_class_agnostic=False,
             loss_cls=dict(
                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
-            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))))
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
 # model training and testing settings
 train_cfg = dict(
     rpn=dict(

configs/_base_/models/faster_rcnn_r50_fpn.py

@@ -26,12 +26,12 @@
         target_stds=[1.0, 1.0, 1.0, 1.0],
         loss_cls=dict(
             type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
-        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
     roi_head=dict(
         type='StandardRoIHead',
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=dict(
@@ -45,7 +45,7 @@
             reg_class_agnostic=False,
             loss_cls=dict(
                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
-            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))))
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
 # model training and testing settings
 train_cfg = dict(
     rpn=dict(
@@ -68,8 +68,8 @@
     rpn_proposal=dict(
         nms_across_levels=False,
         nms_pre=2000,
-        nms_post=2000,
-        max_num=2000,
+        nms_post=1000,
+        max_num=1000,
         nms_thr=0.7,
         min_bbox_size=0),
     rcnn=dict(

configs/_base_/models/mask_rcnn_r50_caffe_c4.py

@@ -25,7 +25,7 @@
         target_stds=[1.0, 1.0, 1.0, 1.0],
         loss_cls=dict(
             type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
-        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
     roi_head=dict(
         type='StandardRoIHead',
         shared_head=dict(
@@ -39,7 +39,7 @@
             norm_eval=True),
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=1024,
             featmap_strides=[16]),
         bbox_head=dict(
@@ -53,7 +53,7 @@
             reg_class_agnostic=False,
             loss_cls=dict(
                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
-            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
         mask_roi_extractor=None,
         mask_head=dict(
             type='FCNMaskHead',

configs/_base_/models/mask_rcnn_r50_fpn.py

@@ -27,12 +27,12 @@
         target_stds=[1.0, 1.0, 1.0, 1.0],
         loss_cls=dict(
             type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
-        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
     roi_head=dict(
         type='StandardRoIHead',
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=dict(
@@ -46,10 +46,10 @@
             reg_class_agnostic=False,
             loss_cls=dict(
                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
-            loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
         mask_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         mask_head=dict(
@@ -82,8 +82,8 @@
     rpn_proposal=dict(
         nms_across_levels=False,
         nms_pre=2000,
-        nms_post=2000,
-        max_num=2000,
+        nms_post=1000,
+        max_num=1000,
         nms_thr=0.7,
         min_bbox_size=0),
     rcnn=dict(
@@ -92,7 +92,7 @@
             pos_iou_thr=0.5,
             neg_iou_thr=0.5,
             min_pos_iou=0.5,
-            match_low_quality=False,
+            match_low_quality=True,
             ignore_iof_thr=-1),
         sampler=dict(
             type='RandomSampler',

configs/_base_/schedules/schedule_1x.py

@@ -1,11 +1,11 @@
 # optimizer
 optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
-optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
+optimizer_config = dict(grad_clip=None)
 # learning policy
 lr_config = dict(
     policy='step',
     warmup='linear',
     warmup_iters=500,
-    warmup_ratio=1.0 / 3,
+    warmup_ratio=0.001,
     step=[8, 11])
 total_epochs = 12

configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py

@@ -21,13 +21,13 @@
 # optimizer
 # lr is set for a batch size of 8
 optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
-optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
+optimizer_config = dict(grad_clip=None)
 # learning policy
 lr_config = dict(
     policy='step',
     warmup='linear',
     warmup_iters=500,
-    warmup_ratio=1.0 / 3,
+    warmup_ratio=0.001,
     # [7] yields higher performance than [6]
     step=[7])
 total_epochs = 8  # actual epoch = 8 * 8 = 64

configs/cityscapes/mask_rcnn_r50_fpn_1x_cityscapes.py

@@ -28,13 +28,13 @@
 # optimizer
 # lr is set for a batch size of 8
 optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
-optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
+optimizer_config = dict(grad_clip=None)
 # learning policy
 lr_config = dict(
     policy='step',
     warmup='linear',
     warmup_iters=500,
-    warmup_ratio=1.0 / 3,
+    warmup_ratio=0.001,
     # [7] yields higher performance than [6]
     step=[7])
 total_epochs = 8  # actual epoch = 8 * 8 = 64

configs/faster_rcnn/faster_rcnn_r50_caffe_c4_1x_coco.py

@@ -5,7 +5,7 @@
 ]
 # use caffe img_norm
 img_norm_cfg = dict(
-    mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
 train_pipeline = [
     dict(type='LoadImageFromFile'),
     dict(type='LoadAnnotations', with_bbox=True),

configs/faster_rcnn/faster_rcnn_r50_caffe_fpn_1x_coco.py

@@ -5,7 +5,7 @@
         norm_cfg=dict(requires_grad=False), norm_eval=True, style='caffe'))
 # use caffe img_norm
 img_norm_cfg = dict(
-    mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
 train_pipeline = [
     dict(type='LoadImageFromFile'),
     dict(type='LoadAnnotations', with_bbox=True),

configs/grid_rcnn/grid_rcnn_r50_fpn_gn-head_2x_coco.py

@@ -35,7 +35,7 @@
         type='GridRoIHead',
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=dict(
@@ -50,7 +50,7 @@
             reg_class_agnostic=False),
         grid_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         grid_head=dict(

configs/htc/htc_r50_fpn_1x_coco.py

@@ -3,7 +3,7 @@
     roi_head=dict(
         semantic_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=256,
             featmap_strides=[8]),
         semantic_head=dict(

configs/htc/htc_without_semantic_r50_fpn_1x_coco.py

@@ -40,7 +40,7 @@
         stage_loss_weights=[1, 0.5, 0.25],
         bbox_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         bbox_head=[
@@ -91,7 +91,7 @@
         ],
         mask_roi_extractor=dict(
             type='SingleRoIExtractor',
-            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
+            roi_layer=dict(type='RoIAlign', out_size=14, sample_num=0),
             out_channels=256,
             featmap_strides=[4, 8, 16, 32]),
         mask_head=[

configs/mask_rcnn/mask_rcnn_r50_caffe_c4_1x_coco.py

@@ -5,7 +5,7 @@
 ]
 # use caffe img_norm
 img_norm_cfg = dict(
-    mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
 train_pipeline = [
     dict(type='LoadImageFromFile'),
     dict(type='LoadAnnotations', with_bbox=True, with_mask=True),

configs/mask_rcnn/mask_rcnn_r50_caffe_fpn_1x_coco.py

@@ -4,7 +4,7 @@
     backbone=dict(norm_cfg=dict(requires_grad=False), style='caffe'))
 # use caffe img_norm
 img_norm_cfg = dict(
-    mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
 train_pipeline = [
     dict(type='LoadImageFromFile'),
     dict(type='LoadAnnotations', with_bbox=True, with_mask=True),

configs/mask_rcnn/mask_rcnn_r50_caffe_fpn_detectron2-poly_1x_coco.py

@@ -0,0 +1,45 @@
+_base_ = './mask_rcnn_r50_fpn_1x_coco.py'
+model = dict(
+    pretrained='open-mmlab://resnet50_caffe',
+    backbone=dict(norm_cfg=dict(requires_grad=False), style='caffe'))
+# use caffe img_norm
+img_norm_cfg = dict(
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='LoadAnnotations',
+        with_bbox=True,
+        with_mask=True,
+        poly2mask=False),
+    dict(
+        type='Resize',
+        img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),
+                   (1333, 768), (1333, 800)],
+        multiscale_mode="value",
+        keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))