[Feat]: Support DAB-DETR

configs/_base_/datasets/coco_detection.py

@@ -1,6 +1,8 @@
 # dataset settings
 dataset_type = 'CocoDataset'
-data_root = 'data/coco/'
+# data_root = '/home/ps/ssd/big_data/xqz/mmdetection/data/coco/'
+data_root = '/home/ps/ssd/big_data/xqz/mmdetection/data/coco_debug/'
+# data_root = 'E:/mmdet-dev/debug_local/coco/'  # for local debug only
 img_norm_cfg = dict(
     mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
 train_pipeline = [

configs/dab_detr/dab_detr_r50_8x2_50e_coco.py

@@ -0,0 +1,187 @@
+_base_ = [
+    '../_base_/datasets/coco_detection.py', '../_base_/default_runtime.py'
+]
+model = dict(
+    type='DABDETR',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(3, ),
+        frozen_stages=1,
+        norm_cfg=dict(
+            type='FrozenBN2d',
+            requires_grad=False),  # register torch.ops.FrozenBatchNorm2d
+        norm_eval=True,
+        style='pytorch',
+        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
+    bbox_head=dict(
+        type='DABDETRHead',
+        num_query=300,
+        num_classes=80,  # default 80, 91 to align with official repo
+        in_channels=2048,
+        iter_update=True,
+        random_refpoints_xy=False,
+        transformer=dict(
+            type='DabDetrTransformer',
+            num_patterns=0,
+            encoder=dict(
+                type='DabDetrTransformerEncoder',
+                num_layers=6,
+                transformerlayers=dict(
+                    type='BaseTransformerLayer',
+                    attn_cfgs=[
+                        dict(
+                            type='MultiheadAttention',
+                            embed_dims=256,
+                            num_heads=8,
+                            dropout=0.)  # default 0.
+                    ],
+                    feedforward_channels=2048,
+                    ffn_dropout=0.,  # default 0.
+                    act_cfg=dict(type='PReLU'),  # default PReLU
+                    operation_order=('self_attn', 'norm', 'ffn', 'norm'))),
+            decoder=dict(
+                type='DabDetrTransformerDecoder',
+                query_dim=4,
+                num_layers=6,
+                query_scale_type='cond_elewise',
+                modulate_hw_attn=True,
+                bbox_embed_diff_each_layer=False,
+                transformerlayers=dict(
+                    type='DabDetrTransformerDecoderLayer',
+                    attn_cfgs=dict(
+                        type='ConditionalAttention',
+                        embed_dims=256,
+                        num_heads=8,
+                        dropout=0.),  # default 0.
+                    feedforward_channels=2048,
+                    ffn_dropout=0.,  # default 0.
+                    sa_dropout=0.,  # default 0.
+                    ca_dropout=0.,  # default 0.
+                    keep_query_pos=False,
+                    act_cfg=dict(type='PReLU'),  # default PReLU
+                    operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
+                                     'ffn', 'norm')),
+            )),
+        positional_encoding=dict(
+            type='SinePositionalEncodingHW',
+            num_feats=128,
+            temperatureH=20,
+            temperatureW=20,
+            normalize=True),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0
+        ),  # weights of cls cost and cls loss can't be different for DETRHead
+        loss_bbox=dict(type='L1Loss', loss_weight=5.0),
+        loss_iou=dict(type='GIoULoss', loss_weight=2.0)),
+    # training and testing settings
+    train_cfg=dict(
+        assigner=dict(
+            type='HungarianAssigner',
+            cls_cost=dict(type='FocalLossCost', weight=2.,
+                          eps=1e-8),  # default eps=1e-8
+            reg_cost=dict(type='BBoxL1Cost', weight=5.0, box_format='xywh'),
+            iou_cost=dict(type='IoUCost', iou_mode='giou', weight=2.0))),
+    test_cfg=dict(max_per_img=300))  # default 300
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375],
+    to_rgb=True)  # default True, False for pillow
+# img_norm_cfg = dict(
+#     mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], to_rgb=True)
+# train_pipeline, NOTE the img_scale and the Pad's size_divisor is different
+# from the default setting in mmdet.
+train_pipeline = [
+    dict(type='LoadImageFromFile', channel_order='bgr'
+         ),  # default 'bgr', 'rgb' to align with official repo
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='AutoAugment',
+        policies=[[
+            dict(
+                type='Resize',
+                img_scale=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                           (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                           (736, 1333), (768, 1333), (800, 1333)],
+                multiscale_mode='value',
+                keep_ratio=True)
+        ],
+                  [
+                      dict(
+                          type='Resize',
+                          img_scale=[(400, 1333), (500, 1333), (600, 1333)],
+                          multiscale_mode='value',
+                          keep_ratio=True),
+                      dict(
+                          type='RandomCrop',
+                          crop_type='absolute_range',
+                          crop_size=(384, 600),
+                          allow_negative_crop=True),
+                      dict(
+                          type='Resize',
+                          img_scale=[(480, 1333), (512, 1333), (544, 1333),
+                                     (576, 1333), (608, 1333), (640, 1333),
+                                     (672, 1333), (704, 1333), (736, 1333),
+                                     (768, 1333), (800, 1333)],
+                          multiscale_mode='value',
+                          override=True,
+                          keep_ratio=True)
+                  ]]),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=1),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
+]
+# test_pipeline, NOTE the Pad's size_divisor is different from the default
+# setting (size_divisor=32). While there is little effect on the performance
+# whether we use the default setting or use size_divisor=1.
+test_pipeline = [
+    dict(type='LoadImageFromFile',
+         channel_order='bgr'),  # default 'bgr', 'rgb' for pillow
+    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=1),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=1,  # default 2 * 8, use 4 * 4, used in train dataset only
+    workers_per_gpu=0,
+    # default 2, 0 for debug, used in train & val & test datasets
+    train=dict(
+        continuous_categories=True,
+        # default True, False to align with official repo
+        pipeline=train_pipeline),
+    val=dict(continuous_categories=True,
+             pipeline=test_pipeline),  # add 'samples_per_gpu' here
+    test=dict(continuous_categories=True,
+              pipeline=test_pipeline))  # add 'samples_per_gpu' here
+evaluation = dict(interval=1, metric='bbox', save_best='bbox_mAP')
+checkpoint_config = dict(interval=10)
+# optimizer
+optimizer = dict(
+    type='AdamW',
+    lr=0.0001,
+    weight_decay=0.0001,
+    paramwise_cfg=dict(
+        custom_keys={'backbone': dict(lr_mult=0.1, decay_mult=1.0)}))
+optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2))
+# learning policy
+lr_config = dict(
+    policy='step', step=[1])  # default 40 epochs, 1 epoch for debug train
+runner = dict(
+    type='EpochBasedRunner',
+    max_epochs=1)  # default 50 epochs, 1 epoch for debug train
+auto_scale_lr = dict(enable=True, base_batch_size=16)

configs/dab_detr/dab_detr_r50_dc5_8x2_50e_coco.py

@@ -0,0 +1,181 @@
+_base_ = [
+    '../_base_/datasets/coco_detection.py', '../_base_/default_runtime.py'
+]
+model = dict(
+    type='DABDETR',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(3, ),
+        strides=(1, 2, 2, 1),
+        dilations=(1, 1, 1, 2),  # dc5
+        frozen_stages=1,
+        norm_cfg=dict(
+            type='FrozenBN2d',
+            requires_grad=False),  # register torch.ops.FrozenBatchNorm2d
+        norm_eval=True,
+        style='pytorch',
+        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
+    bbox_head=dict(
+        type='DABDETRHead',
+        num_query=300,
+        num_classes=80,  # default 80, 91 to align with official repo / debug
+        in_channels=2048,
+        iter_update=True,
+        random_refpoints_xy=False,
+        transformer=dict(
+            type='DabDetrTransformer',
+            num_patterns=0,
+            encoder=dict(
+                type='DabDetrTransformerEncoder',
+                num_layers=6,
+                transformerlayers=dict(
+                    type='BaseTransformerLayer',
+                    attn_cfgs=[
+                        dict(
+                            type='MultiheadAttention',
+                            embed_dims=256,
+                            num_heads=8,
+                            dropout=0.1)
+                    ],
+                    feedforward_channels=2048,
+                    ffn_dropout=0.1,
+                    act_cfg=dict(type='PReLU'),  # default PReLU
+                    operation_order=('self_attn', 'norm', 'ffn', 'norm'))),
+            decoder=dict(
+                type='DabDetrTransformerDecoder',
+                query_dim=4,
+                num_layers=6,
+                query_scale_type='cond_elewise',
+                modulate_hw_attn=True,
+                bbox_embed_diff_each_layer=False,
+                transformerlayers=dict(
+                    type='DabDetrTransformerDecoderLayer',
+                    attn_cfgs=dict(
+                        type='ConditionalAttention',
+                        embed_dims=256,
+                        num_heads=8,
+                        dropout=0.1),
+                    feedforward_channels=2048,
+                    ffn_dropout=0.1,
+                    sa_dropout=0.1,
+                    ca_dropout=0.1,
+                    keep_query_pos=False,
+                    act_cfg=dict(type='PReLU'),  # default PReLU
+                    operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
+                                     'ffn', 'norm')),
+            )),
+        positional_encoding=dict(
+            type='SinePositionalEncodingHW',
+            num_feats=128,
+            temperatureH=20,
+            temperatureW=20,
+            normalize=True),
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0
+        ),  # weights of cls cost and cls loss can't be different for DETRHead
+        loss_bbox=dict(type='L1Loss', loss_weight=5.0),
+        loss_iou=dict(type='GIoULoss', loss_weight=2.0)),
+    # training and testing settings
+    train_cfg=dict(
+        assigner=dict(
+            type='HungarianAssigner',
+            cls_cost=dict(type='FocalLossCost', weight=2.,
+                          eps=1e-8),  # default eps=1e-8
+            reg_cost=dict(type='BBoxL1Cost', weight=5.0, box_format='xywh'),
+            iou_cost=dict(type='IoUCost', iou_mode='giou', weight=2.0))),
+    test_cfg=dict(max_per_img=100))
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+# train_pipeline, NOTE the img_scale and the Pad's size_divisor is different
+# from the default setting in mmdet.
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(
+        type='AutoAugment',
+        policies=[[
+            dict(
+                type='Resize',
+                img_scale=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                           (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                           (736, 1333), (768, 1333), (800, 1333)],
+                multiscale_mode='value',
+                keep_ratio=True)
+        ],
+                  [
+                      dict(
+                          type='Resize',
+                          img_scale=[(400, 1333), (500, 1333), (600, 1333)],
+                          multiscale_mode='value',
+                          keep_ratio=True),
+                      dict(
+                          type='RandomCrop',
+                          crop_type='absolute_range',
+                          crop_size=(384, 600),
+                          allow_negative_crop=True),
+                      dict(
+                          type='Resize',
+                          img_scale=[(480, 1333), (512, 1333), (544, 1333),
+                                     (576, 1333), (608, 1333), (640, 1333),
+                                     (672, 1333), (704, 1333), (736, 1333),
+                                     (768, 1333), (800, 1333)],
+                          multiscale_mode='value',
+                          override=True,
+                          keep_ratio=True)
+                  ]]),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=1),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
+]
+# test_pipeline, NOTE the Pad's size_divisor is different from the default
+# setting (size_divisor=32). While there is little effect on the performance
+# whether we use the default setting or use size_divisor=1.
+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=1),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+data = dict(
+    samples_per_gpu=2,  # default batch size is 2 * 8, use 2 * 5
+    workers_per_gpu=2,  # default 2, 0 for debug
+    train=dict(
+        continuous_categories=True,
+        # default True, False to align with official repo
+        pipeline=train_pipeline),
+    val=dict(continuous_categories=True, pipeline=test_pipeline),
+    test=dict(continuous_categories=True, pipeline=test_pipeline))
+evaluation = dict(interval=1, metric='bbox', save_best='bbox_mAP')
+checkpoint_config = dict(interval=10)
+# optimizer
+optimizer = dict(
+    type='AdamW',
+    lr=0.0001,
+    weight_decay=0.0001,
+    paramwise_cfg=dict(
+        custom_keys={'backbone': dict(lr_mult=0.1, decay_mult=1.0)}))
+optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2))
+# learning policy
+lr_config = dict(
+    policy='step', step=[40])  # default 40 epochs, 1 epoch for debug train
+runner = dict(
+    type='EpochBasedRunner',
+    max_epochs=50)  # default 50 epochs, 1 epoch for debug train
+auto_scale_lr = dict(enable=True, base_batch_size=16)