[PaddleDetection] add deformable models (#3016)

* add deformable models
* fix dcn v2
* fix log_iter

configs/dcn/faster_rcnn_dcn_r101_vd_fpn_1x.yml

+architecture: FasterRCNN
+train_feed: FasterRCNNTrainFeed
+eval_feed: FasterRCNNEvalFeed
+test_feed: FasterRCNNTestFeed
+max_iters: 90000
+snapshot_iter: 10000
+use_gpu: true
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar
+weights: output/faster_rcnn_dcn_r101_vd_fpn_1x/model_final
+metric: COCO
+num_classes: 81
+
+FasterRCNN:
+  backbone: ResNet
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: BBoxHead
+  bbox_assigner: BBoxAssigner
+
+ResNet:
+  depth: 101
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  norm_type: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  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
+    post_nms_top_n: 2000
+    pre_nms_top_n: 2000
+  test_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    post_nms_top_n: 1000
+    pre_nms_top_n: 1000
+
+FPNRoIAlign:
+  canconical_level: 4
+  canonical_size: 224
+  max_level: 5
+  min_level: 2
+  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_hi: 0.5
+  bg_thresh_lo: 0.0
+  fg_fraction: 0.25
+  fg_thresh: 0.5
+
+BBoxHead:
+  head: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [60000, 80000]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2
+
+FasterRCNNTrainFeed:
+  # batch size per device
+  batch_size: 2
+  dataset:
+    dataset_dir: dataset/coco
+    image_dir: train2017
+    annotation: annotations/instances_train2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+
+FasterRCNNEvalFeed:
+  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
+
+FasterRCNNTestFeed:
+  batch_size: 1
+  dataset:
+    annotation: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

configs/dcn/faster_rcnn_dcn_r50_fpn_1x.yml

+architecture: FasterRCNN
+train_feed: FasterRCNNTrainFeed
+eval_feed: FasterRCNNEvalFeed
+test_feed: FasterRCNNTestFeed
+max_iters: 90000
+use_gpu: true
+snapshot_iter: 10000
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar
+metric: COCO
+weights: output/faster_rcnn_dcn_r50_fpn_1x/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
+  norm_type: bn
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  min_level: 2
+  max_level: 6
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [60000, 80000]
+  - !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: 2
+
+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/dcn/faster_rcnn_dcn_r50_vd_fpn_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
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+weights: output/faster_rcnn_dcn_r50_vd_fpn_2x/model_final
+metric: COCO
+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: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  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
+    post_nms_top_n: 2000
+    pre_nms_top_n: 2000
+  test_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    post_nms_top_n: 1000
+    pre_nms_top_n: 1000
+
+FPNRoIAlign:
+  canconical_level: 4
+  canonical_size: 224
+  max_level: 5
+  min_level: 2
+  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_hi: 0.5
+  bg_thresh_lo: 0.0
+  fg_fraction: 0.25
+  fg_thresh: 0.5
+
+BBoxHead:
+  head: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+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 per device
+  batch_size: 2
+  dataset:
+    dataset_dir: dataset/coco
+    image_dir: train2017
+    annotation: annotations/instances_train2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+
+FasterRCNNEvalFeed:
+  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
+
+FasterRCNNTestFeed:
+  batch_size: 1
+  dataset:
+    annotation: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

configs/dcn/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.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
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
+weights: output/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x/model_final
+metric: COCO
+num_classes: 81
+
+FasterRCNN:
+  backbone: ResNeXt
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: BBoxHead
+  bbox_assigner: BBoxAssigner
+
+ResNeXt:
+  depth: 101
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  group_width: 4
+  groups: 64
+  norm_type: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  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
+    post_nms_top_n: 2000
+    pre_nms_top_n: 2000
+  test_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    post_nms_top_n: 1000
+    pre_nms_top_n: 1000
+
+FPNRoIAlign:
+  canconical_level: 4
+  canonical_size: 224
+  max_level: 5
+  min_level: 2
+  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_hi: 0.5
+  bg_thresh_lo: 0.0
+  fg_fraction: 0.25
+  fg_thresh: 0.5
+
+BBoxHead:
+  head: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.01
+  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 per device
+  batch_size: 1
+  dataset:
+    dataset_dir: dataset/coco
+    image_dir: train2017
+    annotation: annotations/instances_train2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+  shuffle: true
+
+FasterRCNNEvalFeed:
+  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
+
+FasterRCNNTestFeed:
+  batch_size: 1
+  dataset:
+    annotation: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2
+  shuffle: false

configs/dcn/mask_rcnn_dcn_r101_vd_fpn_1x.yml

+architecture: MaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+max_iters: 180000
+snapshot_iter: 10000
+use_gpu: true
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar 
+weights: output/mask_rcnn_dcn_r101_vd_fpn_1x/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: 101
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  norm_type: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  num_chan_reduced: 256
+  num_convs: 4
+  resolution: 28
+
+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: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.01
+  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
+
+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: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

configs/dcn/mask_rcnn_dcn_r50_fpn_1x.yml

+architecture: MaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+use_gpu: true
+max_iters: 180000
+snapshot_iter: 10000
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar
+metric: COCO
+weights: output/mask_rcnn_dcn_r50_fpn_1x/model_final/
+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: bn
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  num_chan_reduced: 256
+  num_convs: 4
+  resolution: 28
+
+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: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.01
+  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
+
+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: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

configs/dcn/mask_rcnn_dcn_r50_vd_fpn_2x.yml

+architecture: MaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+use_gpu: true
+max_iters: 360000
+snapshot_iter: 10000
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+metric: COCO
+weights: output/mask_rcnn_dcn_r50_vd_fpn_2x/model_final/
+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: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  box_resolution: 7
+  sampling_ratio: 2
+  mask_resolution: 14
+
+MaskHead:
+  dilation: 1
+  num_chan_reduced: 256
+  num_convs: 4
+  resolution: 28
+
+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: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [240000, 320000]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2
+
+MaskRCNNTrainFeed:
+  # batch size per device
+  batch_size: 1
+  dataset:
+    dataset_dir: dataset/coco
+    image_dir: train2017
+    annotation: annotations/instances_train2017.json
+  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: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

configs/dcn/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml

+architecture: MaskRCNN
+train_feed: MaskRCNNTrainFeed
+eval_feed: MaskRCNNEvalFeed
+test_feed: MaskRCNNTestFeed
+max_iters: 180000
+snapshot_iter: 10000
+use_gpu: true
+log_smooth_window: 20
+log_iter: 20
+save_dir: output
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar 
+weights: output/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x/model_final
+metric: COCO
+num_classes: 81
+
+MaskRCNN:
+  backbone: ResNeXt
+  fpn: FPN
+  rpn_head: FPNRPNHead
+  roi_extractor: FPNRoIAlign
+  bbox_head: BBoxHead
+  bbox_assigner: BBoxAssigner
+
+ResNeXt:
+  depth: 101
+  feature_maps: [2, 3, 4, 5]
+  freeze_at: 2
+  group_width: 4
+  groups: 64
+  norm_type: bn
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+FPN:
+  max_level: 6
+  min_level: 2
+  num_chan: 256
+  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
+
+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
+  num_chan_reduced: 256
+  num_convs: 4
+  resolution: 28
+
+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: TwoFCHead
+  nms:
+    keep_top_k: 100
+    nms_threshold: 0.5
+    score_threshold: 0.05
+
+TwoFCHead:
+  num_chan: 1024
+
+LearningRate:
+  base_lr: 0.01
+  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
+
+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: annotations/instances_val2017.json
+  batch_transforms:
+  - !PadBatch
+    pad_to_stride: 32
+  num_workers: 2

docs/MODEL_ZOO.md

@@ -24,7 +24,7 @@
 
 The backbone models pretrained on ImageNet are available. All backbone models are pretrained on standard ImageNet-1k dataset and can be downloaded [here](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#supported-models-and-performances).
 
-- Notes:  The ResNet50 model was trained with cosine LR decay schedule and can be downloaded [here](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar).
+- **Notes:**  The ResNet50 model was trained with cosine LR decay schedule and can be downloaded [here](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar).
 
 ## Baselines
 
@@ -58,6 +58,24 @@ The backbone models pretrained on ImageNet are available. All backbone models ar
 | SENet154-vd-FPN      | Faster         |    1    |  1.44x  |  42.9  |    -    | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_se154_vd_fpn_s1x.tar) |
 | SENet154-vd-FPN      | Mask           |    1    |  1.44x  |  44.0  |  38.7   | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_se154_vd_fpn_s1x.tar) |
 
+### Deformable ConvNets v2
+
+| Backbone             | Type           | Conv    | Image/gpu | Lr schd | Box AP | Mask AP |                           Download                           |
+| :------------------- | :------------- | :-----: |:--------: | :-----: | :----: | :-----: | :----------------------------------------------------------: |
+| ResNet50-FPN         | Faster         | c3-c5   |    2      |   1x    |  41.0  |    -    | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_fpn_1x.tar) |
+| ResNet50-vd-FPN      | Faster         | c3-c5   |    2      |   2x    |  42.4  |    -    | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_vd_fpn_2x.tar) |
+| ResNet101-vd-FPN     | Faster         | c3-c5   |    2      |   1x    |  44.1  |    -    | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r101_vd_fpn_1x.tar) |
+| ResNeXt101-vd-FPN    | Faster         | c3-c5   |    1      |   1x    |  45.2  |    -    | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
+| ResNet50-FPN         | Mask           | c3-c5   |    1      |   1x    |  41.9  |  37.3   | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_fpn_1x.tar) |
+| ResNet50-vd-FPN      | Mask           | c3-c5   |    1      |   2x    |  42.9  |  38.0   | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_vd_fpn_2x.tar) |
+| ResNet101-vd-FPN     | Mask           | c3-c5   |    1      |   1x    |  44.6  |  39.2   | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r101_vd_fpn_1x.tar) |
+| ResNeXt101-vd-FPN    | Mask           | c3-c5   |    1      |   1x    |  46.2  |  40.4   | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
+
+#### Notes:
+- Deformable ConvNets v2(dcn_v2) reference from [Deformable ConvNets v2](https://arxiv.org/abs/1811.11168).
+- `c3-c5` means adding `dcn` in resnet stage 3 to 5.
+- Detailed configuration file in [configs/dcn](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/PaddleDetection/configs/dcn)
+
 ### Yolo v3
 
 | Backbone     | Size | Image/gpu | Lr schd | Box AP | Download  |
@@ -86,7 +104,7 @@ The backbone models pretrained on ImageNet are available. All backbone models ar
 | ResNet34     | 416  |    8    |   270e  |  81.9  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) |
 | ResNet34     | 320  |    8    |   270e  |  80.1  | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) |
 
-**NOTE**: Yolo v3 is trained in 8 GPU with total batch size as 64 and trained 270 epoches. Yolo v3 training data augmentations: mixup,
+**Notes:** Yolo v3 is trained in 8 GPU with total batch size as 64 and trained 270 epoches. Yolo v3 training data augmentations: mixup,
 randomly color distortion, randomly cropping, randomly expansion, randomly interpolation method, randomly flippling. Yolo v3 used randomly
 reshaped minibatch in training, inferences can be performed on different image sizes with the same model weights, and we provided evaluation
 results of image size 608/416/320 above.
@@ -106,5 +124,5 @@ results of image size 608/416/320 above.
 | :----------- | :--: | :-----: | :-----: | :----: | :-------: |
 | MobileNet v1 | 300  |    32   |   120e  |  73.13  | [model](https://paddlemodels.bj.bcebos.com/object_detection/ssd_mobilenet_v1_voc.tar) |
 
-**NOTE**: SSD is trained in 2 GPU with totoal batch size as 64 and trained 120 epoches. SSD training data augmentations: randomly color distortion,
+**Notes:** SSD is trained in 2 GPU with totoal batch size as 64 and trained 120 epoches. SSD training data augmentations: randomly color distortion,
 randomly cropping, randomly expansion, randomly flipping.

docs/MODEL_ZOO_cn.md

@@ -56,6 +56,24 @@ Paddle提供基于ImageNet的骨架网络预训练模型。所有预训练模型
 | SENet154-vd-FPN      | Faster         |    1    |  1.44x  |  42.9  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_se154_vd_fpn_s1x.tar) |
 | SENet154-vd-FPN      | Mask           |    1    |  1.44x  |  44.0  |  38.7   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_se154_vd_fpn_s1x.tar) |
 
+### Deformable 卷积网络v2
+
+| 骨架网络             | 网络类型           | 卷积    | 每张GPU图片个数 | 学习率策略 | Box AP | Mask AP |                           下载                           |
+| :------------------- | :------------- | :-----: |:--------: | :-----: | :----: | :-----: | :----------------------------------------------------------: |
+| ResNet50-FPN         | Faster         | c3-c5   |    2      |   1x    |  41.0  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_fpn_1x.tar) |
+| ResNet50-vd-FPN      | Faster         | c3-c5   |    2      |   2x    |  42.4  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_vd_fpn_2x.tar) |
+| ResNet101-vd-FPN     | Faster         | c3-c5   |    2      |   1x    |  44.1  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r101_vd_fpn_1x.tar) |
+| ResNeXt101-vd-FPN    | Faster         | c3-c5   |    1      |   1x    |  45.2  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
+| ResNet50-FPN         | Mask           | c3-c5   |    1      |   1x    |  41.9  |  37.3   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_fpn_1x.tar) |
+| ResNet50-vd-FPN      | Mask           | c3-c5   |    1      |   2x    |  42.9  |  38.0   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_vd_fpn_2x.tar) |
+| ResNet101-vd-FPN     | Mask           | c3-c5   |    1      |   1x    |  44.6  |  39.2   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r101_vd_fpn_1x.tar) |
+| ResNeXt101-vd-FPN    | Mask           | c3-c5   |    1      |   1x    |  46.2  |  40.4   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
+
+#### 注意事项:
+- Deformable卷积网络v2(dcn_v2)参考自论文[Deformable ConvNets v2](https://arxiv.org/abs/1811.11168).
+- `c3-c5`意思是在resnet模块的3到5阶段增加`dcn`.
+- 详细的配置文件在[configs/dcn](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/PaddleDetection/configs/dcn)
+
 ### Yolo v3
 
 | 骨架网络     | 输入尺寸 | 每张GPU图片个数 | 学习率策略 | Box AP | 下载 |

ppdet/modeling/backbones/resnet.py

@@ -22,6 +22,7 @@ from paddle import fluid
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.framework import Variable
 from paddle.fluid.regularizer import L2Decay
+from paddle.fluid.initializer import Constant
 
 from ppdet.core.workspace import register, serializable
 from numbers import Integral
@@ -44,6 +45,7 @@ class ResNet(object):
         norm_decay (float): weight decay for normalization layer weights
         variant (str): ResNet variant, supports 'a', 'b', 'c', 'd' currently
         feature_maps (list): index of stages whose feature maps are returned
+        dcn_v2_stages (list): index of stages who select deformable conv v2
     """
 
     def __init__(self,
@@ -53,7 +55,8 @@ class ResNet(object):
                  freeze_norm=True,
                  norm_decay=0.,
                  variant='b',
-                 feature_maps=[2, 3, 4, 5]):
+                 feature_maps=[2, 3, 4, 5],
+                 dcn_v2_stages=[]):
         super(ResNet, self).__init__()
 
         if isinstance(feature_maps, Integral):
@@ -74,6 +77,7 @@ class ResNet(object):
         self.variant = variant
         self._model_type = 'ResNet'
         self.feature_maps = feature_maps
+        self.dcn_v2_stages = dcn_v2_stages
         self.depth_cfg = {
             18: ([2, 2, 2, 2], self.basicblock),
             34: ([3, 4, 6, 3], self.basicblock),
@@ -85,6 +89,19 @@ class ResNet(object):
         self._c1_out_chan_num = 64
         self.na = NameAdapter(self)
 
+    def _conv_offset(self, input, filter_size, stride, padding, act=None, name=None):
+        out_channel = filter_size * filter_size * 3
+        out = fluid.layers.conv2d(input,
+            num_filters=out_channel,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            param_attr=ParamAttr(initializer=Constant(0.0)),
+            bias_attr=ParamAttr(initializer=Constant(0.0)),
+            act=act,
+            name=name)
+        return out
+
     def _conv_norm(self,
                    input,
                    num_filters,
@@ -92,18 +109,50 @@ class ResNet(object):
                    stride=1,
                    groups=1,
                    act=None,
-                   name=None):
-        conv = fluid.layers.conv2d(
-            input=input,
-            num_filters=num_filters,
-            filter_size=filter_size,
-            stride=stride,
-            padding=(filter_size - 1) // 2,
-            groups=groups,
-            act=None,
-            param_attr=ParamAttr(name=name + "_weights"),
-            bias_attr=False,
-            name=name + '.conv2d.output.1')
+                   name=None,
+                   dcn_v2=False):
+        if not dcn_v2:
+            conv = fluid.layers.conv2d(
+                input=input,
+                num_filters=num_filters,
+                filter_size=filter_size,
+                stride=stride,
+                padding=(filter_size - 1) // 2,
+                groups=groups,
+                act=None,
+                param_attr=ParamAttr(name=name + "_weights"),
+                bias_attr=False,
+                name=name + '.conv2d.output.1')
+        else:
+            # select deformable conv"
+            offset_mask = self._conv_offset(
+                input=input,
+                filter_size=filter_size,
+                stride=stride,
+                padding=(filter_size - 1) // 2,
+                act=None,
+                name=name + "_conv_offset")
+            offset_channel = filter_size ** 2 * 2
+            mask_channel = filter_size ** 2
+            offset, mask = fluid.layers.split(
+                input=offset_mask,
+                num_or_sections=[offset_channel, mask_channel],
+                dim=1)
+            mask = fluid.layers.sigmoid(mask)
+            conv = fluid.layers.deformable_conv(
+                input=input,
+                offset=offset,
+                mask=mask,
+                num_filters=num_filters,
+                filter_size=filter_size,
+                stride=stride,
+                padding=(filter_size - 1) // 2,
+                groups=groups,
+                deformable_groups=1,
+                im2col_step=1,
+                param_attr=ParamAttr(name=name + "_weights"),
+                bias_attr=False,
+                name=name + ".conv2d.output.1")
 
         bn_name = self.na.fix_conv_norm_name(name)
 
@@ -168,7 +217,7 @@ class ResNet(object):
         else:
             return input
 
-    def bottleneck(self, input, num_filters, stride, is_first, name):
+    def bottleneck(self, input, num_filters, stride, is_first, name, dcn_v2=False):
         if self.variant == 'a':
             stride1, stride2 = stride, 1
         else:
@@ -192,7 +241,7 @@ class ResNet(object):
                     [num_filters * expand, 1, 1, None, 1, conv_name3]]
 
         residual = input
-        for (c, k, s, act, g, _name) in conv_def:
+        for i, (c, k, s, act, g, _name) in enumerate(conv_def):
             residual = self._conv_norm(
                 input=residual,
                 num_filters=c,
@@ -200,7 +249,8 @@ class ResNet(object):
                 stride=s,
                 act=act,
                 groups=g,
-                name=_name)
+                name=_name,
+                dcn_v2=(i==1 and dcn_v2))
         short = self._shortcut(
             input,
             num_filters * expand,
@@ -214,7 +264,8 @@ class ResNet(object):
         return fluid.layers.elementwise_add(
             x=short, y=residual, act='relu', name=name + ".add.output.5")
 
-    def basicblock(self, input, num_filters, stride, is_first, name):
+    def basicblock(self, input, num_filters, stride, is_first, name, dcn_v2=False):
+        assert dcn_v2 is False, "Not implemented yet."
         conv0 = self._conv_norm(
             input=input,
             num_filters=num_filters,
@@ -248,6 +299,7 @@ class ResNet(object):
 
         ch_out = self.stage_filters[stage_num - 2]
         is_first = False if stage_num != 2 else True
+        dcn_v2 = True if stage_num in self.dcn_v2_stages else False
         # Make the layer name and parameter name consistent
         # with ImageNet pre-trained model
         conv = input
@@ -260,7 +312,8 @@ class ResNet(object):
                 num_filters=ch_out,
                 stride=2 if i == 0 and stage_num != 2 else 1,
                 is_first=is_first,
-                name=conv_name)
+                name=conv_name,
+                dcn_v2=dcn_v2)
         return conv
 
     def c1_stage(self, input):

ppdet/modeling/backbones/resnext.py

@@ -37,6 +37,7 @@ class ResNeXt(ResNet):
         norm_decay (float): weight decay for normalization layer weights
         variant (str): ResNet variant, supports 'a', 'b', 'c', 'd' currently
         feature_maps (list): index of the stages whose feature maps are returned
+        dcn_v2_stages (list): index of stages who select deformable conv v2
     """
 
     def __init__(self,
@@ -48,7 +49,8 @@ class ResNeXt(ResNet):
                  freeze_norm=True,
                  norm_decay=True,
                  variant='a',
-                 feature_maps=[2, 3, 4, 5]):
+                 feature_maps=[2, 3, 4, 5],
+                 dcn_v2_stages=[]):
         assert depth in [50, 101, 152], "depth {} should be 50, 101 or 152"
         super(ResNeXt, self).__init__(depth, freeze_at, norm_type, freeze_norm,
                                       norm_decay, variant, feature_maps)
@@ -61,6 +63,7 @@ class ResNeXt(ResNet):
         self.groups = groups
         self.group_width = group_width
         self._model_type = 'ResNeXt'
+        self.dcn_v2_stages = dcn_v2_stages
 
 
 @register

ppdet/modeling/backbones/senet.py

@@ -42,6 +42,7 @@ class SENet(ResNeXt):
         norm_decay (float): weight decay for normalization layer weights
         variant (str): ResNet variant, supports 'a', 'b', 'c', 'd' currently
         feature_maps (list): index of the stages whose feature maps are returned
+        dcn_v2_stages (list): index of stages who select deformable conv v2
     """
 
     def __init__(self,
@@ -53,7 +54,8 @@ class SENet(ResNeXt):
                  freeze_norm=True,
                  norm_decay=0.,
                  variant='d',
-                 feature_maps=[2, 3, 4, 5]):
+                 feature_maps=[2, 3, 4, 5],
+                 dcn_v2_stages=[]):
         super(SENet, self).__init__(depth, groups, group_width, freeze_at,
                                     norm_type, freeze_norm, norm_decay, variant,
                                     feature_maps)
@@ -64,6 +66,7 @@ class SENet(ResNeXt):
         self.reduction_ratio = 16
         self._c1_out_chan_num = 128
         self._model_type = 'SEResNeXt'
+        self.dcn_v2_stages = dcn_v2_stages
 
     def _squeeze_excitation(self, input, num_channels, name=None):
         pool = fluid.layers.pool2d(