[Dygraph] add ResNet-vd, ResNet-34, ResNeXt-101 (#2071)

* add configs and modelzoo

* remove trainreader of batchsize 2

* add vd r34 x101

* fix typo

* add vd r34 x101

* add resnet dcn, clean code

* fix x101 width in py2

* add faster mask rcnn dcn configs and modelzoo

* fix dcn modelzoo

* clean dcn configs code

dygraph/configs/dcn/README.md

+### Deformable ConvNets v2
+
+| 骨架网络             | 网络类型           | 卷积    | 每张GPU图片个数 | 学习率策略 |推理时间(fps)| Box AP | Mask AP |                           下载                           | 配置文件 |
+| :------------------- | :------------- | :-----: |:--------: | :-----: | :-----------: |:----: | :-----: | :----------------------------------------------------------: | :----: |
+| ResNet50-FPN         | Faster         | c3-c5   |    2      |   1x    |    -     |  41.3  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_dcn_r50_fpn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/faster_rcnn_dcn_r50_fpn_1x_coco.yml) |
+| ResNet50-vd-FPN      | Faster         | c3-c5   |    2      |   2x    |    -     |  42.4  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_dcn_r50_vd_fpn_2x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/faster_rcnn_dcn_r50_vd_fpn_2x.yml) |
+| ResNet101-vd-FPN     | Faster         | c3-c5   |    2      |   1x    |    -     |  44.1  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_dcn_r101_vd_fpn_1x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/faster_rcnn_dcn_r101_vd_fpn_1x.yml) |
+| ResNeXt101-vd-FPN    | Faster         | c3-c5   |    1      |   1x    |    -     |  45.2  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.pdparams) |[配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml) |
+| ResNet50-FPN         | Mask           | c3-c5   |    1      |   1x    |    -     |  41.9  |  37.3   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/mask_rcnn_dcn_r50_fpn_1x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/mask_rcnn_dcn_r50_fpn_1x.yml) |
+| ResNet50-vd-FPN      | Mask           | c3-c5   |    1      |   2x    |    -     |  42.9  |  38.0   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/mask_rcnn_dcn_r50_vd_fpn_2x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/mask_rcnn_dcn_r50_vd_fpn_2x.yml) |
+| ResNet101-vd-FPN     | Mask           | c3-c5   |    1      |   1x    |    -     |  44.6  |  39.2   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/mask_rcnn_dcn_r101_vd_fpn_1x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/mask_rcnn_dcn_r101_vd_fpn_1x.yml) |
+| ResNeXt101-vd-FPN    | Mask           | c3-c5   |    1      |   1x    |     -    |  46.2  |  40.4   | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/dcn/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.yml) |
+
+**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`.
+
+## Citations
+```
+@inproceedings{dai2017deformable,
+  title={Deformable Convolutional Networks},
+  author={Dai, Jifeng and Qi, Haozhi and Xiong, Yuwen and Li, Yi and Zhang, Guodong and Hu, Han and Wei, Yichen},
+  booktitle={Proceedings of the IEEE international conference on computer vision},
+  year={2017}
+}
+@article{zhu2018deformable,
+  title={Deformable ConvNets v2: More Deformable, Better Results},
+  author={Zhu, Xizhou and Hu, Han and Lin, Stephen and Dai, Jifeng},
+  journal={arXiv preprint arXiv:1811.11168},
+  year={2018}
+}
+```

dygraph/configs/dcn/faster_rcnn_dcn_r101_vd_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/faster_rcnn_dcn_r101_fpn_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+TrainReader:
+  batch_size: 2

dygraph/configs/dcn/faster_rcnn_dcn_r50_fpn_1x_coco.yml

+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '../faster_rcnn/_base_/optimizer_1x.yml',
+  '../faster_rcnn/_base_/faster_rcnn_r50_fpn.yml',
+  '../faster_rcnn/_base_/faster_fpn_reader.yml',
+]
+weights: output/faster_rcnn_dcn_r50_fpn_1x_coco/model_final
+
+ResNet:
+  depth: 50
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+TrainReader:
+  batch_size: 2

dygraph/configs/dcn/faster_rcnn_dcn_r50_vd_fpn_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+weights: output/faster_rcnn_dcn_r50_vd_fpn_2x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 24
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+TrainReader:
+  batch_size: 2

dygraph/configs/dcn/faster_rcnn_dcn_x101_vd_64x4d_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+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_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/dcn/mask_rcnn_dcn_r101_vd_fpn_1x_coco.yml

+_BASE_: [
+  'mask_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/mask_rcnn_dcn_r101_fpn_1x_coco/model_final.pdparams
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/dcn/mask_rcnn_dcn_r50_fpn_1x_coco.yml

+_BASE_: [
+  '../datasets/coco_instance.yml',
+  '../runtime.yml',
+  '../mask_rcnn/_base_/optimizer_1x.yml',
+  '../mask_rcnn/_base_/mask_rcnn_r50_fpn.yml',
+  '../mask_rcnn/_base_/mask_fpn_reader.yml',
+]
+weights: output/mask_rcnn_dcn_r50_fpn_1x_coco/model_final
+
+ResNet:
+  depth: 50
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/dcn/mask_rcnn_dcn_r50_vd_fpn_2x_coco.yml

+_BASE_: [
+  'mask_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+weights: output/mask_rcnn_dcn_r50_vd_fpn_2x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/dcn/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x_coco.yml

+_BASE_: [
+  'mask_rcnn_dcn_r50_fpn_1x_coco.yml',
+]
+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_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  variant: d
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+  dcn_v2_stages: [1,2,3]
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/faster_rcnn/README.md

@@ -4,8 +4,11 @@
 
 | 骨架网络             | 网络类型       | 每张GPU图片个数 | 学习率策略 |推理时间(fps) | Box AP | Mask AP |                           下载                          | 配置文件 |
 | :------------------- | :------------- | :-----: | :-----: | :------------: | :-----: | :-----: | :-----------------------------------------------------: | :-----: |
-| ResNet50             | Faster         |    1    |   1x    |     ----     |  35.1  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r50_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn_r50_1x_coco.yml) |
-| ResNet50-FPN         | Faster         |    1    |   1x    |     ----     |  37.0  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r50_fpn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn_r50_fpn_1x_coco.yml) |
+| ResNet50             | Faster         |    1    |   1x    |     ----     |  35.1  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r50_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn/faster_rcnn_r50_1x_coco.yml) |
+| ResNet50-FPN         | Faster         |    1    |   1x    |     ----     |  37.0  |    -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r50_fpn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml) |
+| ResNet34-FPN         | Faster         |    2    |   1x    |     ----     |  36.7  |   -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r34_fpn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn/faster_rcnn_r34_fpn_1x_coco.yml) |
+| ResNet101            | Faster         |    1    |   1x    |     ----     |  38.3  |   -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r101_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn/faster_rcnn_r101_1x_coco.yml) |
+| ResNet101-FPN        | Faster         |    1    |   1x    |     ----     |  38.7  |   -    | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/faster_rcnn_r101_fpn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/dygraph/configs/faster_rcnn/faster_rcnn_r101_fpn_1x_coco.yml) |
 
 ## Citations
 ```

dygraph/configs/faster_rcnn/faster_rcnn_r101_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/faster_rcnn_r101_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4

dygraph/configs/faster_rcnn/faster_rcnn_r101_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/faster_rcnn_r101_fpn_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4

dygraph/configs/faster_rcnn/faster_rcnn_r101_fpn_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/faster_rcnn_r101_fpn_2x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500

dygraph/configs/faster_rcnn/faster_rcnn_r101_vd_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar
+weights: output/faster_rcnn_r101_vd_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 1000

dygraph/configs/faster_rcnn/faster_rcnn_r101_vd_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar
+weights: output/faster_rcnn_r101_vd_fpn_1x_coco/model_final.pdparams
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 1000

dygraph/configs/faster_rcnn/faster_rcnn_r34_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar
+weights: output/faster_rcnn_r34_fpn_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 34
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+FPN:
+  in_channels: [64, 128, 256, 512]
+  out_channel: 256
+  min_level: 0
+  max_level: 4
+  spatial_scale: [0.25, 0.125, 0.0625, 0.03125]
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+TrainReader:
+  batch_size: 2

dygraph/configs/faster_rcnn/faster_rcnn_r34_vd_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_vd_pretrained.tar
+weights: output/faster_rcnn_r34_vd_fpn_1x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 34
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+FPN:
+  in_channels: [64, 128, 256, 512]
+  out_channel: 256
+  min_level: 0
+  max_level: 4
+  spatial_scale: [0.25, 0.125, 0.0625, 0.03125]
+
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000
+
+TrainReader:
+  batch_size: 2

dygraph/configs/faster_rcnn/faster_rcnn_r50_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_1x_coco.yml',
+]
+weights: output/faster_rcnn_r50_2x_coco/model_final
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500

dygraph/configs/faster_rcnn/faster_rcnn_r50_fpn_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+weights: output/faster_rcnn_r50_fpn_2x_coco/model_final
+
+epoch: 24
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500
+
+TrainReader:
+  batch_size: 2

dygraph/configs/faster_rcnn/faster_rcnn_r50_vd_fpn_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+weights: output/faster_rcnn_r50_vd_fpn_2x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 24
+LearningRate:
+  base_lr: 0.02
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500
+
+TrainReader:
+  batch_size: 2

dygraph/configs/faster_rcnn/faster_rcnn_x101_vd_64x4d_fpn_1x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
+weights: output/faster_rcnn_x101_vd_64x4d_fpn_1x_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/faster_rcnn/faster_rcnn_x101_vd_64x4d_fpn_2x_coco.yml

+_BASE_: [
+  'faster_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
+weights: output/faster_rcnn_x101_vd_64x4d_fpn_2x_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/mask_rcnn/mask_rcnn_r101_fpn_1x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_fpn_1x_coco.yml',
+]
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar
+weights: output/mask_rcnn_r101_fpn_1x_coco/model_final.pdparams
+
+ResNet:
+  # index 0 stands for res2
+  depth: 101
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4

dygraph/configs/mask_rcnn/mask_rcnn_r50_2x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_1x_coco.yml',
+]
+weights: output/mask_rcnn_r50_2x_coco/model_final
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500

dygraph/configs/mask_rcnn/mask_rcnn_r50_fpn_2x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_fpn_1x_coco.yml',
+]
+weights: output/mask_rcnn_r50_fpn_2x_coco/model_final
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500

dygraph/configs/mask_rcnn/mask_rcnn_r50_vd_fpn_2x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar
+weights: output/mask_rcnn_r50_vd_fpn_2x_coco/model_final
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.3333333333333333
+    steps: 500

dygraph/configs/mask_rcnn/mask_rcnn_x101_vd_64x4d_fpn_1x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
+weights: output/mask_rcnn_x101_vd_64x4d_fpn_1x_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  variant: d
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 12
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [8, 11]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/configs/mask_rcnn/mask_rcnn_x101_vd_64x4d_fpn_2x_coco.yml

+_BASE_: [
+  'mask_rcnn_r50_fpn_1x_coco.yml',
+]
+
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
+weights: output/mask_rcnn_x101_vd_64x4d_fpn_2x_coco/model_final.pdparams
+
+ResNet:
+  # for ResNeXt: groups, base_width, base_channels
+  depth: 101
+  variant: d
+  groups: 64
+  base_width: 4
+  base_channels: 64
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [0,1,2,3]
+  num_stages: 4
+
+epoch: 24
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [16, 22]
+  - !LinearWarmup
+    start_factor: 0.1
+    steps: 1000

dygraph/ppdet/modeling/backbones/name_adapter.py

@@ -41,6 +41,17 @@ class NameAdapter(object):
             shortcut_name = name + "_branch1"
         return conv_name1, conv_name2, conv_name3, shortcut_name
 
+    def fix_basicblock_name(self, name):
+        if self.model_type == 'SEResNeXt':
+            conv_name1 = 'conv' + name + '_x1'
+            conv_name2 = 'conv' + name + '_x2'
+            shortcut_name = name
+        else:
+            conv_name1 = name + "_branch2a"
+            conv_name2 = name + "_branch2b"
+            shortcut_name = name + "_branch1"
+        return conv_name1, conv_name2, shortcut_name
+
     def fix_layer_warp_name(self, stage_num, count, i):
         name = 'res' + str(stage_num)
         if count > 10 and stage_num == 4:

dygraph/ppdet/modeling/backbones/resnet.py

@@ -12,16 +12,12 @@
 # See the License for the specific language governing permissions and   
 # limitations under the License.
 
-import numpy as np
-from paddle import ParamAttr
+import math
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
-from paddle.nn import Conv2D, BatchNorm
-from paddle.nn import MaxPool2D
-
+from paddle import ParamAttr
 from ppdet.core.workspace import register, serializable
-
 from paddle.regularizer import L2Decay
 from .name_adapter import NameAdapter
 from numbers import Integral
@@ -29,6 +25,14 @@ from ppdet.modeling.layers import DeformableConvV2
 
 __all__ = ['ResNet', 'Res5Head']
 
+ResNet_cfg = {
+    18: [2, 2, 2, 2],
+    34: [3, 4, 6, 3],
+    50: [3, 4, 6, 3],
+    101: [3, 4, 23, 3],
+    152: [3, 8, 36, 3],
+}
+
 
 class ConvNormLayer(nn.Layer):
     def __init__(self,
@@ -37,6 +41,7 @@ class ConvNormLayer(nn.Layer):
                  filter_size,
                  stride,
                  name_adapter,
+                 groups=1,
                  act=None,
                  norm_type='bn',
                  norm_decay=0.,
@@ -50,13 +55,13 @@ class ConvNormLayer(nn.Layer):
         self.act = act
 
         if not dcn_v2:
-            self.conv = Conv2D(
+            self.conv = nn.Conv2D(
                 in_channels=ch_in,
                 out_channels=ch_out,
                 kernel_size=filter_size,
                 stride=stride,
                 padding=(filter_size - 1) // 2,
-                groups=1,
+                groups=groups,
                 weight_attr=ParamAttr(
                     learning_rate=lr, name=name + "_weights"),
                 bias_attr=False)
@@ -67,7 +72,7 @@ class ConvNormLayer(nn.Layer):
                 kernel_size=filter_size,
                 stride=stride,
                 padding=(filter_size - 1) // 2,
-                groups=1,
+                groups=groups,
                 weight_attr=ParamAttr(
                     learning_rate=lr, name=name + '_weights'),
                 bias_attr=False,
@@ -87,14 +92,18 @@ class ConvNormLayer(nn.Layer):
             trainable=False if freeze_norm else True)
 
         global_stats = True if freeze_norm else False
-        self.norm = BatchNorm(
-            ch_out,
-            act=act,
-            param_attr=param_attr,
-            bias_attr=bias_attr,
-            use_global_stats=global_stats,
-            moving_mean_name=bn_name + '_mean',
-            moving_variance_name=bn_name + '_variance')
+        if norm_type == 'sync_bn':
+            self.norm = nn.SyncBatchNorm(
+                ch_out, weight_attr=param_attr, bias_attr=bias_attr)
+        else:
+            self.norm = nn.BatchNorm(
+                ch_out,
+                act=None,
+                param_attr=param_attr,
+                bias_attr=bias_attr,
+                use_global_stats=global_stats,
+                moving_mean_name=bn_name + '_mean',
+                moving_variance_name=bn_name + '_variance')
         norm_params = self.norm.parameters()
 
         if freeze_norm:
@@ -103,8 +112,104 @@ class ConvNormLayer(nn.Layer):
 
     def forward(self, inputs):
         out = self.conv(inputs)
-        if self.norm_type == 'bn':
+        if self.norm_type in ['bn', 'sync_bn']:
             out = self.norm(out)
+        if self.act:
+            out = getattr(F, self.act)(out)
+        return out
+
+
+class BasicBlock(nn.Layer):
+    def __init__(self,
+                 ch_in,
+                 ch_out,
+                 stride,
+                 shortcut,
+                 name_adapter,
+                 name,
+                 variant='b',
+                 lr=1.0,
+                 norm_type='bn',
+                 norm_decay=0.,
+                 freeze_norm=True,
+                 dcn_v2=False):
+        super(BasicBlock, self).__init__()
+        assert dcn_v2 is False, "Not implemented yet."
+        conv_name1, conv_name2, shortcut_name = name_adapter.fix_basicblock_name(
+            name)
+
+        self.shortcut = shortcut
+        if not shortcut:
+            if variant == 'd' and stride == 2:
+                self.short = nn.Sequential()
+                self.short.add_sublayer(
+                    'pool',
+                    nn.AvgPool2D(
+                        kernel_size=2, stride=2, padding=0, ceil_mode=True))
+                self.short.add_sublayer(
+                    'conv',
+                    ConvNormLayer(
+                        ch_in=ch_in,
+                        ch_out=ch_out,
+                        filter_size=1,
+                        stride=1,
+                        name_adapter=name_adapter,
+                        norm_type=norm_type,
+                        norm_decay=norm_decay,
+                        freeze_norm=freeze_norm,
+                        lr=lr,
+                        name=shortcut_name))
+            else:
+                self.short = ConvNormLayer(
+                    ch_in=ch_in,
+                    ch_out=ch_out,
+                    filter_size=1,
+                    stride=stride,
+                    name_adapter=name_adapter,
+                    norm_type=norm_type,
+                    norm_decay=norm_decay,
+                    freeze_norm=freeze_norm,
+                    lr=lr,
+                    name=shortcut_name)
+
+        self.branch2a = ConvNormLayer(
+            ch_in=ch_in,
+            ch_out=ch_out,
+            filter_size=3,
+            stride=stride,
+            name_adapter=name_adapter,
+            act='relu',
+            norm_type=norm_type,
+            norm_decay=norm_decay,
+            freeze_norm=freeze_norm,
+            lr=lr,
+            name=conv_name1)
+
+        self.branch2b = ConvNormLayer(
+            ch_in=ch_out,
+            ch_out=ch_out,
+            filter_size=3,
+            stride=1,
+            name_adapter=name_adapter,
+            act=None,
+            norm_type=norm_type,
+            norm_decay=norm_decay,
+            freeze_norm=freeze_norm,
+            lr=lr,
+            name=conv_name2)
+
+    def forward(self, inputs):
+        out = self.branch2a(inputs)
+        out = self.branch2b(out)
+
+        if self.shortcut:
+            short = inputs
+        else:
+            short = self.short(inputs)
+
+        out = paddle.add(x=out, y=short)
+        out = F.relu(out)
+
         return out
 
 
@@ -117,6 +222,9 @@ class BottleNeck(nn.Layer):
                  name_adapter,
                  name,
                  variant='b',
+                 groups=1,
+                 base_width=4,
+                 base_channels=64,
                  lr=1.0,
                  norm_type='bn',
                  norm_decay=0.,
@@ -128,29 +236,58 @@ class BottleNeck(nn.Layer):
         else:
             stride1, stride2 = 1, stride
 
+        # ResNeXt
+        if groups == 1:
+            width = ch_out
+        else:
+            width = int(
+                math.floor(ch_out * (base_width * 1.0 / base_channels)) *
+                groups)
+
         conv_name1, conv_name2, conv_name3, \
             shortcut_name = name_adapter.fix_bottleneck_name(name)
 
         self.shortcut = shortcut
         if not shortcut:
-            self.short = ConvNormLayer(
-                ch_in=ch_in,
-                ch_out=ch_out * 4,
-                filter_size=1,
-                stride=stride,
-                name_adapter=name_adapter,
-                norm_type=norm_type,
-                norm_decay=norm_decay,
-                freeze_norm=freeze_norm,
-                lr=lr,
-                name=shortcut_name)
+            if variant == 'd' and stride == 2:
+                self.short = nn.Sequential()
+                self.short.add_sublayer(
+                    'pool',
+                    nn.AvgPool2D(
+                        kernel_size=2, stride=2, padding=0, ceil_mode=True))
+                self.short.add_sublayer(
+                    'conv',
+                    ConvNormLayer(
+                        ch_in=ch_in,
+                        ch_out=ch_out * 4,
+                        filter_size=1,
+                        stride=1,
+                        name_adapter=name_adapter,
+                        norm_type=norm_type,
+                        norm_decay=norm_decay,
+                        freeze_norm=freeze_norm,
+                        lr=lr,
+                        name=shortcut_name))
+            else:
+                self.short = ConvNormLayer(
+                    ch_in=ch_in,
+                    ch_out=ch_out * 4,
+                    filter_size=1,
+                    stride=stride,
+                    name_adapter=name_adapter,
+                    norm_type=norm_type,
+                    norm_decay=norm_decay,
+                    freeze_norm=freeze_norm,
+                    lr=lr,
+                    name=shortcut_name)
 
         self.branch2a = ConvNormLayer(
             ch_in=ch_in,
-            ch_out=ch_out,
+            ch_out=width,
             filter_size=1,
             stride=stride1,
             name_adapter=name_adapter,
+            groups=1,
             act='relu',
             norm_type=norm_type,
             norm_decay=norm_decay,
@@ -159,11 +296,12 @@ class BottleNeck(nn.Layer):
             name=conv_name1)
 
         self.branch2b = ConvNormLayer(
-            ch_in=ch_out,
-            ch_out=ch_out,
+            ch_in=width,
+            ch_out=width,
             filter_size=3,
             stride=stride2,
             name_adapter=name_adapter,
+            groups=groups,
             act='relu',
             norm_type=norm_type,
             norm_decay=norm_decay,
@@ -173,11 +311,12 @@ class BottleNeck(nn.Layer):
             name=conv_name2)
 
         self.branch2c = ConvNormLayer(
-            ch_in=ch_out,
+            ch_in=width,
             ch_out=ch_out * 4,
             filter_size=1,
             stride=1,
             name_adapter=name_adapter,
+            groups=1,
             norm_type=norm_type,
             norm_decay=norm_decay,
             freeze_norm=freeze_norm,
@@ -202,11 +341,16 @@ class BottleNeck(nn.Layer):
 
 class Blocks(nn.Layer):
     def __init__(self,
+                 depth,
                  ch_in,
                  ch_out,
                  count,
                  name_adapter,
                  stage_num,
+                 variant='b',
+                 groups=1,
+                 base_width=-1,
+                 base_channels=-1,
                  lr=1.0,
                  norm_type='bn',
                  norm_decay=0.,
@@ -217,22 +361,42 @@ class Blocks(nn.Layer):
         self.blocks = []
         for i in range(count):
             conv_name = name_adapter.fix_layer_warp_name(stage_num, count, i)
-
-            block = self.add_sublayer(
-                conv_name,
-                BottleNeck(
-                    ch_in=ch_in if i == 0 else ch_out * 4,
-                    ch_out=ch_out,
-                    stride=2 if i == 0 and stage_num != 2 else 1,
-                    shortcut=False if i == 0 else True,
-                    name_adapter=name_adapter,
-                    name=conv_name,
-                    variant=name_adapter.variant,
-                    lr=lr,
-                    norm_type=norm_type,
-                    norm_decay=norm_decay,
-                    freeze_norm=freeze_norm,
-                    dcn_v2=dcn_v2))
+            if depth >= 50:
+                block = self.add_sublayer(
+                    conv_name,
+                    BottleNeck(
+                        ch_in=ch_in if i == 0 else ch_out * 4,
+                        ch_out=ch_out,
+                        stride=2 if i == 0 and stage_num != 2 else 1,
+                        shortcut=False if i == 0 else True,
+                        name_adapter=name_adapter,
+                        name=conv_name,
+                        variant=variant,
+                        groups=groups,
+                        base_width=base_width,
+                        base_channels=base_channels,
+                        lr=lr,
+                        norm_type=norm_type,
+                        norm_decay=norm_decay,
+                        freeze_norm=freeze_norm,
+                        dcn_v2=dcn_v2))
+            else:
+                ch_in = ch_in // 4 if i > 0 else ch_in
+                block = self.add_sublayer(
+                    conv_name,
+                    BasicBlock(
+                        ch_in=ch_in if i == 0 else ch_out,
+                        ch_out=ch_out,
+                        stride=2 if i == 0 and stage_num != 2 else 1,
+                        shortcut=False if i == 0 else True,
+                        name_adapter=name_adapter,
+                        name=conv_name,
+                        variant=variant,
+                        lr=lr,
+                        norm_type=norm_type,
+                        norm_decay=norm_decay,
+                        freeze_norm=freeze_norm,
+                        dcn_v2=dcn_v2))
             self.blocks.append(block)
 
     def forward(self, inputs):
@@ -242,16 +406,18 @@ class Blocks(nn.Layer):
         return block_out
 
 
-ResNet_cfg = {50: [3, 4, 6, 3], 101: [3, 4, 23, 3], 152: [3, 8, 36, 3]}
-
-
 @register
 @serializable
 class ResNet(nn.Layer):
+    __shared__ = ['norm_type']
+
     def __init__(self,
                  depth=50,
                  variant='b',
-                 lr_mult=1.,
+                 lr_mult_list=[1.0, 1.0, 1.0, 1.0],
+                 groups=1,
+                 base_width=-1,
+                 base_channels=-1,
                  norm_type='bn',
                  norm_decay=0,
                  freeze_norm=True,
@@ -260,6 +426,8 @@ class ResNet(nn.Layer):
                  dcn_v2_stages=[-1],
                  num_stages=4):
         super(ResNet, self).__init__()
+        self._model_type = 'ResNet' if groups == 1 else 'ResNeXt'
+        assert num_stages >= 1 and num_stages <= 4
         self.depth = depth
         self.variant = variant
         self.norm_type = norm_type
@@ -274,6 +442,11 @@ class ResNet(nn.Layer):
             'is {}'.format(max(return_idx), num_stages)
         self.return_idx = return_idx
         self.num_stages = num_stages
+        assert len(lr_mult_list) == 4, \
+            "lr_mult_list length must be 4 but got {}".format(len(lr_mult_list))
+        if isinstance(dcn_v2_stages, Integral):
+            dcn_v2_stages = [dcn_v2_stages]
+        assert max(dcn_v2_stages) < num_stages
 
         if isinstance(dcn_v2_stages, Integral):
             dcn_v2_stages = [dcn_v2_stages]
@@ -302,30 +475,38 @@ class ResNet(nn.Layer):
                     filter_size=k,
                     stride=s,
                     name_adapter=na,
+                    groups=1,
                     act='relu',
                     norm_type=norm_type,
                     norm_decay=norm_decay,
                     freeze_norm=freeze_norm,
-                    lr=lr_mult,
+                    lr=1.0,
                     name=_name))
 
-        self.pool = MaxPool2D(kernel_size=3, stride=2, padding=1)
+        self.pool = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
 
         ch_in_list = [64, 256, 512, 1024]
         ch_out_list = [64, 128, 256, 512]
 
         self.res_layers = []
         for i in range(num_stages):
+            lr_mult = lr_mult_list[i]
             stage_num = i + 2
             res_name = "res{}".format(stage_num)
             res_layer = self.add_sublayer(
                 res_name,
                 Blocks(
-                    ch_in_list[i],
+                    depth,
+                    ch_in_list[i] // 4
+                    if i > 0 and depth < 50 else ch_in_list[i],
                     ch_out_list[i],
                     count=block_nums[i],
                     name_adapter=na,
                     stage_num=stage_num,
+                    variant=variant,
+                    groups=groups,
+                    base_width=base_width,
+                    base_channels=base_channels,
                     lr=lr_mult,
                     norm_type=norm_type,
                     norm_decay=norm_decay,