[transformer] add readme and deformable configs (#3720)

configs/deformable_detr/README.md

+# Deformable DETR
+
+## Introduction
+
+
+Deformable DETR is an object detection model based on DETR. We reproduced the model of the paper.
+
+
+## Model Zoo
+
+| Backbone | Model | Images/GPU  | Inf time (fps) | Box AP | Config | Download |
+|:------:|:--------:|:--------:|:--------------:|:------:|:------:|:--------:|
+| R-50 | Deformable DETR  | 2 | --- | 44.1 | [config](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/deformable_detr/deformable_detr_r50_1x_coco.yml) | [model](https://paddledet.bj.bcebos.com/models/deformable_detr_r50_1x_coco.pdparams) |
+
+**Notes:**
+
+- Deformable DETR is trained on COCO train2017 dataset and evaluated on val2017 results of `mAP(IoU=0.5:0.95)`.
+- Deformable DETR uses 8GPU to train 50 epochs.
+
+## Citations
+```
+@inproceedings{
+zhu2021deformable,
+title={Deformable DETR: Deformable Transformers for End-to-End Object Detection},
+author={Xizhou Zhu and Weijie Su and Lewei Lu and Bin Li and Xiaogang Wang and Jifeng Dai},
+booktitle={International Conference on Learning Representations},
+year={2021},
+url={https://openreview.net/forum?id=gZ9hCDWe6ke}
+}
+```

configs/deformable_detr/_base_/deformable_detr_r50.yml

+architecture: DETR
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vb_normal_pretrained.pdparams
+hidden_dim: 256
+use_focal_loss: True
+
+
+DETR:
+  backbone: ResNet
+  transformer: DeformableTransformer
+  detr_head: DeformableDETRHead
+  post_process: DETRBBoxPostProcess
+
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [1, 2, 3]
+  lr_mult_list: [0.0, 0.1, 0.1, 0.1]
+  num_stages: 4
+
+
+DeformableTransformer:
+  num_queries: 300
+  position_embed_type: sine
+  nhead: 8
+  num_encoder_layers: 6
+  num_decoder_layers: 6
+  dim_feedforward: 1024
+  dropout: 0.1
+  activation: relu
+  num_feature_levels: 4
+  num_encoder_points: 4
+  num_decoder_points: 4
+
+
+DeformableDETRHead:
+  num_mlp_layers: 3
+
+
+DETRLoss:
+  loss_coeff: {class: 2, bbox: 5, giou: 2, mask: 1, dice: 1}
+  aux_loss: True
+
+
+HungarianMatcher:
+  matcher_coeff: {class: 2, bbox: 5, giou: 2}

configs/deformable_detr/_base_/deformable_detr_reader.yml

+worker_num: 0
+TrainReader:
+  sample_transforms:
+  - Decode: {}
+  - RandomFlip: {prob: 0.5}
+  - RandomSelect: { transforms1: [ RandomShortSideResize: { short_side_sizes: [ 480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800 ], max_size: 1333 } ],
+                    transforms2: [
+                        RandomShortSideResize: { short_side_sizes: [ 400, 500, 600 ] },
+                        RandomSizeCrop: { min_size: 384, max_size: 600 },
+                        RandomShortSideResize: { short_side_sizes: [ 480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800 ], max_size: 1333 } ]
+  }
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - NormalizeBox: {}
+  - BboxXYXY2XYWH: {}
+  - Permute: {}
+  batch_transforms:
+  - PadMaskBatch: {pad_to_stride: -1, return_pad_mask: true}
+  batch_size: 2
+  shuffle: true
+  drop_last: true
+  collate_batch: false
+  use_shared_memory: false
+
+
+EvalReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadMaskBatch: {pad_to_stride: -1, return_pad_mask: true}
+  batch_size: 1
+  shuffle: false
+  drop_last: false
+  drop_empty: false
+
+
+TestReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadMaskBatch: {pad_to_stride: -1, return_pad_mask: true}
+  batch_size: 1
+  shuffle: false
+  drop_last: false

configs/deformable_detr/_base_/deformable_optimizer_1x.yml

+epoch: 50
+
+LearningRate:
+  base_lr: 0.0002
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [40]
+    use_warmup: false
+
+OptimizerBuilder:
+  clip_grad_by_norm: 0.1
+  regularizer: false
+  optimizer:
+    type: AdamW
+    weight_decay: 0.0001

configs/deformable_detr/deformable_detr_r50_1x_coco.yml

+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/deformable_optimizer_1x.yml',
+  '_base_/deformable_detr_r50.yml',
+  '_base_/deformable_detr_reader.yml',
+]
+weights: output/deformable_detr_r50_1x_coco/model_final

configs/detr/README.md

+# DETR
+
+## Introduction
+
+
+DETR is an object detection model based on transformer. We reproduced the model of the paper.
+
+
+## Model Zoo
+
+| Backbone | Model | Images/GPU  | Inf time (fps) | Box AP | Config | Download |
+|:------:|:--------:|:--------:|:--------------:|:------:|:------:|:--------:|
+| R-50 | DETR  | 4 | --- | 42.3 | [config](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/detr/detr_r50_1x_coco.yml) | [model](https://paddledet.bj.bcebos.com/models/detr_r50_1x_coco.pdparams) |
+
+**Notes:**
+
+- DETR is trained on COCO train2017 dataset and evaluated on val2017 results of `mAP(IoU=0.5:0.95)`.
+- DETR uses 8GPU to train 500 epochs.
+
+## Citations
+```
+@inproceedings{detr,
+  author    = {Nicolas Carion and
+               Francisco Massa and
+               Gabriel Synnaeve and
+               Nicolas Usunier and
+               Alexander Kirillov and
+               Sergey Zagoruyko},
+  title     = {End-to-End Object Detection with Transformers},
+  booktitle = {ECCV},
+  year      = {2020}
+}
+```

configs/detr/_base_/detr_r50.yml

 architecture: DETR
-pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_cos_pretrained.pdparams
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vb_normal_pretrained.pdparams
 hidden_dim: 256