keypoint petr (#7774)

* petr train ok train ok refix augsize affine size fix update msdeformable fix flip/affine fix clip add resize area add distortion debug mode fix pos_inds update edge joints update word mistake * delete extra codes;adapt transformer modify;update code format * reverse old transformer modify * integrate datasets

keypoint petr (#7774)
* petr train ok train ok refix augsize affine size fix update msdeformable fix flip/affine fix clip add resize area add distortion debug mode fix pos_inds update edge joints update word mistake * delete extra codes;adapt transformer modify;update code format * reverse old transformer modify * integrate datasets
4c04ce74 · zhiboniu · GitHub · 2685fb56 · 4c04ce74 · 4c04ce74
27 changed file
--- a/.gitignore
+++ b/.gitignore
@@ -18,9 +18,9 @@ __pycache__/
 # Distribution / packaging
 /bin/
-/build/
+*build/
 /develop-eggs/
-/dist/
+*dist/
 /eggs/
 /lib/
 /lib64/
@@ -30,7 +30,7 @@ __pycache__/
 /parts/
 /sdist/
 /var/
-/*.egg-info/
+*.egg-info/
 /.installed.cfg
 /*.egg
 /.eggs

--- a/configs/keypoint/README.md
+++ b/configs/keypoint/README.md
@@ -56,8 +56,10 @@ PaddleDetection 中的关键点检测部分紧跟最先进的算法，包括 Top
 ## 模型库
 COCO数据集
 | 模型              |  方案              |输入尺寸 | AP(coco val) |                           模型下载                           | 配置文件 |  
 | :---------------- | -------- | :----------: | :----------------------------------------------------------: | ----------------------------------------------------| ------- |
+| PETR_Res50       |One-Stage| 512      |     65.5     | [petr_res50.pdparams](https://bj.bcebos.com/v1/paddledet/models/keypoint/petr_resnet50_16x2_coco.pdparams) | [config](./petr/petr_resnet50_16x2_coco.yml)       |
 | HigherHRNet-w32       |Bottom-Up| 512      |     67.1     | [higherhrnet_hrnet_w32_512.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_512.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_512.yml)       |
 | HigherHRNet-w32       | Bottom-Up| 640      |     68.3     | [higherhrnet_hrnet_w32_640.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_640.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_640.yml)       |
 | HigherHRNet-w32+SWAHR |Bottom-Up|  512      |     68.9     | [higherhrnet_hrnet_w32_512_swahr.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_512_swahr.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_512_swahr.yml) |

--- a/configs/keypoint/README_en.md
+++ b/configs/keypoint/README_en.md
@@ -62,6 +62,7 @@ At the same time, PaddleDetection provides a self-developed real-time keypoint d
 COCO Dataset
 | Model              | Input Size | AP(coco val) |                           Model Download                           | Config File                                                    |
 | :---------------- | -------- | :----------: | :----------------------------------------------------------: | ----------------------------------------------------------- |
+| PETR_Res50       |One-Stage| 512      |     65.5     | [petr_res50.pdparams](https://bj.bcebos.com/v1/paddledet/models/keypoint/petr_resnet50_16x2_coco.pdparams) | [config](./petr/petr_resnet50_16x2_coco.yml)       |
 | HigherHRNet-w32       | 512      |     67.1     | [higherhrnet_hrnet_w32_512.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_512.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_512.yml)       |
 | HigherHRNet-w32       | 640      |     68.3     | [higherhrnet_hrnet_w32_640.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_640.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_640.yml)       |
 | HigherHRNet-w32+SWAHR | 512      |     68.9     | [higherhrnet_hrnet_w32_512_swahr.pdparams](https://paddledet.bj.bcebos.com/models/keypoint/higherhrnet_hrnet_w32_512_swahr.pdparams) | [config](./higherhrnet/higherhrnet_hrnet_w32_512_swahr.yml) |

--- a/configs/keypoint/petr/petr_resnet50_16x2_coco.yml
+++ b/configs/keypoint/petr/petr_resnet50_16x2_coco.yml
+use_gpu: true
+log_iter: 50
+save_dir: output
+snapshot_epoch: 1
+weights: output/petr_resnet50_16x2_coco/model_final
+epoch: 100
+num_joints: &num_joints 17
+pixel_std: &pixel_std 200
+metric: COCO
+num_classes: 1
+trainsize: &trainsize 512
+flip_perm: &flip_perm [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
+find_unused_parameters: False
+#####model
+architecture: PETR
+pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/PETR_pretrained.pdparams
+PETR:
+  backbone:
+    name: ResNet
+    depth: 50
+    variant: b
+    norm_type: bn
+    freeze_norm: True
+    freeze_at: 0
+    return_idx: [1,2,3]
+    num_stages: 4
+    lr_mult_list: [0.1, 0.1, 0.1, 0.1]
+  neck:
+    name: ChannelMapper
+    in_channels: [512, 1024, 2048]
+    kernel_size: 1
+    out_channels: 256
+    norm_type: "gn"
+    norm_groups: 32
+    act: None
+    num_outs: 4
+  bbox_head:
+    name: PETRHead
+    num_query: 300
+    num_classes: 1  # only person
+    in_channels: 2048
+    sync_cls_avg_factor: true
+    with_kpt_refine: true
+    transformer:
+      name: PETRTransformer
+      as_two_stage: true
+      encoder:
+        name: TransformerEncoder
+        encoder_layer:
+          name: TransformerEncoderLayer
+          d_model: 256
+          attn:
+            name: MSDeformableAttention
+            embed_dim: 256
+            num_heads: 8
+            num_levels: 4
+            num_points: 4
+          dim_feedforward: 1024
+          dropout: 0.1
+        num_layers: 6
+      decoder:
+        name: PETR_TransformerDecoder
+        num_layers: 3
+        return_intermediate: true
+        decoder_layer:
+          name: PETR_TransformerDecoderLayer
+          d_model: 256
+          dim_feedforward: 1024
+          dropout: 0.1
+          self_attn:
+            name: MultiHeadAttention
+            embed_dim: 256
+            num_heads: 8
+            dropout: 0.1
+          cross_attn:
+            name: MultiScaleDeformablePoseAttention
+            embed_dims: 256
+            num_heads: 8
+            num_levels: 4
+            num_points: 17
+      hm_encoder:
+        name: TransformerEncoder
+        encoder_layer:
+          name: TransformerEncoderLayer
+          d_model: 256
+          attn:
+            name: MSDeformableAttention
+            embed_dim: 256
+            num_heads: 8
+            num_levels: 1
+            num_points: 4
+          dim_feedforward: 1024
+          dropout: 0.1
+        num_layers: 1
+      refine_decoder:
+        name: PETR_DeformableDetrTransformerDecoder
+        num_layers: 2
+        return_intermediate: true
+        decoder_layer:
+          name: PETR_TransformerDecoderLayer
+          d_model: 256
+          dim_feedforward: 1024
+          dropout: 0.1
+          self_attn:
+            name: MultiHeadAttention
+            embed_dim: 256
+            num_heads: 8
+            dropout: 0.1
+          cross_attn:
+            name: MSDeformableAttention
+            embed_dim: 256
+            num_levels: 4
+    positional_encoding:
+      name: PositionEmbedding
+      num_pos_feats: 128
+      normalize: true
+      offset: -0.5
+    loss_cls:
+        name: Weighted_FocalLoss
+        use_sigmoid: true
+        gamma: 2.0
+        alpha: 0.25
+        loss_weight: 2.0
+        reduction: "mean"
+    loss_kpt:
+      name: L1Loss
+      loss_weight: 70.0
+    loss_kpt_rpn:
+      name: L1Loss
+      loss_weight: 70.0
+    loss_oks:
+      name: OKSLoss
+      loss_weight: 2.0
+    loss_hm:
+      name: CenterFocalLoss
+      loss_weight: 4.0
+    loss_kpt_refine:
+      name: L1Loss
+      loss_weight: 80.0
+    loss_oks_refine:
+      name: OKSLoss
+      loss_weight: 3.0
+    assigner:
+        name: PoseHungarianAssigner
+        cls_cost:
+          name: FocalLossCost
+          weight: 2.0
+        kpt_cost:
+          name: KptL1Cost
+          weight: 70.0
+        oks_cost:
+          name: OksCost
+          weight: 7.0
+#####optimizer
+LearningRate:
+  base_lr: 0.0002
+  schedulers:
+  - !PiecewiseDecay
+    milestones: [80]
+    gamma: 0.1
+    use_warmup: false
+  # - !LinearWarmup
+  #   start_factor: 0.001
+  #   steps: 1000
+OptimizerBuilder:
+  clip_grad_by_norm: 0.1
+  optimizer:
+    type: AdamW
+  regularizer:
+    factor: 0.0001
+    type: L2
+#####data
+TrainDataset:
+  !KeypointBottomUpCocoDataset
+    image_dir: train2017
+    anno_path: annotations/person_keypoints_train2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    return_mask: false
+EvalDataset:
+  !KeypointBottomUpCocoDataset
+    image_dir: val2017
+    anno_path: annotations/person_keypoints_val2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    test_mode: true
+    return_mask: false
+TestDataset:
+  !ImageFolder
+    anno_path: dataset/coco/keypoint_imagelist.txt
+worker_num: 2
+global_mean: &global_mean [0.485, 0.456, 0.406]
+global_std: &global_std [0.229, 0.224, 0.225]
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - PhotoMetricDistortion:
+        brightness_delta: 32
+        contrast_range: [0.5, 1.5]
+        saturation_range: [0.5, 1.5]
+        hue_delta: 18
+    - KeyPointFlip:
+        flip_prob: 0.5
+        flip_permutation: *flip_perm
+    - RandomAffine:
+        max_degree: 30
+        scale: [1.0, 1.0]
+        max_shift: 0.
+        trainsize: -1
+    - RandomSelect: { transforms1: [ RandomShortSideRangeResize: { scales: [[400, 1400], [1400, 1400]]} ],
+                      transforms2: [
+                          RandomShortSideResize: { short_side_sizes: [ 400, 500, 600 ] },
+                          RandomSizeCrop: { min_size: 384, max_size: 600},
+                          RandomShortSideRangeResize: { scales: [[400, 1400], [1400, 1400]]} ]}
+  batch_transforms:
+    - NormalizeImage: {mean: *global_mean, std: *global_std, is_scale: True}
+    - PadGT: {pad_img: True, minimum_gtnum: 1}
+    - Permute: {}
+  batch_size: 2
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+  collate_batch: true
+EvalReader:
+  sample_transforms:
+    - PETR_Resize: {img_scale: [[800, 1333]], keep_ratio: True}
+    # - MultiscaleTestResize: {origin_target_size: [[800, 1333]], use_flip: false}
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 1
+TestReader:
+  sample_transforms:
+    - Decode: {}
+    - EvalAffine:
+        size: *trainsize
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 1
--- a/docs/tutorials/data/PrepareKeypointDataSet.md
+++ b/docs/tutorials/data/PrepareKeypointDataSet.md
@@ -82,7 +82,7 @@ MPII keypoint indexes:
 ```
 {
    'joints_vis': [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-    'joints': [
+    'gt_joints': [
        [-1.0, -1.0],
        [-1.0, -1.0],
        [-1.0, -1.0],

--- a/docs/tutorials/data/PrepareKeypointDataSet_en.md
+++ b/docs/tutorials/data/PrepareKeypointDataSet_en.md
@@ -82,7 +82,7 @@ The following example takes a parsed annotation information to illustrate the co
 ```
 {
    'joints_vis': [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-    'joints': [
+    'gt_joints': [
        [-1.0, -1.0],
        [-1.0, -1.0],
        [-1.0, -1.0],

--- a/ppdet/data/source/keypoint_coco.py
+++ b/ppdet/data/source/keypoint_coco.py
@@ -80,7 +80,8 @@ class KeypointBottomUpBaseDataset(DetDataset):
        records = copy.deepcopy(self._get_imganno(idx))
        records['image'] = cv2.imread(records['image_file'])
        records['image'] = cv2.cvtColor(records['image'], cv2.COLOR_BGR2RGB)
-        records['mask'] = (records['mask'] + 0).astype('uint8')
+        if 'mask' in records:
+            records['mask'] = (records['mask'] + 0).astype('uint8')
        records = self.transform(records)
        return records
@@ -135,24 +136,37 @@ class KeypointBottomUpCocoDataset(KeypointBottomUpBaseDataset):
                 num_joints,
                 transform=[],
                 shard=[0, 1],
-                 test_mode=False):
+                 test_mode=False,
+                 return_mask=True,
+                 return_bbox=True,
+                 return_area=True,
+                 return_class=True):
        super().__init__(dataset_dir, image_dir, anno_path, num_joints,
                         transform, shard, test_mode)
        self.ann_file = os.path.join(dataset_dir, anno_path)
        self.shard = shard
        self.test_mode = test_mode
+        self.return_mask = return_mask
+        self.return_bbox = return_bbox
+        self.return_area = return_area
+        self.return_class = return_class
    def parse_dataset(self):
        self.coco = COCO(self.ann_file)
        self.img_ids = self.coco.getImgIds()
        if not self.test_mode:
-            self.img_ids = [
+            self.img_ids_tmp = []
-                img_id for img_id in self.img_ids
+            for img_id in self.img_ids:
-                if len(self.coco.getAnnIds(
+                ann_ids = self.coco.getAnnIds(imgIds=img_id)
-                    imgIds=img_id, iscrowd=None)) > 0
+                anno = self.coco.loadAnns(ann_ids)
-            ]
+                anno = [obj for obj in anno if obj['iscrowd'] == 0]
+                if len(anno) == 0:
+                    continue
+                self.img_ids_tmp.append(img_id)
+            self.img_ids = self.img_ids_tmp
        blocknum = int(len(self.img_ids) / self.shard[1])
        self.img_ids = self.img_ids[(blocknum * self.shard[0]):(blocknum * (
            self.shard[0] + 1))]
@@ -199,21 +213,31 @@ class KeypointBottomUpCocoDataset(KeypointBottomUpBaseDataset):
        ann_ids = coco.getAnnIds(imgIds=img_id)
        anno = coco.loadAnns(ann_ids)
-        mask = self._get_mask(anno, idx)
        anno = [
            obj for obj in anno
-            if obj['iscrowd'] == 0 or obj['num_keypoints'] > 0
+            if obj['iscrowd'] == 0 and obj['num_keypoints'] > 0
        ]
+        db_rec = {}
        joints, orgsize = self._get_joints(anno, idx)
+        db_rec['gt_joints'] = joints
+        db_rec['im_shape'] = orgsize
+        if self.return_bbox:
+            db_rec['gt_bbox'] = self._get_bboxs(anno, idx)
+        if self.return_class:
+            db_rec['gt_class'] = self._get_labels(anno, idx)
+        if self.return_area:
+            db_rec['gt_areas'] = self._get_areas(anno, idx)
+        if self.return_mask:
+            db_rec['mask'] = self._get_mask(anno, idx)
-        db_rec = {}
        db_rec['im_id'] = img_id
        db_rec['image_file'] = os.path.join(self.img_prefix,
                                            self.id2name[img_id])
-        db_rec['mask'] = mask
-        db_rec['joints'] = joints
-        db_rec['im_shape'] = orgsize
        return db_rec
@@ -229,12 +253,41 @@ class KeypointBottomUpCocoDataset(KeypointBottomUpBaseDataset):
                np.array(obj['keypoints']).reshape([-1, 3])
        img_info = self.coco.loadImgs(self.img_ids[idx])[0]
-        joints[..., 0] /= img_info['width']
+        orgsize = np.array([img_info['height'], img_info['width'], 1])
-        joints[..., 1] /= img_info['height']
-        orgsize = np.array([img_info['height'], img_info['width']])
        return joints, orgsize
+    def _get_bboxs(self, anno, idx):
+        num_people = len(anno)
+        gt_bboxes = np.zeros((num_people, 4), dtype=np.float32)
+        for idx, obj in enumerate(anno):
+            if 'bbox' in obj:
+                gt_bboxes[idx, :] = obj['bbox']
+        gt_bboxes[:, 2] += gt_bboxes[:, 0]
+        gt_bboxes[:, 3] += gt_bboxes[:, 1]
+        return gt_bboxes
+    def _get_labels(self, anno, idx):
+        num_people = len(anno)
+        gt_labels = np.zeros((num_people, 1), dtype=np.float32)
+        for idx, obj in enumerate(anno):
+            if 'category_id' in obj:
+                catid = obj['category_id']
+                gt_labels[idx, 0] = self.catid2clsid[catid]
+        return gt_labels
+    def _get_areas(self, anno, idx):
+        num_people = len(anno)
+        gt_areas = np.zeros((num_people, ), dtype=np.float32)
+        for idx, obj in enumerate(anno):
+            if 'area' in obj:
+                gt_areas[idx, ] = obj['area']
+        return gt_areas
    def _get_mask(self, anno, idx):
        """Get ignore masks to mask out losses."""
        coco = self.coco
@@ -506,7 +559,7 @@ class KeypointTopDownCocoDataset(KeypointTopDownBaseDataset):
                    'image_file': os.path.join(self.img_prefix, file_name),
                    'center': center,
                    'scale': scale,
-                    'joints': joints,
+                    'gt_joints': joints,
                    'joints_vis': joints_vis,
                    'im_id': im_id,
                })
@@ -570,7 +623,7 @@ class KeypointTopDownCocoDataset(KeypointTopDownBaseDataset):
                'center': center,
                'scale': scale,
                'score': score,
-                'joints': joints,
+                'gt_joints': joints,
                'joints_vis': joints_vis,
            })
@@ -647,8 +700,8 @@ class KeypointTopDownMPIIDataset(KeypointTopDownBaseDataset):
                (self.ann_info['num_joints'], 3), dtype=np.float32)
            joints_vis = np.zeros(
                (self.ann_info['num_joints'], 3), dtype=np.float32)
-            if 'joints' in a:
+            if 'gt_joints' in a:
-                joints_ = np.array(a['joints'])
+                joints_ = np.array(a['gt_joints'])
                joints_[:, 0:2] = joints_[:, 0:2] - 1
                joints_vis_ = np.array(a['joints_vis'])
                assert len(joints_) == self.ann_info[
@@ -664,7 +717,7 @@ class KeypointTopDownMPIIDataset(KeypointTopDownBaseDataset):
                'im_id': im_id,
                'center': c,
                'scale': s,
-                'joints': joints,
+                'gt_joints': joints,
                'joints_vis': joints_vis
            })
        print("number length: {}".format(len(gt_db)))

--- a/ppdet/data/transform/batch_operators.py
+++ b/ppdet/data/transform/batch_operators.py
@@ -1102,13 +1102,115 @@ class PadGT(BaseOperator):
                                1 means bbox, 0 means no bbox.
    """
-    def __init__(self, return_gt_mask=True):
+    def __init__(self, return_gt_mask=True, pad_img=False, minimum_gtnum=0):
        super(PadGT, self).__init__()
        self.return_gt_mask = return_gt_mask
+        self.pad_img = pad_img
+        self.minimum_gtnum = minimum_gtnum
+    def _impad(self, img: np.ndarray,
+            *,
+            shape = None,
+            padding = None,
+            pad_val = 0,
+            padding_mode = 'constant') -> np.ndarray:
+        """Pad the given image to a certain shape or pad on all sides with
+        specified padding mode and padding value.
+        Args:
+            img (ndarray): Image to be padded.
+            shape (tuple[int]): Expected padding shape (h, w). Default: None.
+            padding (int or tuple[int]): Padding on each border. If a single int is
+                provided this is used to pad all borders. If tuple of length 2 is
+                provided this is the padding on left/right and top/bottom
+                respectively. If a tuple of length 4 is provided this is the
+                padding for the left, top, right and bottom borders respectively.
+                Default: None. Note that `shape` and `padding` can not be both
+                set.
+            pad_val (Number | Sequence[Number]): Values to be filled in padding
+                areas when padding_mode is 'constant'. Default: 0.
+            padding_mode (str): Type of padding. Should be: constant, edge,
+                reflect or symmetric. Default: constant.
+                - constant: pads with a constant value, this value is specified
+                with pad_val.
+                - edge: pads with the last value at the edge of the image.
+                - reflect: pads with reflection of image without repeating the last
+                value on the edge. For example, padding [1, 2, 3, 4] with 2
+                elements on both sides in reflect mode will result in
+                [3, 2, 1, 2, 3, 4, 3, 2].
+                - symmetric: pads with reflection of image repeating the last value
+                on the edge. For example, padding [1, 2, 3, 4] with 2 elements on
+                both sides in symmetric mode will result in
+                [2, 1, 1, 2, 3, 4, 4, 3]
+        Returns:
+            ndarray: The padded image.
+        """
+        assert (shape is not None) ^ (padding is not None)
+        if shape is not None:
+            width = max(shape[1] - img.shape[1], 0)
+            height = max(shape[0] - img.shape[0], 0)
+            padding = (0, 0, int(width), int(height))
+        # check pad_val
+        import numbers
+        if isinstance(pad_val, tuple):
+            assert len(pad_val) == img.shape[-1]
+        elif not isinstance(pad_val, numbers.Number):
+            raise TypeError('pad_val must be a int or a tuple. '
+                            f'But received {type(pad_val)}')
+        # check padding
+        if isinstance(padding, tuple) and len(padding) in [2, 4]:
+            if len(padding) == 2:
+                padding = (padding[0], padding[1], padding[0], padding[1])
+        elif isinstance(padding, numbers.Number):
+            padding = (padding, padding, padding, padding)
+        else:
+            raise ValueError('Padding must be a int or a 2, or 4 element tuple.'
+                            f'But received {padding}')
+        # check padding mode
+        assert padding_mode in ['constant', 'edge', 'reflect', 'symmetric']
+        border_type = {
+            'constant': cv2.BORDER_CONSTANT,
+            'edge': cv2.BORDER_REPLICATE,
+            'reflect': cv2.BORDER_REFLECT_101,
+            'symmetric': cv2.BORDER_REFLECT
+        }
+        img = cv2.copyMakeBorder(
+            img,
+            padding[1],
+            padding[3],
+            padding[0],
+            padding[2],
+            border_type[padding_mode],
+            value=pad_val)
+        return img
+    def checkmaxshape(self, samples):
+        maxh, maxw = 0, 0
+        for sample in samples:
+            h,w = sample['im_shape']
+            if h>maxh:
+                maxh = h
+            if w>maxw:
+                maxw = w
+        return (maxh, maxw)
    def __call__(self, samples, context=None):
        num_max_boxes = max([len(s['gt_bbox']) for s in samples])
+        num_max_boxes = max(self.minimum_gtnum, num_max_boxes)
+        if self.pad_img:
+            maxshape = self.checkmaxshape(samples)
        for sample in samples:
+            if self.pad_img:
+                img = sample['image']
+                padimg = self._impad(img, shape=maxshape)
+                sample['image'] = padimg
            if self.return_gt_mask:
                sample['pad_gt_mask'] = np.zeros(
                    (num_max_boxes, 1), dtype=np.float32)
@@ -1142,6 +1244,17 @@ class PadGT(BaseOperator):
                if num_gt > 0:
                    pad_diff[:num_gt] = sample['difficult']
                sample['difficult'] = pad_diff
+            if 'gt_joints' in sample:
+                num_joints = sample['gt_joints'].shape[1]
+                pad_gt_joints = np.zeros((num_max_boxes, num_joints, 3), dtype=np.float32)
+                if num_gt > 0:
+                    pad_gt_joints[:num_gt] = sample['gt_joints']
+                sample['gt_joints'] = pad_gt_joints
+            if 'gt_areas' in sample:
+                pad_gt_areas = np.zeros((num_max_boxes, 1), dtype=np.float32)
+                if num_gt > 0:
+                    pad_gt_areas[:num_gt, 0] = sample['gt_areas']
+                sample['gt_areas'] = pad_gt_areas
        return samples

--- a/ppdet/data/transform/keypoint_operators.py
+++ b/ppdet/data/transform/keypoint_operators.py
@@ -41,7 +41,7 @@ __all__ = [
    'TopDownAffine', 'ToHeatmapsTopDown', 'ToHeatmapsTopDown_DARK',
    'ToHeatmapsTopDown_UDP', 'TopDownEvalAffine',
    'AugmentationbyInformantionDropping', 'SinglePoseAffine', 'NoiseJitter',
-    'FlipPose'
+    'FlipPose', 'PETR_Resize'
 ]
@@ -65,38 +65,77 @@ class KeyPointFlip(object):
    """
-    def __init__(self, flip_permutation, hmsize, flip_prob=0.5):
+    def __init__(self, flip_permutation, hmsize=None, flip_prob=0.5):
        super(KeyPointFlip, self).__init__()
        assert isinstance(flip_permutation, Sequence)
        self.flip_permutation = flip_permutation
        self.flip_prob = flip_prob
        self.hmsize = hmsize
-    def __call__(self, records):
+    def _flipjoints(self, records, sizelst):
-        image = records['image']
+        '''
-        kpts_lst = records['joints']
+        records['gt_joints'] is Sequence in higherhrnet
-        mask_lst = records['mask']
+        '''
-        flip = np.random.random() < self.flip_prob
+        if not ('gt_joints' in records and records['gt_joints'].size > 0):
-        if flip:
+            return records
-            image = image[:, ::-1]
-            for idx, hmsize in enumerate(self.hmsize):
+        kpts_lst = records['gt_joints']
-                if len(mask_lst) > idx:
+        if isinstance(kpts_lst, Sequence):
-                    mask_lst[idx] = mask_lst[idx][:, ::-1]
+            for idx, hmsize in enumerate(sizelst):
                if kpts_lst[idx].ndim == 3:
                    kpts_lst[idx] = kpts_lst[idx][:, self.flip_permutation]
                else:
                    kpts_lst[idx] = kpts_lst[idx][self.flip_permutation]
                kpts_lst[idx][..., 0] = hmsize - kpts_lst[idx][..., 0]
-                kpts_lst[idx] = kpts_lst[idx].astype(np.int64)
+        else:
-                kpts_lst[idx][kpts_lst[idx][..., 0] >= hmsize, 2] = 0
+            hmsize = sizelst[0]
-                kpts_lst[idx][kpts_lst[idx][..., 1] >= hmsize, 2] = 0
+            if kpts_lst.ndim == 3:
-                kpts_lst[idx][kpts_lst[idx][..., 0] < 0, 2] = 0
+                kpts_lst = kpts_lst[:, self.flip_permutation]
-                kpts_lst[idx][kpts_lst[idx][..., 1] < 0, 2] = 0
+            else:
-        records['image'] = image
+                kpts_lst = kpts_lst[self.flip_permutation]
-        records['joints'] = kpts_lst
+            kpts_lst[..., 0] = hmsize - kpts_lst[..., 0]
+        records['gt_joints'] = kpts_lst
+        return records
+    def _flipmask(self, records, sizelst):
+        if not 'mask' in records:
+            return records
+        mask_lst = records['mask']
+        for idx, hmsize in enumerate(sizelst):
+            if len(mask_lst) > idx:
+                mask_lst[idx] = mask_lst[idx][:, ::-1]
        records['mask'] = mask_lst
        return records
+    def _flipbbox(self, records, sizelst):
+        if not 'gt_bbox' in records:
+            return records
+        bboxes = records['gt_bbox']
+        hmsize = sizelst[0]
+        bboxes[:, 0::2] = hmsize - bboxes[:, 0::2][:, ::-1]
+        bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, hmsize)
+        records['gt_bbox'] = bboxes
+        return records
+    def __call__(self, records):
+        flip = np.random.random() < self.flip_prob
+        if flip:
+            image = records['image']
+            image = image[:, ::-1]
+            records['image'] = image
+            if self.hmsize is None:
+                sizelst = [image.shape[1]]
+            else:
+                sizelst = self.hmsize
+            self._flipjoints(records, sizelst)
+            self._flipmask(records, sizelst)
+            self._flipbbox(records, sizelst)
+        return records
 @register_keypointop
 class RandomAffine(object):
@@ -121,9 +160,10 @@ class RandomAffine(object):
                 max_degree=30,
                 scale=[0.75, 1.5],
                 max_shift=0.2,
-                 hmsize=[128, 256],
+                 hmsize=None,
                 trainsize=512,
-                 scale_type='short'):
+                 scale_type='short',
+                 boldervalue=[114, 114, 114]):
        super(RandomAffine, self).__init__()
        self.max_degree = max_degree
        self.min_scale = scale[0]
@@ -132,8 +172,9 @@ class RandomAffine(object):
        self.hmsize = hmsize
        self.trainsize = trainsize
        self.scale_type = scale_type
+        self.boldervalue = boldervalue
-    def _get_affine_matrix(self, center, scale, res, rot=0):
+    def _get_affine_matrix_old(self, center, scale, res, rot=0):
        """Generate transformation matrix."""
        h = scale
        t = np.zeros((3, 3), dtype=np.float32)
@@ -159,21 +200,94 @@ class RandomAffine(object):
            t = np.dot(t_inv, np.dot(rot_mat, np.dot(t_mat, t)))
        return t
+    def _get_affine_matrix(self, center, scale, res, rot=0):
+        """Generate transformation matrix."""
+        w, h = scale
+        t = np.zeros((3, 3), dtype=np.float32)
+        t[0, 0] = float(res[0]) / w
+        t[1, 1] = float(res[1]) / h
+        t[0, 2] = res[0] * (-float(center[0]) / w + .5)
+        t[1, 2] = res[1] * (-float(center[1]) / h + .5)
+        t[2, 2] = 1
+        if rot != 0:
+            rot = -rot  # To match direction of rotation from cropping
+            rot_mat = np.zeros((3, 3), dtype=np.float32)
+            rot_rad = rot * np.pi / 180
+            sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+            rot_mat[0, :2] = [cs, -sn]
+            rot_mat[1, :2] = [sn, cs]
+            rot_mat[2, 2] = 1
+            # Need to rotate around center
+            t_mat = np.eye(3)
+            t_mat[0, 2] = -res[0] / 2
+            t_mat[1, 2] = -res[1] / 2
+            t_inv = t_mat.copy()
+            t_inv[:2, 2] *= -1
+            t = np.dot(t_inv, np.dot(rot_mat, np.dot(t_mat, t)))
+        return t
+    def _affine_joints_mask(self,
+                            degree,
+                            center,
+                            roi_size,
+                            dsize,
+                            keypoints=None,
+                            heatmap_mask=None,
+                            gt_bbox=None):
+        kpts = None
+        mask = None
+        bbox = None
+        mask_affine_mat = self._get_affine_matrix(center, roi_size, dsize,
+                                                  degree)[:2]
+        if heatmap_mask is not None:
+            mask = cv2.warpAffine(heatmap_mask, mask_affine_mat, dsize)
+            mask = ((mask / 255) > 0.5).astype(np.float32)
+        if keypoints is not None:
+            kpts = copy.deepcopy(keypoints)
+            kpts[..., 0:2] = warp_affine_joints(kpts[..., 0:2].copy(),
+                                                mask_affine_mat)
+            kpts[(kpts[..., 0]) > dsize[0], :] = 0
+            kpts[(kpts[..., 1]) > dsize[1], :] = 0
+            kpts[(kpts[..., 0]) < 0, :] = 0
+            kpts[(kpts[..., 1]) < 0, :] = 0
+        if gt_bbox is not None:
+            temp_bbox = gt_bbox[:, [0, 3, 2, 1]]
+            cat_bbox = np.concatenate((gt_bbox, temp_bbox), axis=-1)
+            gt_bbox_warped = warp_affine_joints(cat_bbox, mask_affine_mat)
+            bbox = np.zeros_like(gt_bbox)
+            bbox[:, 0] = gt_bbox_warped[:, 0::2].min(1).clip(0, dsize[0])
+            bbox[:, 2] = gt_bbox_warped[:, 0::2].max(1).clip(0, dsize[0])
+            bbox[:, 1] = gt_bbox_warped[:, 1::2].min(1).clip(0, dsize[1])
+            bbox[:, 3] = gt_bbox_warped[:, 1::2].max(1).clip(0, dsize[1])
+        return kpts, mask, bbox
    def __call__(self, records):
        image = records['image']
-        keypoints = records['joints']
+        shape = np.array(image.shape[:2][::-1])
-        heatmap_mask = records['mask']
+        keypoints = None
+        heatmap_mask = None
+        gt_bbox = None
+        if 'gt_joints' in records:
+            keypoints = records['gt_joints']
+        if 'mask' in records:
+            heatmap_mask = records['mask']
+            heatmap_mask *= 255
+        if 'gt_bbox' in records:
+            gt_bbox = records['gt_bbox']
        degree = (np.random.random() * 2 - 1) * self.max_degree
-        shape = np.array(image.shape[:2][::-1])
        center = center = np.array((np.array(shape) / 2))
        aug_scale = np.random.random() * (self.max_scale - self.min_scale
                                          ) + self.min_scale
        if self.scale_type == 'long':
-            scale = max(shape[0], shape[1]) / 1.0
+            scale = np.array([max(shape[0], shape[1]) / 1.0] * 2)
        elif self.scale_type == 'short':
-            scale = min(shape[0], shape[1]) / 1.0
+            scale = np.array([min(shape[0], shape[1]) / 1.0] * 2)
+        elif self.scale_type == 'wh':
+            scale = shape
        else:
            raise ValueError('Unknown scale type: {}'.format(self.scale_type))
        roi_size = aug_scale * scale
@@ -181,44 +295,55 @@ class RandomAffine(object):
        dy = int(0)
        if self.max_shift > 0:
-            dx = np.random.randint(-self.max_shift * roi_size,
+            dx = np.random.randint(-self.max_shift * roi_size[0],
-                                   self.max_shift * roi_size)
+                                   self.max_shift * roi_size[0])
-            dy = np.random.randint(-self.max_shift * roi_size,
+            dy = np.random.randint(-self.max_shift * roi_size[0],
-                                   self.max_shift * roi_size)
+                                   self.max_shift * roi_size[1])
        center += np.array([dx, dy])
        input_size = 2 * center
+        if self.trainsize != -1:
+            dsize = self.trainsize
+            imgshape = (dsize, dsize)
+        else:
+            dsize = scale
+            imgshape = (shape.tolist())
-        keypoints[..., :2] *= shape
+        image_affine_mat = self._get_affine_matrix(center, roi_size, dsize,
-        heatmap_mask *= 255
+                                                   degree)[:2]
-        kpts_lst = []
-        mask_lst = []
-        image_affine_mat = self._get_affine_matrix(
-            center, roi_size, (self.trainsize, self.trainsize), degree)[:2]
        image = cv2.warpAffine(
            image,
-            image_affine_mat, (self.trainsize, self.trainsize),
+            image_affine_mat,
-            flags=cv2.INTER_LINEAR)
+            imgshape,
+            flags=cv2.INTER_LINEAR,
+            borderValue=self.boldervalue)
+        if self.hmsize is None:
+            kpts, mask, gt_bbox = self._affine_joints_mask(
+                degree, center, roi_size, dsize, keypoints, heatmap_mask,
+                gt_bbox)
+            records['image'] = image
+            if kpts is not None: records['gt_joints'] = kpts
+            if mask is not None: records['mask'] = mask
+            if gt_bbox is not None: records['gt_bbox'] = gt_bbox
+            return records
+        kpts_lst = []
+        mask_lst = []
        for hmsize in self.hmsize:
-            kpts = copy.deepcopy(keypoints)
+            kpts, mask, gt_bbox = self._affine_joints_mask(
-            mask_affine_mat = self._get_affine_matrix(
+                degree, center, roi_size, [hmsize, hmsize], keypoints,
-                center, roi_size, (hmsize, hmsize), degree)[:2]
+                heatmap_mask, gt_bbox)
-            if heatmap_mask is not None:
-                mask = cv2.warpAffine(heatmap_mask, mask_affine_mat,
-                                      (hmsize, hmsize))
-                mask = ((mask / 255) > 0.5).astype(np.float32)
-            kpts[..., 0:2] = warp_affine_joints(kpts[..., 0:2].copy(),
-                                                mask_affine_mat)
-            kpts[np.trunc(kpts[..., 0]) >= hmsize, 2] = 0
-            kpts[np.trunc(kpts[..., 1]) >= hmsize, 2] = 0
-            kpts[np.trunc(kpts[..., 0]) < 0, 2] = 0
-            kpts[np.trunc(kpts[..., 1]) < 0, 2] = 0
            kpts_lst.append(kpts)
            mask_lst.append(mask)
        records['image'] = image
-        records['joints'] = kpts_lst
-        records['mask'] = mask_lst
+        if 'gt_joints' in records:
+            records['gt_joints'] = kpts_lst
+        if 'mask' in records:
+            records['mask'] = mask_lst
+        if 'gt_bbox' in records:
+            records['gt_bbox'] = gt_bbox
        return records
@@ -251,8 +376,8 @@ class EvalAffine(object):
        if mask is not None:
            mask = cv2.warpAffine(mask, trans, size_resized)
            records['mask'] = mask
-        if 'joints' in records:
+        if 'gt_joints' in records:
-            del records['joints']
+            del records['gt_joints']
        records['image'] = image_resized
        return records
@@ -303,7 +428,7 @@ class TagGenerate(object):
        self.num_joints = num_joints
    def __call__(self, records):
-        kpts_lst = records['joints']
+        kpts_lst = records['gt_joints']
        kpts = kpts_lst[0]
        tagmap = np.zeros((self.max_people, self.num_joints, 4), dtype=np.int64)
        inds = np.where(kpts[..., 2] > 0)
@@ -315,7 +440,7 @@ class TagGenerate(object):
        tagmap[p, j, 2] = visible[..., 0]  # x
        tagmap[p, j, 3] = 1
        records['tagmap'] = tagmap
-        del records['joints']
+        del records['gt_joints']
        return records
@@ -349,7 +474,7 @@ class ToHeatmaps(object):
        self.gaussian = np.exp(-((x - x0)**2 + (y - y0)**2) / (2 * sigma**2))
    def __call__(self, records):
-        kpts_lst = records['joints']
+        kpts_lst = records['gt_joints']
        mask_lst = records['mask']
        for idx, hmsize in enumerate(self.hmsize):
            mask = mask_lst[idx]
@@ -470,7 +595,7 @@ class RandomFlipHalfBodyTransform(object):
    def __call__(self, records):
        image = records['image']
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        c = records['center']
        s = records['scale']
@@ -493,7 +618,7 @@ class RandomFlipHalfBodyTransform(object):
                joints, joints_vis, image.shape[1], self.flip_pairs)
            c[0] = image.shape[1] - c[0] - 1
        records['image'] = image
-        records['joints'] = joints
+        records['gt_joints'] = joints
        records['joints_vis'] = joints_vis
        records['center'] = c
        records['scale'] = s
@@ -553,7 +678,7 @@ class AugmentationbyInformantionDropping(object):
    def __call__(self, records):
        img = records['image']
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        if np.random.rand() < self.prob_cutout:
            img = self._cutout(img, joints, joints_vis)
@@ -581,7 +706,7 @@ class TopDownAffine(object):
    def __call__(self, records):
        image = records['image']
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        rot = records['rotate'] if "rotate" in records else 0
        if self.use_udp:
@@ -606,7 +731,7 @@ class TopDownAffine(object):
                    joints[i, 0:2] = affine_transform(joints[i, 0:2], trans)
        records['image'] = image
-        records['joints'] = joints
+        records['gt_joints'] = joints
        return records
@@ -842,7 +967,7 @@ class ToHeatmapsTopDown(object):
            https://github.com/leoxiaobin/deep-high-resolution-net.pytorch
            Copyright (c) Microsoft, under the MIT License.
        """
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        num_joints = joints.shape[0]
        image_size = np.array(
@@ -885,7 +1010,7 @@ class ToHeatmapsTopDown(object):
                    0]:g_y[1], g_x[0]:g_x[1]]
        records['target'] = target
        records['target_weight'] = target_weight
-        del records['joints'], records['joints_vis']
+        del records['gt_joints'], records['joints_vis']
        return records
@@ -910,7 +1035,7 @@ class ToHeatmapsTopDown_DARK(object):
        self.sigma = sigma
    def __call__(self, records):
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        num_joints = joints.shape[0]
        image_size = np.array(
@@ -943,7 +1068,7 @@ class ToHeatmapsTopDown_DARK(object):
                    (x - mu_x)**2 + (y - mu_y)**2) / (2 * self.sigma**2))
        records['target'] = target
        records['target_weight'] = target_weight
-        del records['joints'], records['joints_vis']
+        del records['gt_joints'], records['joints_vis']
        return records
@@ -972,7 +1097,7 @@ class ToHeatmapsTopDown_UDP(object):
        self.sigma = sigma
    def __call__(self, records):
-        joints = records['joints']
+        joints = records['gt_joints']
        joints_vis = records['joints_vis']
        num_joints = joints.shape[0]
        image_size = np.array(
@@ -1017,6 +1142,472 @@ class ToHeatmapsTopDown_UDP(object):
                    0]:g_y[1], g_x[0]:g_x[1]]
        records['target'] = target
        records['target_weight'] = target_weight
-        del records['joints'], records['joints_vis']
+        del records['gt_joints'], records['joints_vis']
        return records
+from typing import Optional, Tuple, Union, List
+import numbers
+def _scale_size(
+        size: Tuple[int, int],
+        scale: Union[float, int, tuple], ) -> Tuple[int, int]:
+    """Rescale a size by a ratio.
+    Args:
+        size (tuple[int]): (w, h).
+        scale (float | tuple(float)): Scaling factor.
+    Returns:
+        tuple[int]: scaled size.
+    """
+    if isinstance(scale, (float, int)):
+        scale = (scale, scale)
+    w, h = size
+    return int(w * float(scale[0]) + 0.5), int(h * float(scale[1]) + 0.5)
+def rescale_size(old_size: tuple,
+                 scale: Union[float, int, tuple],
+                 return_scale: bool=False) -> tuple:
+    """Calculate the new size to be rescaled to.
+    Args:
+        old_size (tuple[int]): The old size (w, h) of image.
+        scale (float | tuple[int]): The scaling factor or maximum size.
+            If it is a float number, then the image will be rescaled by this
+            factor, else if it is a tuple of 2 integers, then the image will
+            be rescaled as large as possible within the scale.
+        return_scale (bool): Whether to return the scaling factor besides the
+            rescaled image size.
+    Returns:
+        tuple[int]: The new rescaled image size.
+    """
+    w, h = old_size
+    if isinstance(scale, (float, int)):
+        if scale <= 0:
+            raise ValueError(f'Invalid scale {scale}, must be positive.')
+        scale_factor = scale
+    elif isinstance(scale, list):
+        max_long_edge = max(scale)
+        max_short_edge = min(scale)
+        scale_factor = min(max_long_edge / max(h, w),
+                           max_short_edge / min(h, w))
+    else:
+        raise TypeError(
+            f'Scale must be a number or tuple of int, but got {type(scale)}')
+    new_size = _scale_size((w, h), scale_factor)
+    if return_scale:
+        return new_size, scale_factor
+    else:
+        return new_size
+def imrescale(img: np.ndarray,
+              scale: Union[float, Tuple[int, int]],
+              return_scale: bool=False,
+              interpolation: str='bilinear',
+              backend: Optional[str]=None) -> Union[np.ndarray, Tuple[
+                  np.ndarray, float]]:
+    """Resize image while keeping the aspect ratio.
+    Args:
+        img (ndarray): The input image.
+        scale (float | tuple[int]): The scaling factor or maximum size.
+            If it is a float number, then the image will be rescaled by this
+            factor, else if it is a tuple of 2 integers, then the image will
+            be rescaled as large as possible within the scale.
+        return_scale (bool): Whether to return the scaling factor besides the
+            rescaled image.
+        interpolation (str): Same as :func:`resize`.
+        backend (str | None): Same as :func:`resize`.
+    Returns:
+        ndarray: The rescaled image.
+    """
+    h, w = img.shape[:2]
+    new_size, scale_factor = rescale_size((w, h), scale, return_scale=True)
+    rescaled_img = imresize(
+        img, new_size, interpolation=interpolation, backend=backend)
+    if return_scale:
+        return rescaled_img, scale_factor
+    else:
+        return rescaled_img
+def imresize(
+        img: np.ndarray,
+        size: Tuple[int, int],
+        return_scale: bool=False,
+        interpolation: str='bilinear',
+        out: Optional[np.ndarray]=None,
+        backend: Optional[str]=None,
+        interp=cv2.INTER_LINEAR, ) -> Union[Tuple[np.ndarray, float, float],
+                                            np.ndarray]:
+    """Resize image to a given size.
+    Args:
+        img (ndarray): The input image.
+        size (tuple[int]): Target size (w, h).
+        return_scale (bool): Whether to return `w_scale` and `h_scale`.
+        interpolation (str): Interpolation method, accepted values are
+            "nearest", "bilinear", "bicubic", "area", "lanczos" for 'cv2'
+            backend, "nearest", "bilinear" for 'pillow' backend.
+        out (ndarray): The output destination.
+        backend (str | None): The image resize backend type. Options are `cv2`,
+            `pillow`, `None`. If backend is None, the global imread_backend
+            specified by ``mmcv.use_backend()`` will be used. Default: None.
+    Returns:
+        tuple | ndarray: (`resized_img`, `w_scale`, `h_scale`) or
+        `resized_img`.
+    """
+    h, w = img.shape[:2]
+    if backend is None:
+        backend = imread_backend
+    if backend not in ['cv2', 'pillow']:
+        raise ValueError(f'backend: {backend} is not supported for resize.'
+                         f"Supported backends are 'cv2', 'pillow'")
+    if backend == 'pillow':
+        assert img.dtype == np.uint8, 'Pillow backend only support uint8 type'
+        pil_image = Image.fromarray(img)
+        pil_image = pil_image.resize(size, pillow_interp_codes[interpolation])
+        resized_img = np.array(pil_image)
+    else:
+        resized_img = cv2.resize(img, size, dst=out, interpolation=interp)
+    if not return_scale:
+        return resized_img
+    else:
+        w_scale = size[0] / w
+        h_scale = size[1] / h
+        return resized_img, w_scale, h_scale
+class PETR_Resize:
+    """Resize images & bbox & mask.
+    This transform resizes the input image to some scale. Bboxes and masks are
+    then resized with the same scale factor. If the input dict contains the key
+    "scale", then the scale in the input dict is used, otherwise the specified
+    scale in the init method is used. If the input dict contains the key
+    "scale_factor" (if MultiScaleFlipAug does not give img_scale but
+    scale_factor), the actual scale will be computed by image shape and
+    scale_factor.
+    `img_scale` can either be a tuple (single-scale) or a list of tuple
+    (multi-scale). There are 3 multiscale modes:
+    - ``ratio_range is not None``: randomly sample a ratio from the ratio \
+      range and multiply it with the image scale.
+    - ``ratio_range is None`` and ``multiscale_mode == "range"``: randomly \
+      sample a scale from the multiscale range.
+    - ``ratio_range is None`` and ``multiscale_mode == "value"``: randomly \
+      sample a scale from multiple scales.
+    Args:
+        img_scale (tuple or list[tuple]): Images scales for resizing.
+        multiscale_mode (str): Either "range" or "value".
+        ratio_range (tuple[float]): (min_ratio, max_ratio)
+        keep_ratio (bool): Whether to keep the aspect ratio when resizing the
+            image.
+        bbox_clip_border (bool, optional): Whether to clip the objects outside
+            the border of the image. In some dataset like MOT17, the gt bboxes
+            are allowed to cross the border of images. Therefore, we don't
+            need to clip the gt bboxes in these cases. Defaults to True.
+        backend (str): Image resize backend, choices are 'cv2' and 'pillow'.
+            These two backends generates slightly different results. Defaults
+            to 'cv2'.
+        interpolation (str): Interpolation method, accepted values are
+            "nearest", "bilinear", "bicubic", "area", "lanczos" for 'cv2'
+            backend, "nearest", "bilinear" for 'pillow' backend.
+        override (bool, optional): Whether to override `scale` and
+            `scale_factor` so as to call resize twice. Default False. If True,
+            after the first resizing, the existed `scale` and `scale_factor`
+            will be ignored so the second resizing can be allowed.
+            This option is a work-around for multiple times of resize in DETR.
+            Defaults to False.
+    """
+    def __init__(self,
+                 img_scale=None,
+                 multiscale_mode='range',
+                 ratio_range=None,
+                 keep_ratio=True,
+                 bbox_clip_border=True,
+                 backend='cv2',
+                 interpolation='bilinear',
+                 override=False,
+                 keypoint_clip_border=True):
+        if img_scale is None:
+            self.img_scale = None
+        else:
+            if isinstance(img_scale, list):
+                self.img_scale = img_scale
+            else:
+                self.img_scale = [img_scale]
+            assert isinstance(self.img_scale, list)
+        if ratio_range is not None:
+            # mode 1: given a scale and a range of image ratio
+            assert len(self.img_scale) == 1
+        else:
+            # mode 2: given multiple scales or a range of scales
+            assert multiscale_mode in ['value', 'range']
+        self.backend = backend
+        self.multiscale_mode = multiscale_mode
+        self.ratio_range = ratio_range
+        self.keep_ratio = keep_ratio
+        # TODO: refactor the override option in Resize
+        self.interpolation = interpolation
+        self.override = override
+        self.bbox_clip_border = bbox_clip_border
+        self.keypoint_clip_border = keypoint_clip_border
+    @staticmethod
+    def random_select(img_scales):
+        """Randomly select an img_scale from given candidates.
+        Args:
+            img_scales (list[tuple]): Images scales for selection.
+        Returns:
+            (tuple, int): Returns a tuple ``(img_scale, scale_dix)``, \
+                where ``img_scale`` is the selected image scale and \
+                ``scale_idx`` is the selected index in the given candidates.
+        """
+        assert isinstance(img_scales, list)
+        scale_idx = np.random.randint(len(img_scales))
+        img_scale = img_scales[scale_idx]
+        return img_scale, scale_idx
+    @staticmethod
+    def random_sample(img_scales):
+        """Randomly sample an img_scale when ``multiscale_mode=='range'``.
+        Args:
+            img_scales (list[tuple]): Images scale range for sampling.
+                There must be two tuples in img_scales, which specify the lower
+                and upper bound of image scales.
+        Returns:
+            (tuple, None): Returns a tuple ``(img_scale, None)``, where \
+                ``img_scale`` is sampled scale and None is just a placeholder \
+                to be consistent with :func:`random_select`.
+        """
+        assert isinstance(img_scales, list) and len(img_scales) == 2
+        img_scale_long = [max(s) for s in img_scales]
+        img_scale_short = [min(s) for s in img_scales]
+        long_edge = np.random.randint(
+            min(img_scale_long), max(img_scale_long) + 1)
+        short_edge = np.random.randint(
+            min(img_scale_short), max(img_scale_short) + 1)
+        img_scale = (long_edge, short_edge)
+        return img_scale, None
+    @staticmethod
+    def random_sample_ratio(img_scale, ratio_range):
+        """Randomly sample an img_scale when ``ratio_range`` is specified.
+        A ratio will be randomly sampled from the range specified by
+        ``ratio_range``. Then it would be multiplied with ``img_scale`` to
+        generate sampled scale.
+        Args:
+            img_scale (list): Images scale base to multiply with ratio.
+            ratio_range (tuple[float]): The minimum and maximum ratio to scale
+                the ``img_scale``.
+        Returns:
+            (tuple, None): Returns a tuple ``(scale, None)``, where \
+                ``scale`` is sampled ratio multiplied with ``img_scale`` and \
+                None is just a placeholder to be consistent with \
+                :func:`random_select`.
+        """
+        assert isinstance(img_scale, list) and len(img_scale) == 2
+        min_ratio, max_ratio = ratio_range
+        assert min_ratio <= max_ratio
+        ratio = np.random.random_sample() * (max_ratio - min_ratio) + min_ratio
+        scale = int(img_scale[0] * ratio), int(img_scale[1] * ratio)
+        return scale, None
+    def _random_scale(self, results):
+        """Randomly sample an img_scale according to ``ratio_range`` and
+        ``multiscale_mode``.
+        If ``ratio_range`` is specified, a ratio will be sampled and be
+        multiplied with ``img_scale``.
+        If multiple scales are specified by ``img_scale``, a scale will be
+        sampled according to ``multiscale_mode``.
+        Otherwise, single scale will be used.
+        Args:
+            results (dict): Result dict from :obj:`dataset`.
+        Returns:
+            dict: Two new keys 'scale` and 'scale_idx` are added into \
+                ``results``, which would be used by subsequent pipelines.
+        """
+        if self.ratio_range is not None:
+            scale, scale_idx = self.random_sample_ratio(self.img_scale[0],
+                                                        self.ratio_range)
+        elif len(self.img_scale) == 1:
+            scale, scale_idx = self.img_scale[0], 0
+        elif self.multiscale_mode == 'range':
+            scale, scale_idx = self.random_sample(self.img_scale)
+        elif self.multiscale_mode == 'value':
+            scale, scale_idx = self.random_select(self.img_scale)
+        else:
+            raise NotImplementedError
+        results['scale'] = scale
+        results['scale_idx'] = scale_idx
+    def _resize_img(self, results):
+        """Resize images with ``results['scale']``."""
+        for key in ['image'] if 'image' in results else []:
+            if self.keep_ratio:
+                img, scale_factor = imrescale(
+                    results[key],
+                    results['scale'],
+                    return_scale=True,
+                    interpolation=self.interpolation,
+                    backend=self.backend)
+                # the w_scale and h_scale has minor difference
+                # a real fix should be done in the imrescale in the future
+                new_h, new_w = img.shape[:2]
+                h, w = results[key].shape[:2]
+                w_scale = new_w / w
+                h_scale = new_h / h
+            else:
+                img, w_scale, h_scale = imresize(
+                    results[key],
+                    results['scale'],
+                    return_scale=True,
+                    interpolation=self.interpolation,
+                    backend=self.backend)
+            scale_factor = np.array(
+                [w_scale, h_scale, w_scale, h_scale], dtype=np.float32)
+            results['im_shape'] = np.array(img.shape)
+            # in case that there is no padding
+            results['pad_shape'] = img.shape
+            results['scale_factor'] = scale_factor
+            results['keep_ratio'] = self.keep_ratio
+            # img_pad = self.impad(img, shape=results['scale'])
+            results[key] = img
+    def _resize_bboxes(self, results):
+        """Resize bounding boxes with ``results['scale_factor']``."""
+        for key in ['gt_bbox'] if 'gt_bbox' in results else []:
+            bboxes = results[key] * results['scale_factor']
+            if self.bbox_clip_border:
+                img_shape = results['im_shape']
+                bboxes[:, 0::2] = np.clip(bboxes[:, 0::2], 0, img_shape[1])
+                bboxes[:, 1::2] = np.clip(bboxes[:, 1::2], 0, img_shape[0])
+            results[key] = bboxes
+    def _resize_masks(self, results):
+        """Resize masks with ``results['scale']``"""
+        for key in ['mask'] if 'mask' in results else []:
+            if results[key] is None:
+                continue
+            if self.keep_ratio:
+                results[key] = results[key].rescale(results['scale'])
+            else:
+                results[key] = results[key].resize(results['im_shape'][:2])
+    def _resize_seg(self, results):
+        """Resize semantic segmentation map with ``results['scale']``."""
+        for key in ['seg'] if 'seg' in results else []:
+            if self.keep_ratio:
+                gt_seg = imrescale(
+                    results[key],
+                    results['scale'],
+                    interpolation='nearest',
+                    backend=self.backend)
+            else:
+                gt_seg = imresize(
+                    results[key],
+                    results['scale'],
+                    interpolation='nearest',
+                    backend=self.backend)
+            results[key] = gt_seg
+    def _resize_keypoints(self, results):
+        """Resize keypoints with ``results['scale_factor']``."""
+        for key in ['gt_joints'] if 'gt_joints' in results else []:
+            keypoints = results[key].copy()
+            keypoints[..., 0] = keypoints[..., 0] * results['scale_factor'][0]
+            keypoints[..., 1] = keypoints[..., 1] * results['scale_factor'][1]
+            if self.keypoint_clip_border:
+                img_shape = results['im_shape']
+                keypoints[..., 0] = np.clip(keypoints[..., 0], 0, img_shape[1])
+                keypoints[..., 1] = np.clip(keypoints[..., 1], 0, img_shape[0])
+            results[key] = keypoints
+    def _resize_areas(self, results):
+        """Resize mask areas with ``results['scale_factor']``."""
+        for key in ['gt_areas'] if 'gt_areas' in results else []:
+            areas = results[key].copy()
+            areas = areas * results['scale_factor'][0] * results[
+                'scale_factor'][1]
+            results[key] = areas
+    def __call__(self, results):
+        """Call function to resize images, bounding boxes, masks, semantic
+        segmentation map.
+        Args:
+            results (dict): Result dict from loading pipeline.
+        Returns:
+            dict: Resized results, 'im_shape', 'pad_shape', 'scale_factor', \
+                'keep_ratio' keys are added into result dict.
+        """
+        if 'scale' not in results:
+            if 'scale_factor' in results:
+                img_shape = results['image'].shape[:2]
+                scale_factor = results['scale_factor']
+                assert isinstance(scale_factor, float)
+                results['scale'] = tuple(
+                    [int(x * scale_factor) for x in img_shape][::-1])
+            else:
+                self._random_scale(results)
+        else:
+            if not self.override:
+                assert 'scale_factor' not in results, (
+                    'scale and scale_factor cannot be both set.')
+            else:
+                results.pop('scale')
+                if 'scale_factor' in results:
+                    results.pop('scale_factor')
+                self._random_scale(results)
+        self._resize_img(results)
+        self._resize_bboxes(results)
+        self._resize_masks(results)
+        self._resize_seg(results)
+        self._resize_keypoints(results)
+        self._resize_areas(results)
+        return results
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(img_scale={self.img_scale}, '
+        repr_str += f'multiscale_mode={self.multiscale_mode}, '
+        repr_str += f'ratio_range={self.ratio_range}, '
+        repr_str += f'keep_ratio={self.keep_ratio}, '
+        repr_str += f'bbox_clip_border={self.bbox_clip_border})'
+        repr_str += f'keypoint_clip_border={self.keypoint_clip_border})'
+        return repr_str
--- a/ppdet/data/transform/operators.py
+++ b/ppdet/data/transform/operators.py
@@ -594,6 +594,108 @@ class RandomDistort(BaseOperator):
        return sample
+@register_op
+class PhotoMetricDistortion(BaseOperator):
+    """Apply photometric distortion to image sequentially, every transformation
+    is applied with a probability of 0.5. The position of random contrast is in
+    second or second to last.
+    1. random brightness
+    2. random contrast (mode 0)
+    3. convert color from BGR to HSV
+    4. random saturation
+    5. random hue
+    6. convert color from HSV to BGR
+    7. random contrast (mode 1)
+    8. randomly swap channels
+    Args:
+        brightness_delta (int): delta of brightness.
+        contrast_range (tuple): range of contrast.
+        saturation_range (tuple): range of saturation.
+        hue_delta (int): delta of hue.
+    """
+    def __init__(self,
+                 brightness_delta=32,
+                 contrast_range=(0.5, 1.5),
+                 saturation_range=(0.5, 1.5),
+                 hue_delta=18):
+        super(PhotoMetricDistortion, self).__init__()
+        self.brightness_delta = brightness_delta
+        self.contrast_lower, self.contrast_upper = contrast_range
+        self.saturation_lower, self.saturation_upper = saturation_range
+        self.hue_delta = hue_delta
+    def apply(self, results, context=None):
+        """Call function to perform photometric distortion on images.
+        Args:
+            results (dict): Result dict from loading pipeline.
+        Returns:
+            dict: Result dict with images distorted.
+        """
+        img = results['image']
+        img = img.astype(np.float32)
+        # random brightness
+        if np.random.randint(2):
+            delta = np.random.uniform(-self.brightness_delta,
+                                      self.brightness_delta)
+            img += delta
+        # mode == 0 --> do random contrast first
+        # mode == 1 --> do random contrast last
+        mode = np.random.randint(2)
+        if mode == 1:
+            if np.random.randint(2):
+                alpha = np.random.uniform(self.contrast_lower,
+                                          self.contrast_upper)
+                img *= alpha
+        # convert color from BGR to HSV
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
+        # random saturation
+        if np.random.randint(2):
+            img[..., 1] *= np.random.uniform(self.saturation_lower,
+                                             self.saturation_upper)
+        # random hue
+        if np.random.randint(2):
+            img[..., 0] += np.random.uniform(-self.hue_delta, self.hue_delta)
+            img[..., 0][img[..., 0] > 360] -= 360
+            img[..., 0][img[..., 0] < 0] += 360
+        # convert color from HSV to BGR
+        img = cv2.cvtColor(img, cv2.COLOR_HSV2BGR)
+        # random contrast
+        if mode == 0:
+            if np.random.randint(2):
+                alpha = np.random.uniform(self.contrast_lower,
+                                          self.contrast_upper)
+                img *= alpha
+        # randomly swap channels
+        if np.random.randint(2):
+            img = img[..., np.random.permutation(3)]
+        results['image'] = img
+        return results
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += f'(\nbrightness_delta={self.brightness_delta},\n'
+        repr_str += 'contrast_range='
+        repr_str += f'{(self.contrast_lower, self.contrast_upper)},\n'
+        repr_str += 'saturation_range='
+        repr_str += f'{(self.saturation_lower, self.saturation_upper)},\n'
+        repr_str += f'hue_delta={self.hue_delta})'
+        return repr_str
 @register_op
 class AutoAugment(BaseOperator):
    def __init__(self, autoaug_type="v1"):
@@ -771,6 +873,19 @@ class Resize(BaseOperator):
        bbox[:, 1::2] = np.clip(bbox[:, 1::2], 0, resize_h)
        return bbox
+    def apply_area(self, area, scale):
+        im_scale_x, im_scale_y = scale
+        return area * im_scale_x * im_scale_y
+    def apply_joints(self, joints, scale, size):
+        im_scale_x, im_scale_y = scale
+        resize_w, resize_h = size
+        joints[..., 0] *= im_scale_x
+        joints[..., 1] *= im_scale_y
+        joints[..., 0] = np.clip(joints[..., 0], 0, resize_w)
+        joints[..., 1] = np.clip(joints[..., 1], 0, resize_h)
+        return joints
    def apply_segm(self, segms, im_size, scale):
        def _resize_poly(poly, im_scale_x, im_scale_y):
            resized_poly = np.array(poly).astype('float32')
@@ -833,8 +948,8 @@ class Resize(BaseOperator):
            im_scale = min(target_size_min / im_size_min,
                           target_size_max / im_size_max)
-            resize_h = im_scale * float(im_shape[0])
+            resize_h = int(im_scale * float(im_shape[0]) + 0.5)
-            resize_w = im_scale * float(im_shape[1])
+            resize_w = int(im_scale * float(im_shape[1]) + 0.5)
            im_scale_x = im_scale
            im_scale_y = im_scale
@@ -878,6 +993,11 @@ class Resize(BaseOperator):
                                                [im_scale_x, im_scale_y],
                                                [resize_w, resize_h])
+        # apply areas
+        if 'gt_areas' in sample:
+            sample['gt_areas'] = self.apply_area(sample['gt_areas'],
+                                                 [im_scale_x, im_scale_y])
        # apply polygon
        if 'gt_poly' in sample and len(sample['gt_poly']) > 0:
            sample['gt_poly'] = self.apply_segm(sample['gt_poly'], im_shape[:2],
@@ -911,6 +1031,11 @@ class Resize(BaseOperator):
            ]
            sample['gt_segm'] = np.asarray(masks).astype(np.uint8)
+        if 'gt_joints' in sample:
+            sample['gt_joints'] = self.apply_joints(sample['gt_joints'],
+                                                    [im_scale_x, im_scale_y],
+                                                    [resize_w, resize_h])
        return sample
@@ -1362,7 +1487,8 @@ class RandomCrop(BaseOperator):
                 num_attempts=50,
                 allow_no_crop=True,
                 cover_all_box=False,
-                 is_mask_crop=False):
+                 is_mask_crop=False,
+                 ioumode="iou"):
        super(RandomCrop, self).__init__()
        self.aspect_ratio = aspect_ratio
        self.thresholds = thresholds
@@ -1371,6 +1497,7 @@ class RandomCrop(BaseOperator):
        self.allow_no_crop = allow_no_crop
        self.cover_all_box = cover_all_box
        self.is_mask_crop = is_mask_crop
+        self.ioumode = ioumode
    def crop_segms(self, segms, valid_ids, crop, height, width):
        def _crop_poly(segm, crop):
@@ -1516,9 +1643,14 @@ class RandomCrop(BaseOperator):
                crop_y = np.random.randint(0, h - crop_h)
                crop_x = np.random.randint(0, w - crop_w)
                crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
-                iou = self._iou_matrix(
+                if self.ioumode == "iof":
-                    gt_bbox, np.array(
+                    iou = self._gtcropiou_matrix(
-                        [crop_box], dtype=np.float32))
+                        gt_bbox, np.array(
+                            [crop_box], dtype=np.float32))
+                elif self.ioumode == "iou":
+                    iou = self._iou_matrix(
+                        gt_bbox, np.array(
+                            [crop_box], dtype=np.float32))
                if iou.max() < thresh:
                    continue
@@ -1582,6 +1714,10 @@ class RandomCrop(BaseOperator):
                    sample['difficult'] = np.take(
                        sample['difficult'], valid_ids, axis=0)
+                if 'gt_joints' in sample:
+                    sample['gt_joints'] = self._crop_joints(sample['gt_joints'],
+                                                            crop_box)
                return sample
        return sample
@@ -1596,6 +1732,16 @@ class RandomCrop(BaseOperator):
        area_o = (area_a[:, np.newaxis] + area_b - area_i)
        return area_i / (area_o + 1e-10)
+    def _gtcropiou_matrix(self, a, b):
+        tl_i = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+        br_i = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+        area_i = np.prod(br_i - tl_i, axis=2) * (tl_i < br_i).all(axis=2)
+        area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+        area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
+        area_o = (area_a[:, np.newaxis] + area_b - area_i)
+        return area_i / (area_a + 1e-10)
    def _crop_box_with_center_constraint(self, box, crop):
        cropped_box = box.copy()
@@ -1620,6 +1766,16 @@ class RandomCrop(BaseOperator):
        x1, y1, x2, y2 = crop
        return segm[:, y1:y2, x1:x2]
+    def _crop_joints(self, joints, crop):
+        x1, y1, x2, y2 = crop
+        joints[joints[..., 0] > x2, :] = 0
+        joints[joints[..., 1] > y2, :] = 0
+        joints[joints[..., 0] < x1, :] = 0
+        joints[joints[..., 1] < y1, :] = 0
+        joints[..., 0] -= x1
+        joints[..., 1] -= y1
+        return joints
 @register_op
 class RandomScaledCrop(BaseOperator):
@@ -1648,8 +1804,8 @@ class RandomScaledCrop(BaseOperator):
        random_dim = int(dim * random_scale)
        dim_max = max(h, w)
        scale = random_dim / dim_max
-        resize_w = w * scale
+        resize_w = int(w * scale + 0.5)
-        resize_h = h * scale
+        resize_h = int(h * scale + 0.5)
        offset_x = int(max(0, np.random.uniform(0., resize_w - dim)))
        offset_y = int(max(0, np.random.uniform(0., resize_h - dim)))
@@ -2316,25 +2472,26 @@ class RandomResizeCrop(BaseOperator):
        is_mask_crop(bool): whether crop the segmentation.
    """
-    def __init__(
+    def __init__(self,
-            self,
+                 resizes,
-            resizes,
+                 cropsizes,
-            cropsizes,
+                 prob=0.5,
-            prob=0.5,
+                 mode='short',
-            mode='short',
+                 keep_ratio=True,
-            keep_ratio=True,
+                 interp=cv2.INTER_LINEAR,
-            interp=cv2.INTER_LINEAR,
+                 num_attempts=3,
-            num_attempts=3,
+                 cover_all_box=False,
-            cover_all_box=False,
+                 allow_no_crop=False,
-            allow_no_crop=False,
+                 thresholds=[0.3, 0.5, 0.7],
-            thresholds=[0.3, 0.5, 0.7],
+                 is_mask_crop=False,
-            is_mask_crop=False, ):
+                 ioumode="iou"):
        super(RandomResizeCrop, self).__init__()
        self.resizes = resizes
        self.cropsizes = cropsizes
        self.prob = prob
        self.mode = mode
+        self.ioumode = ioumode
        self.resizer = Resize(0, keep_ratio=keep_ratio, interp=interp)
        self.croper = RandomCrop(
@@ -2389,9 +2546,14 @@ class RandomResizeCrop(BaseOperator):
                crop_x = random.randint(0, w - crop_w)
                crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
-                iou = self._iou_matrix(
+                if self.ioumode == "iof":
-                    gt_bbox, np.array(
+                    iou = self._gtcropiou_matrix(
-                        [crop_box], dtype=np.float32))
+                        gt_bbox, np.array(
+                            [crop_box], dtype=np.float32))
+                elif self.ioumode == "iou":
+                    iou = self._iou_matrix(
+                        gt_bbox, np.array(
+                            [crop_box], dtype=np.float32))
                if iou.max() < thresh:
                    continue
@@ -2447,6 +2609,14 @@ class RandomResizeCrop(BaseOperator):
                if 'is_crowd' in sample:
                    sample['is_crowd'] = np.take(
                        sample['is_crowd'], valid_ids, axis=0)
+                if 'gt_areas' in sample:
+                    sample['gt_areas'] = np.take(
+                        sample['gt_areas'], valid_ids, axis=0)
+                if 'gt_joints' in sample:
+                    gt_joints = self._crop_joints(sample['gt_joints'], crop_box)
+                    sample['gt_joints'] = gt_joints[valid_ids]
                return sample
        return sample
@@ -2479,8 +2649,8 @@ class RandomResizeCrop(BaseOperator):
                im_scale = max(target_size_min / im_size_min,
                               target_size_max / im_size_max)
-            resize_h = im_scale * float(im_shape[0])
+            resize_h = int(im_scale * float(im_shape[0]) + 0.5)
-            resize_w = im_scale * float(im_shape[1])
+            resize_w = int(im_scale * float(im_shape[1]) + 0.5)
            im_scale_x = im_scale
            im_scale_y = im_scale
@@ -2540,6 +2710,11 @@ class RandomResizeCrop(BaseOperator):
            ]
            sample['gt_segm'] = np.asarray(masks).astype(np.uint8)
+        if 'gt_joints' in sample:
+            sample['gt_joints'] = self.apply_joints(sample['gt_joints'],
+                                                    [im_scale_x, im_scale_y],
+                                                    [resize_w, resize_h])
        return sample
@@ -2612,10 +2787,10 @@ class RandomShortSideResize(BaseOperator):
        if w < h:
            ow = size
-            oh = int(size * h / w)
+            oh = int(round(size * h / w))
        else:
            oh = size
-            ow = int(size * w / h)
+            ow = int(round(size * w / h))
        return (ow, oh)
@@ -2672,6 +2847,16 @@ class RandomShortSideResize(BaseOperator):
                for gt_segm in sample['gt_segm']
            ]
            sample['gt_segm'] = np.asarray(masks).astype(np.uint8)
+        if 'gt_joints' in sample:
+            sample['gt_joints'] = self.apply_joints(
+                sample['gt_joints'], [im_scale_x, im_scale_y], target_size)
+        # apply areas
+        if 'gt_areas' in sample:
+            sample['gt_areas'] = self.apply_area(sample['gt_areas'],
+                                                 [im_scale_x, im_scale_y])
        return sample
    def apply_bbox(self, bbox, scale, size):
@@ -2683,6 +2868,23 @@ class RandomShortSideResize(BaseOperator):
        bbox[:, 1::2] = np.clip(bbox[:, 1::2], 0, resize_h)
        return bbox.astype('float32')
+    def apply_joints(self, joints, scale, size):
+        im_scale_x, im_scale_y = scale
+        resize_w, resize_h = size
+        joints[..., 0] *= im_scale_x
+        joints[..., 1] *= im_scale_y
+        # joints[joints[..., 0] >= resize_w, :] = 0
+        # joints[joints[..., 1] >= resize_h, :] = 0
+        # joints[joints[..., 0] < 0, :] = 0
+        # joints[joints[..., 1] < 0, :] = 0
+        joints[..., 0] = np.clip(joints[..., 0], 0, resize_w)
+        joints[..., 1] = np.clip(joints[..., 1], 0, resize_h)
+        return joints
+    def apply_area(self, area, scale):
+        im_scale_x, im_scale_y = scale
+        return area * im_scale_x * im_scale_y
    def apply_segm(self, segms, im_size, scale):
        def _resize_poly(poly, im_scale_x, im_scale_y):
            resized_poly = np.array(poly).astype('float32')
@@ -2730,6 +2932,44 @@ class RandomShortSideResize(BaseOperator):
        return self.resize(sample, target_size, self.max_size, interp)
+@register_op
+class RandomShortSideRangeResize(RandomShortSideResize):
+    def __init__(self, scales, interp=cv2.INTER_LINEAR, random_interp=False):
+        """
+        Resize the image randomly according to the short side. If max_size is not None,
+        the long side is scaled according to max_size. The whole process will be keep ratio.
+        Args:
+            short_side_sizes (list|tuple): Image target short side size.
+            interp (int): The interpolation method.
+            random_interp (bool): Whether random select interpolation method.
+        """
+        super(RandomShortSideRangeResize, self).__init__(scales, None, interp,
+                                                         random_interp)
+        assert isinstance(scales,
+                          Sequence), "short_side_sizes must be List or Tuple"
+        self.scales = scales
+    def random_sample(self, img_scales):
+        img_scale_long = [max(s) for s in img_scales]
+        img_scale_short = [min(s) for s in img_scales]
+        long_edge = np.random.randint(
+            min(img_scale_long), max(img_scale_long) + 1)
+        short_edge = np.random.randint(
+            min(img_scale_short), max(img_scale_short) + 1)
+        img_scale = (long_edge, short_edge)
+        return img_scale
+    def apply(self, sample, context=None):
+        long_edge, short_edge = self.random_sample(self.short_side_sizes)
+        # print("target size:{}".format((long_edge, short_edge)))
+        interp = random.choice(
+            self.interps) if self.random_interp else self.interp
+        return self.resize(sample, short_edge, long_edge, interp)
 @register_op
 class RandomSizeCrop(BaseOperator):
    """
@@ -2805,6 +3045,9 @@ class RandomSizeCrop(BaseOperator):
                sample['is_crowd'] = sample['is_crowd'][keep_index] if len(
                    keep_index) > 0 else np.zeros(
                        [0, 1], dtype=np.float32)
+            if 'gt_areas' in sample:
+                sample['gt_areas'] = np.take(
+                    sample['gt_areas'], keep_index, axis=0)
        image_shape = sample['image'].shape[:2]
        sample['image'] = self.paddle_crop(sample['image'], *region)
@@ -2826,6 +3069,12 @@ class RandomSizeCrop(BaseOperator):
            if keep_index is not None and len(keep_index) > 0:
                sample['gt_segm'] = sample['gt_segm'][keep_index]
+        if 'gt_joints' in sample:
+            gt_joints = self._crop_joints(sample['gt_joints'], region)
+            sample['gt_joints'] = gt_joints
+            if keep_index is not None:
+                sample['gt_joints'] = sample['gt_joints'][keep_index]
        return sample
    def apply_bbox(self, bbox, region):
@@ -2836,6 +3085,19 @@ class RandomSizeCrop(BaseOperator):
        crop_bbox = crop_bbox.clip(min=0)
        return crop_bbox.reshape([-1, 4]).astype('float32')
+    def _crop_joints(self, joints, region):
+        y1, x1, h, w = region
+        x2 = x1 + w
+        y2 = y1 + h
+        # x1, y1, x2, y2 = crop
+        joints[..., 0] -= x1
+        joints[..., 1] -= y1
+        joints[joints[..., 0] > w, :] = 0
+        joints[joints[..., 1] > h, :] = 0
+        joints[joints[..., 0] < 0, :] = 0
+        joints[joints[..., 1] < 0, :] = 0
+        return joints
    def apply_segm(self, segms, region, image_shape):
        def _crop_poly(segm, crop):
            xmin, ymin, xmax, ymax = crop

--- a/ppdet/modeling/architectures/__init__.py
+++ b/ppdet/modeling/architectures/__init__.py
@@ -72,3 +72,4 @@ from .yolof import *
 from .pose3d_metro import *
 from .centertrack import *
 from .queryinst import *
+from .keypoint_petr import *
--- a/ppdet/modeling/architectures/keypoint_petr.py
+++ b/ppdet/modeling/architectures/keypoint_petr.py
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. 
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License. 
+# You may obtain a copy of the License at 
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and 
+# limitations under the License.
+"""
+this code is base on https://github.com/hikvision-research/opera/blob/main/opera/models/detectors/petr.py
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import paddle
+from ppdet.core.workspace import register
+from .meta_arch import BaseArch
+from .. import layers as L
+__all__ = ['PETR']
+@register
+class PETR(BaseArch):
+    __category__ = 'architecture'
+    __inject__ = ['backbone', 'neck', 'bbox_head']
+    def __init__(self,
+                 backbone='ResNet',
+                 neck='ChannelMapper',
+                 bbox_head='PETRHead'):
+        """
+        PETR, see https://openaccess.thecvf.com/content/CVPR2022/papers/Shi_End-to-End_Multi-Person_Pose_Estimation_With_Transformers_CVPR_2022_paper.pdf
+        Args:
+            backbone (nn.Layer): backbone instance
+            neck (nn.Layer): neck between backbone and head
+            bbox_head (nn.Layer): model output and loss
+        """
+        super(PETR, self).__init__()
+        self.backbone = backbone
+        if neck is not None:
+            self.with_neck = True
+        self.neck = neck
+        self.bbox_head = bbox_head
+        self.deploy = False
+    def extract_feat(self, img):
+        """Directly extract features from the backbone+neck."""
+        x = self.backbone(img)
+        if self.with_neck:
+            x = self.neck(x)
+        return x
+    def get_inputs(self):
+        img_metas = []
+        gt_bboxes = []
+        gt_labels = []
+        gt_keypoints = []
+        gt_areas = []
+        pad_gt_mask = self.inputs['pad_gt_mask'].astype("bool").squeeze(-1)
+        for idx, im_shape in enumerate(self.inputs['im_shape']):
+            img_meta = {
+                'img_shape': im_shape.astype("int32").tolist() + [1, ],
+                'batch_input_shape': self.inputs['image'].shape[-2:],
+                'image_name': self.inputs['image_file'][idx]
+            }
+            img_metas.append(img_meta)
+            if (not pad_gt_mask[idx].any()):
+                gt_keypoints.append(self.inputs['gt_joints'][idx][:1])
+                gt_labels.append(self.inputs['gt_class'][idx][:1])
+                gt_bboxes.append(self.inputs['gt_bbox'][idx][:1])
+                gt_areas.append(self.inputs['gt_areas'][idx][:1])
+                continue
+            gt_keypoints.append(self.inputs['gt_joints'][idx][pad_gt_mask[idx]])
+            gt_labels.append(self.inputs['gt_class'][idx][pad_gt_mask[idx]])
+            gt_bboxes.append(self.inputs['gt_bbox'][idx][pad_gt_mask[idx]])
+            gt_areas.append(self.inputs['gt_areas'][idx][pad_gt_mask[idx]])
+        return img_metas, gt_bboxes, gt_labels, gt_keypoints, gt_areas
+    def get_loss(self):
+        """
+        Args:
+            img (Tensor): Input images of shape (N, C, H, W).
+                Typically these should be mean centered and std scaled.
+            img_metas (list[dict]): A List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                :class:`mmdet.datasets.pipelines.Collect`.
+            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
+                image in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels (list[Tensor]): Class indices corresponding to each box.
+            gt_keypoints (list[Tensor]): Each item are the truth keypoints for
+                each image in [p^{1}_x, p^{1}_y, p^{1}_v, ..., p^{K}_x,
+                p^{K}_y, p^{K}_v] format.
+            gt_areas (list[Tensor]): mask areas corresponding to each box.
+            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
+                boxes can be ignored when computing the loss.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        img_metas, gt_bboxes, gt_labels, gt_keypoints, gt_areas = self.get_inputs(
+        )
+        gt_bboxes_ignore = getattr(self.inputs, 'gt_bboxes_ignore', None)
+        x = self.extract_feat(self.inputs)
+        losses = self.bbox_head.forward_train(x, img_metas, gt_bboxes,
+                                              gt_labels, gt_keypoints, gt_areas,
+                                              gt_bboxes_ignore)
+        loss = 0
+        for k, v in losses.items():
+            loss += v
+        losses['loss'] = loss
+        return losses
+    def get_pred_numpy(self):
+        """Used for computing network flops.
+        """
+        img = self.inputs['image']
+        batch_size, _, height, width = img.shape
+        dummy_img_metas = [
+            dict(
+                batch_input_shape=(height, width),
+                img_shape=(height, width, 3),
+                scale_factor=(1., 1., 1., 1.)) for _ in range(batch_size)
+        ]
+        x = self.extract_feat(img)
+        outs = self.bbox_head(x, img_metas=dummy_img_metas)
+        bbox_list = self.bbox_head.get_bboxes(
+            *outs, dummy_img_metas, rescale=True)
+        return bbox_list
+    def get_pred(self):
+        """
+        """
+        img = self.inputs['image']
+        batch_size, _, height, width = img.shape
+        img_metas = [
+            dict(
+                batch_input_shape=(height, width),
+                img_shape=(height, width, 3),
+                scale_factor=self.inputs['scale_factor'][i])
+            for i in range(batch_size)
+        ]
+        kptpred = self.simple_test(
+            self.inputs, img_metas=img_metas, rescale=True)
+        keypoints = kptpred[0][1][0]
+        bboxs = kptpred[0][0][0]
+        keypoints[..., 2] = bboxs[:, None, 4]
+        res_lst = [[keypoints, bboxs[:, 4]]]
+        outputs = {'keypoint': res_lst}
+        return outputs
+    def simple_test(self, inputs, img_metas, rescale=False):
+        """Test function without test time augmentation.
+        Args:
+            inputs (list[paddle.Tensor]): List of multiple images.
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+        Returns:
+            list[list[np.ndarray]]: BBox and keypoint results of each image
+                and classes. The outer list corresponds to each image.
+                The inner list corresponds to each class.
+        """
+        batch_size = len(img_metas)
+        assert batch_size == 1, 'Currently only batch_size 1 for inference ' \
+            f'mode is supported. Found batch_size {batch_size}.'
+        feat = self.extract_feat(inputs)
+        results_list = self.bbox_head.simple_test(
+            feat, img_metas, rescale=rescale)
+        bbox_kpt_results = [
+            self.bbox_kpt2result(det_bboxes, det_labels, det_kpts,
+                                 self.bbox_head.num_classes)
+            for det_bboxes, det_labels, det_kpts in results_list
+        ]
+        return bbox_kpt_results
+    def bbox_kpt2result(self, bboxes, labels, kpts, num_classes):
+        """Convert detection results to a list of numpy arrays.
+        Args:
+            bboxes (paddle.Tensor | np.ndarray): shape (n, 5).
+            labels (paddle.Tensor | np.ndarray): shape (n, ).
+            kpts (paddle.Tensor | np.ndarray): shape (n, K, 3).
+            num_classes (int): class number, including background class.
+        Returns:
+            list(ndarray): bbox and keypoint results of each class.
+        """
+        if bboxes.shape[0] == 0:
+            return [np.zeros((0, 5), dtype=np.float32) for i in range(num_classes)], \
+                [np.zeros((0, kpts.size(1), 3), dtype=np.float32)
+                    for i in range(num_classes)]
+        else:
+            if isinstance(bboxes, paddle.Tensor):
+                bboxes = bboxes.numpy()
+                labels = labels.numpy()
+                kpts = kpts.numpy()
+            return [bboxes[labels == i, :] for i in range(num_classes)], \
+                [kpts[labels == i, :, :] for i in range(num_classes)]
--- a/ppdet/modeling/assigners/__init__.py
+++ b/ppdet/modeling/assigners/__init__.py
@@ -31,3 +31,5 @@ from .fcosr_assigner import *
 from .rotated_task_aligned_assigner import *
 from .task_aligned_assigner_cr import *
 from .uniform_assigner import *
+from .hungarian_assigner import *
+from .pose_utils import *
--- a/ppdet/modeling/assigners/hungarian_assigner.py
+++ b/ppdet/modeling/assigners/hungarian_assigner.py
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+try:
+    from scipy.optimize import linear_sum_assignment
+except ImportError:
+    linear_sum_assignment = None
+import paddle
+from ppdet.core.workspace import register
+__all__ = ['PoseHungarianAssigner', 'PseudoSampler']
+class AssignResult:
+    """Stores assignments between predicted and truth boxes.
+    Attributes:
+        num_gts (int): the number of truth boxes considered when computing this
+            assignment
+        gt_inds (LongTensor): for each predicted box indicates the 1-based
+            index of the assigned truth box. 0 means unassigned and -1 means
+            ignore.
+        max_overlaps (FloatTensor): the iou between the predicted box and its
+            assigned truth box.
+        labels (None | LongTensor): If specified, for each predicted box
+            indicates the category label of the assigned truth box.
+    """
+    def __init__(self, num_gts, gt_inds, max_overlaps, labels=None):
+        self.num_gts = num_gts
+        self.gt_inds = gt_inds
+        self.max_overlaps = max_overlaps
+        self.labels = labels
+        # Interface for possible user-defined properties
+        self._extra_properties = {}
+    @property
+    def num_preds(self):
+        """int: the number of predictions in this assignment"""
+        return len(self.gt_inds)
+    def set_extra_property(self, key, value):
+        """Set user-defined new property."""
+        assert key not in self.info
+        self._extra_properties[key] = value
+    def get_extra_property(self, key):
+        """Get user-defined property."""
+        return self._extra_properties.get(key, None)
+    @property
+    def info(self):
+        """dict: a dictionary of info about the object"""
+        basic_info = {
+            'num_gts': self.num_gts,
+            'num_preds': self.num_preds,
+            'gt_inds': self.gt_inds,
+            'max_overlaps': self.max_overlaps,
+            'labels': self.labels,
+        }
+        basic_info.update(self._extra_properties)
+        return basic_info
+@register
+class PoseHungarianAssigner:
+    """Computes one-to-one matching between predictions and ground truth.
+    This class computes an assignment between the targets and the predictions
+    based on the costs. The costs are weighted sum of three components:
+    classification cost, regression L1 cost and regression oks cost. The
+    targets don't include the no_object, so generally there are more
+    predictions than targets. After the one-to-one matching, the un-matched
+    are treated as backgrounds. Thus each query prediction will be assigned
+    with `0` or a positive integer indicating the ground truth index:
+    - 0: negative sample, no assigned gt.
+    - positive integer: positive sample, index (1-based) of assigned gt.
+    Args:
+        cls_weight (int | float, optional): The scale factor for classification
+            cost. Default 1.0.
+        kpt_weight (int | float, optional): The scale factor for regression
+            L1 cost. Default 1.0.
+        oks_weight (int | float, optional): The scale factor for regression
+            oks cost. Default 1.0.
+    """
+    __inject__ = ['cls_cost', 'kpt_cost', 'oks_cost']
+    def __init__(self,
+                 cls_cost='ClassificationCost',
+                 kpt_cost='KptL1Cost',
+                 oks_cost='OksCost'):
+        self.cls_cost = cls_cost
+        self.kpt_cost = kpt_cost
+        self.oks_cost = oks_cost
+    def assign(self,
+               cls_pred,
+               kpt_pred,
+               gt_labels,
+               gt_keypoints,
+               gt_areas,
+               img_meta,
+               eps=1e-7):
+        """Computes one-to-one matching based on the weighted costs.
+        This method assign each query prediction to a ground truth or
+        background. The `assigned_gt_inds` with -1 means don't care,
+        0 means negative sample, and positive number is the index (1-based)
+        of assigned gt.
+        The assignment is done in the following steps, the order matters.
+        1. assign every prediction to -1
+        2. compute the weighted costs
+        3. do Hungarian matching on CPU based on the costs
+        4. assign all to 0 (background) first, then for each matched pair
+           between predictions and gts, treat this prediction as foreground
+           and assign the corresponding gt index (plus 1) to it.
+        Args:
+            cls_pred (Tensor): Predicted classification logits, shape
+                [num_query, num_class].
+            kpt_pred (Tensor): Predicted keypoints with normalized coordinates
+                (x_{i}, y_{i}), which are all in range [0, 1]. Shape
+                [num_query, K*2].
+            gt_labels (Tensor): Label of `gt_keypoints`, shape (num_gt,).
+            gt_keypoints (Tensor): Ground truth keypoints with unnormalized
+                coordinates [p^{1}_x, p^{1}_y, p^{1}_v, ..., \
+                    p^{K}_x, p^{K}_y, p^{K}_v]. Shape [num_gt, K*3].
+            gt_areas (Tensor): Ground truth mask areas, shape (num_gt,).
+            img_meta (dict): Meta information for current image.
+            eps (int | float, optional): A value added to the denominator for
+                numerical stability. Default 1e-7.
+        Returns:
+            :obj:`AssignResult`: The assigned result.
+        """
+        num_gts, num_kpts = gt_keypoints.shape[0], kpt_pred.shape[0]
+        if not gt_keypoints.astype('bool').any():
+            num_gts = 0
+        # 1. assign -1 by default
+        assigned_gt_inds = paddle.full((num_kpts, ), -1, dtype="int64")
+        assigned_labels = paddle.full((num_kpts, ), -1, dtype="int64")
+        if num_gts == 0 or num_kpts == 0:
+            # No ground truth or keypoints, return empty assignment
+            if num_gts == 0:
+                # No ground truth, assign all to background
+                assigned_gt_inds[:] = 0
+            return AssignResult(
+                num_gts, assigned_gt_inds, None, labels=assigned_labels)
+        img_h, img_w, _ = img_meta['img_shape']
+        factor = paddle.to_tensor(
+            [img_w, img_h, img_w, img_h], dtype=gt_keypoints.dtype).reshape(
+                (1, -1))
+        # 2. compute the weighted costs
+        # classification cost
+        cls_cost = self.cls_cost(cls_pred, gt_labels)
+        # keypoint regression L1 cost
+        gt_keypoints_reshape = gt_keypoints.reshape((gt_keypoints.shape[0], -1,
+                                                     3))
+        valid_kpt_flag = gt_keypoints_reshape[..., -1]
+        kpt_pred_tmp = kpt_pred.clone().detach().reshape((kpt_pred.shape[0], -1,
+                                                          2))
+        normalize_gt_keypoints = gt_keypoints_reshape[
+            ..., :2] / factor[:, :2].unsqueeze(0)
+        kpt_cost = self.kpt_cost(kpt_pred_tmp, normalize_gt_keypoints,
+                                 valid_kpt_flag)
+        # keypoint OKS cost
+        kpt_pred_tmp = kpt_pred.clone().detach().reshape((kpt_pred.shape[0], -1,
+                                                          2))
+        kpt_pred_tmp = kpt_pred_tmp * factor[:, :2].unsqueeze(0)
+        oks_cost = self.oks_cost(kpt_pred_tmp, gt_keypoints_reshape[..., :2],
+                                 valid_kpt_flag, gt_areas)
+        # weighted sum of above three costs
+        cost = cls_cost + kpt_cost + oks_cost
+        # 3. do Hungarian matching on CPU using linear_sum_assignment
+        cost = cost.detach().cpu()
+        if linear_sum_assignment is None:
+            raise ImportError('Please run "pip install scipy" '
+                              'to install scipy first.')
+        matched_row_inds, matched_col_inds = linear_sum_assignment(cost)
+        matched_row_inds = paddle.to_tensor(matched_row_inds)
+        matched_col_inds = paddle.to_tensor(matched_col_inds)
+        # 4. assign backgrounds and foregrounds
+        # assign all indices to backgrounds first
+        assigned_gt_inds[:] = 0
+        # assign foregrounds based on matching results
+        assigned_gt_inds[matched_row_inds] = matched_col_inds + 1
+        assigned_labels[matched_row_inds] = gt_labels[matched_col_inds][
+            ..., 0].astype("int64")
+        return AssignResult(
+            num_gts, assigned_gt_inds, None, labels=assigned_labels)
+class SamplingResult:
+    """Bbox sampling result.
+    """
+    def __init__(self, pos_inds, neg_inds, bboxes, gt_bboxes, assign_result,
+                 gt_flags):
+        self.pos_inds = pos_inds
+        self.neg_inds = neg_inds
+        if pos_inds.size > 0:
+            self.pos_bboxes = bboxes[pos_inds]
+            self.neg_bboxes = bboxes[neg_inds]
+            self.pos_is_gt = gt_flags[pos_inds]
+            self.num_gts = gt_bboxes.shape[0]
+            self.pos_assigned_gt_inds = assign_result.gt_inds[pos_inds] - 1
+            if gt_bboxes.numel() == 0:
+                # hack for index error case
+                assert self.pos_assigned_gt_inds.numel() == 0
+                self.pos_gt_bboxes = paddle.zeros(
+                    gt_bboxes.shape, dtype=gt_bboxes.dtype).reshape((-1, 4))
+            else:
+                if len(gt_bboxes.shape) < 2:
+                    gt_bboxes = gt_bboxes.reshape((-1, 4))
+                self.pos_gt_bboxes = paddle.index_select(
+                    gt_bboxes,
+                    self.pos_assigned_gt_inds.astype('int64'),
+                    axis=0)
+            if assign_result.labels is not None:
+                self.pos_gt_labels = assign_result.labels[pos_inds]
+            else:
+                self.pos_gt_labels = None
+    @property
+    def bboxes(self):
+        """paddle.Tensor: concatenated positive and negative boxes"""
+        return paddle.concat([self.pos_bboxes, self.neg_bboxes])
+    def __nice__(self):
+        data = self.info.copy()
+        data['pos_bboxes'] = data.pop('pos_bboxes').shape
+        data['neg_bboxes'] = data.pop('neg_bboxes').shape
+        parts = [f"'{k}': {v!r}" for k, v in sorted(data.items())]
+        body = '    ' + ',\n    '.join(parts)
+        return '{\n' + body + '\n}'
+    @property
+    def info(self):
+        """Returns a dictionary of info about the object."""
+        return {
+            'pos_inds': self.pos_inds,
+            'neg_inds': self.neg_inds,
+            'pos_bboxes': self.pos_bboxes,
+            'neg_bboxes': self.neg_bboxes,
+            'pos_is_gt': self.pos_is_gt,
+            'num_gts': self.num_gts,
+            'pos_assigned_gt_inds': self.pos_assigned_gt_inds,
+        }
+@register
+class PseudoSampler:
+    """A pseudo sampler that does not do sampling actually."""
+    def __init__(self, **kwargs):
+        pass
+    def _sample_pos(self, **kwargs):
+        """Sample positive samples."""
+        raise NotImplementedError
+    def _sample_neg(self, **kwargs):
+        """Sample negative samples."""
+        raise NotImplementedError
+    def sample(self, assign_result, bboxes, gt_bboxes, *args, **kwargs):
+        """Directly returns the positive and negative indices  of samples.
+        Args:
+            assign_result (:obj:`AssignResult`): Assigned results
+            bboxes (paddle.Tensor): Bounding boxes
+            gt_bboxes (paddle.Tensor): Ground truth boxes
+        Returns:
+            :obj:`SamplingResult`: sampler results
+        """
+        pos_inds = paddle.nonzero(
+            assign_result.gt_inds > 0, as_tuple=False).squeeze(-1)
+        neg_inds = paddle.nonzero(
+            assign_result.gt_inds == 0, as_tuple=False).squeeze(-1)
+        gt_flags = paddle.zeros([bboxes.shape[0]], dtype='int32')
+        sampling_result = SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes,
+                                         assign_result, gt_flags)
+        return sampling_result
--- a/ppdet/modeling/assigners/pose_utils.py
+++ b/ppdet/modeling/assigners/pose_utils.py
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import numpy as np
+import paddle
+import paddle.nn.functional as F
+from ppdet.core.workspace import register
+__all__ = ['KptL1Cost', 'OksCost', 'ClassificationCost']
+def masked_fill(x, mask, value):
+    y = paddle.full(x.shape, value, x.dtype)
+    return paddle.where(mask, y, x)
+@register
+class KptL1Cost(object):
+    """KptL1Cost.
+    this function based on: https://github.com/hikvision-research/opera/blob/main/opera/core/bbox/match_costs/match_cost.py
+    Args:
+        weight (int | float, optional): loss_weight.
+    """
+    def __init__(self, weight=1.0):
+        self.weight = weight
+    def __call__(self, kpt_pred, gt_keypoints, valid_kpt_flag):
+        """
+        Args:
+            kpt_pred (Tensor): Predicted keypoints with normalized coordinates
+                (x_{i}, y_{i}), which are all in range [0, 1]. Shape
+                [num_query, K, 2].
+            gt_keypoints (Tensor): Ground truth keypoints with normalized
+                coordinates (x_{i}, y_{i}). Shape [num_gt, K, 2].
+            valid_kpt_flag (Tensor): valid flag of ground truth keypoints.
+                Shape [num_gt, K].
+        Returns:
+            paddle.Tensor: kpt_cost value with weight.
+        """
+        kpt_cost = []
+        for i in range(len(gt_keypoints)):
+            if gt_keypoints[i].size == 0:
+                kpt_cost.append(kpt_pred.sum() * 0)
+            kpt_pred_tmp = kpt_pred.clone()
+            valid_flag = valid_kpt_flag[i] > 0
+            valid_flag_expand = valid_flag.unsqueeze(0).unsqueeze(-1).expand_as(
+                kpt_pred_tmp)
+            if not valid_flag_expand.all():
+                kpt_pred_tmp = masked_fill(kpt_pred_tmp, ~valid_flag_expand, 0)
+            cost = F.pairwise_distance(
+                kpt_pred_tmp.reshape((kpt_pred_tmp.shape[0], -1)),
+                gt_keypoints[i].reshape((-1, )).unsqueeze(0),
+                p=1,
+                keepdim=True)
+            avg_factor = paddle.clip(
+                valid_flag.astype('float32').sum() * 2, 1.0)
+            cost = cost / avg_factor
+            kpt_cost.append(cost)
+        kpt_cost = paddle.concat(kpt_cost, axis=1)
+        return kpt_cost * self.weight
+@register
+class OksCost(object):
+    """OksCost.
+    this function based on: https://github.com/hikvision-research/opera/blob/main/opera/core/bbox/match_costs/match_cost.py
+    Args:
+        num_keypoints (int): number of keypoints
+        weight (int | float, optional): loss_weight.
+    """
+    def __init__(self, num_keypoints=17, weight=1.0):
+        self.weight = weight
+        if num_keypoints == 17:
+            self.sigmas = np.array(
+                [
+                    .26, .25, .25, .35, .35, .79, .79, .72, .72, .62, .62, 1.07,
+                    1.07, .87, .87, .89, .89
+                ],
+                dtype=np.float32) / 10.0
+        elif num_keypoints == 14:
+            self.sigmas = np.array(
+                [
+                    .79, .79, .72, .72, .62, .62, 1.07, 1.07, .87, .87, .89,
+                    .89, .79, .79
+                ],
+                dtype=np.float32) / 10.0
+        else:
+            raise ValueError(f'Unsupported keypoints number {num_keypoints}')
+    def __call__(self, kpt_pred, gt_keypoints, valid_kpt_flag, gt_areas):
+        """
+        Args:
+            kpt_pred (Tensor): Predicted keypoints with unnormalized
+                coordinates (x_{i}, y_{i}). Shape [num_query, K, 2].
+            gt_keypoints (Tensor): Ground truth keypoints with unnormalized
+                coordinates (x_{i}, y_{i}). Shape [num_gt, K, 2].
+            valid_kpt_flag (Tensor): valid flag of ground truth keypoints.
+                Shape [num_gt, K].
+            gt_areas (Tensor): Ground truth mask areas. Shape [num_gt,].
+        Returns:
+            paddle.Tensor: oks_cost value with weight.
+        """
+        sigmas = paddle.to_tensor(self.sigmas)
+        variances = (sigmas * 2)**2
+        oks_cost = []
+        assert len(gt_keypoints) == len(gt_areas)
+        for i in range(len(gt_keypoints)):
+            if gt_keypoints[i].size == 0:
+                oks_cost.append(kpt_pred.sum() * 0)
+            squared_distance = \
+                (kpt_pred[:, :, 0] - gt_keypoints[i, :, 0].unsqueeze(0)) ** 2 + \
+                (kpt_pred[:, :, 1] - gt_keypoints[i, :, 1].unsqueeze(0)) ** 2
+            vis_flag = (valid_kpt_flag[i] > 0).astype('int')
+            vis_ind = vis_flag.nonzero(as_tuple=False)[:, 0]
+            num_vis_kpt = vis_ind.shape[0]
+            # assert num_vis_kpt > 0
+            if num_vis_kpt == 0:
+                oks_cost.append(paddle.zeros((squared_distance.shape[0], 1)))
+                continue
+            area = gt_areas[i]
+            squared_distance0 = squared_distance / (area * variances * 2)
+            squared_distance0 = paddle.index_select(
+                squared_distance0, vis_ind, axis=1)
+            squared_distance1 = paddle.exp(-squared_distance0).sum(axis=1,
+                                                                   keepdim=True)
+            oks = squared_distance1 / num_vis_kpt
+            # The 1 is a constant that doesn't change the matching, so omitted.
+            oks_cost.append(-oks)
+        oks_cost = paddle.concat(oks_cost, axis=1)
+        return oks_cost * self.weight
+@register
+class ClassificationCost:
+    """ClsSoftmaxCost.
+     Args:
+         weight (int | float, optional): loss_weight
+    """
+    def __init__(self, weight=1.):
+        self.weight = weight
+    def __call__(self, cls_pred, gt_labels):
+        """
+        Args:
+            cls_pred (Tensor): Predicted classification logits, shape
+                (num_query, num_class).
+            gt_labels (Tensor): Label of `gt_bboxes`, shape (num_gt,).
+        Returns:
+            paddle.Tensor: cls_cost value with weight
+        """
+        # Following the official DETR repo, contrary to the loss that
+        # NLL is used, we approximate it in 1 - cls_score[gt_label].
+        # The 1 is a constant that doesn't change the matching,
+        # so it can be omitted.
+        cls_score = cls_pred.softmax(-1)
+        cls_cost = -cls_score[:, gt_labels]
+        return cls_cost * self.weight
+@register
+class FocalLossCost:
+    """FocalLossCost.
+     Args:
+         weight (int | float, optional): loss_weight
+         alpha (int | float, optional): focal_loss alpha
+         gamma (int | float, optional): focal_loss gamma
+         eps (float, optional): default 1e-12
+         binary_input (bool, optional): Whether the input is binary,
+            default False.
+    """
+    def __init__(self,
+                 weight=1.,
+                 alpha=0.25,
+                 gamma=2,
+                 eps=1e-12,
+                 binary_input=False):
+        self.weight = weight
+        self.alpha = alpha
+        self.gamma = gamma
+        self.eps = eps
+        self.binary_input = binary_input
+    def _focal_loss_cost(self, cls_pred, gt_labels):
+        """
+        Args:
+            cls_pred (Tensor): Predicted classification logits, shape
+                (num_query, num_class).
+            gt_labels (Tensor): Label of `gt_bboxes`, shape (num_gt,).
+        Returns:
+            paddle.Tensor: cls_cost value with weight
+        """
+        if gt_labels.size == 0:
+            return cls_pred.sum() * 0
+        cls_pred = F.sigmoid(cls_pred)
+        neg_cost = -(1 - cls_pred + self.eps).log() * (
+            1 - self.alpha) * cls_pred.pow(self.gamma)
+        pos_cost = -(cls_pred + self.eps).log() * self.alpha * (
+            1 - cls_pred).pow(self.gamma)
+        cls_cost = paddle.index_select(
+            pos_cost, gt_labels, axis=1) - paddle.index_select(
+                neg_cost, gt_labels, axis=1)
+        return cls_cost * self.weight
+    def _mask_focal_loss_cost(self, cls_pred, gt_labels):
+        """
+        Args:
+            cls_pred (Tensor): Predicted classfication logits
+                in shape (num_query, d1, ..., dn), dtype=paddle.float32.
+            gt_labels (Tensor): Ground truth in shape (num_gt, d1, ..., dn),
+                dtype=paddle.long. Labels should be binary.
+        Returns:
+            Tensor: Focal cost matrix with weight in shape\
+                (num_query, num_gt).
+        """
+        cls_pred = cls_pred.flatten(1)
+        gt_labels = gt_labels.flatten(1).float()
+        n = cls_pred.shape[1]
+        cls_pred = F.sigmoid(cls_pred)
+        neg_cost = -(1 - cls_pred + self.eps).log() * (
+            1 - self.alpha) * cls_pred.pow(self.gamma)
+        pos_cost = -(cls_pred + self.eps).log() * self.alpha * (
+            1 - cls_pred).pow(self.gamma)
+        cls_cost = paddle.einsum('nc,mc->nm', pos_cost, gt_labels) + \
+            paddle.einsum('nc,mc->nm', neg_cost, (1 - gt_labels))
+        return cls_cost / n * self.weight
+    def __call__(self, cls_pred, gt_labels):
+        """
+        Args:
+            cls_pred (Tensor): Predicted classfication logits.
+            gt_labels (Tensor)): Labels.
+        Returns:
+            Tensor: Focal cost matrix with weight in shape\
+                (num_query, num_gt).
+        """
+        if self.binary_input:
+            return self._mask_focal_loss_cost(cls_pred, gt_labels)
+        else:
+            return self._focal_loss_cost(cls_pred, gt_labels)
--- a/ppdet/modeling/backbones/resnet.py
+++ b/ppdet/modeling/backbones/resnet.py
@@ -285,36 +285,6 @@ class BottleNeck(nn.Layer):
        # ResNeXt
        width = int(ch_out * (base_width / 64.)) * groups
-        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 * self.expansion,
-                        filter_size=1,
-                        stride=1,
-                        norm_type=norm_type,
-                        norm_decay=norm_decay,
-                        freeze_norm=freeze_norm,
-                        lr=lr))
-            else:
-                self.short = ConvNormLayer(
-                    ch_in=ch_in,
-                    ch_out=ch_out * self.expansion,
-                    filter_size=1,
-                    stride=stride,
-                    norm_type=norm_type,
-                    norm_decay=norm_decay,
-                    freeze_norm=freeze_norm,
-                    lr=lr)
        self.branch2a = ConvNormLayer(
            ch_in=ch_in,
            ch_out=width,
@@ -351,6 +321,36 @@ class BottleNeck(nn.Layer):
            freeze_norm=freeze_norm,
            lr=lr)
+        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 * self.expansion,
+                        filter_size=1,
+                        stride=1,
+                        norm_type=norm_type,
+                        norm_decay=norm_decay,
+                        freeze_norm=freeze_norm,
+                        lr=lr))
+            else:
+                self.short = ConvNormLayer(
+                    ch_in=ch_in,
+                    ch_out=ch_out * self.expansion,
+                    filter_size=1,
+                    stride=stride,
+                    norm_type=norm_type,
+                    norm_decay=norm_decay,
+                    freeze_norm=freeze_norm,
+                    lr=lr)
        self.std_senet = std_senet
        if self.std_senet:
            self.se = SELayer(ch_out * self.expansion)

--- a/ppdet/modeling/backbones/vision_transformer.py
+++ b/ppdet/modeling/backbones/vision_transformer.py
@@ -284,9 +284,9 @@ class RelativePositionBias(nn.Layer):
    def forward(self):
        relative_position_bias = \
-            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.relative_position_bias_table[self.relative_position_index.reshape([-1])].reshape([
                 self.window_size[0] * self.window_size[1] + 1,
-                 self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH 
+                 self.window_size[0] * self.window_size[1] + 1, -1])  # Wh*Ww,Wh*Ww,nH 
        return relative_position_bias.transpose((2, 0, 1))  # nH, Wh*Ww, Wh*Ww

--- a/ppdet/modeling/heads/__init__.py
+++ b/ppdet/modeling/heads/__init__.py
@@ -67,3 +67,4 @@ from .yolof_head import *
 from .ppyoloe_contrast_head import *
 from .centertrack_head import *
 from .sparse_roi_head import *
+from .petr_head import *
--- a/ppdet/modeling/heads/petr_head.py
+++ b/ppdet/modeling/heads/petr_head.py
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. 
+#   
+# Licensed under the Apache License, Version 2.0 (the "License");   
+# you may not use this file except in compliance with the License.  
+# You may obtain a copy of the License at   
+#   
+#     http://www.apache.org/licenses/LICENSE-2.0    
+#   
+# Unless required by applicable law or agreed to in writing, software   
+# distributed under the License is distributed on an "AS IS" BASIS, 
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.  
+# See the License for the specific language governing permissions and   
+# limitations under the License.
+"""
+this code is base on https://github.com/hikvision-research/opera/blob/main/opera/models/dense_heads/petr_head.py
+"""
+import copy
+import numpy as np
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from ppdet.core.workspace import register
+import paddle.distributed as dist
+from ..transformers.petr_transformer import inverse_sigmoid, masked_fill
+from ..initializer import constant_, normal_
+__all__ = ["PETRHead"]
+from functools import partial
+def bias_init_with_prob(prior_prob: float) -> float:
+    """initialize conv/fc bias value according to a given probability value."""
+    bias_init = float(-np.log((1 - prior_prob) / prior_prob))
+    return bias_init
+def multi_apply(func, *args, **kwargs):
+    """Apply function to a list of arguments.
+    Note:
+        This function applies the ``func`` to multiple inputs and
+        map the multiple outputs of the ``func`` into different
+        list. Each list contains the same type of outputs corresponding
+        to different inputs.
+    Args:
+        func (Function): A function that will be applied to a list of
+            arguments
+    Returns:
+        tuple(list): A tuple containing multiple list, each list contains \
+            a kind of returned results by the function
+    """
+    pfunc = partial(func, **kwargs) if kwargs else func
+    map_results = map(pfunc, *args)
+    res = tuple(map(list, zip(*map_results)))
+    return res
+def reduce_mean(tensor):
+    """"Obtain the mean of tensor on different GPUs."""
+    if not (dist.get_world_size() and dist.is_initialized()):
+        return tensor
+    tensor = tensor.clone()
+    dist.all_reduce(
+        tensor.divide(
+            paddle.to_tensor(
+                dist.get_world_size(), dtype='float32')),
+        op=dist.ReduceOp.SUM)
+    return tensor
+def gaussian_radius(det_size, min_overlap=0.7):
+    """calculate gaussian radius according to object size.
+    """
+    height, width = det_size
+    a1 = 1
+    b1 = (height + width)
+    c1 = width * height * (1 - min_overlap) / (1 + min_overlap)
+    sq1 = paddle.sqrt(b1**2 - 4 * a1 * c1)
+    r1 = (b1 + sq1) / 2
+    a2 = 4
+    b2 = 2 * (height + width)
+    c2 = (1 - min_overlap) * width * height
+    sq2 = paddle.sqrt(b2**2 - 4 * a2 * c2)
+    r2 = (b2 + sq2) / 2
+    a3 = 4 * min_overlap
+    b3 = -2 * min_overlap * (height + width)
+    c3 = (min_overlap - 1) * width * height
+    sq3 = paddle.sqrt(b3**2 - 4 * a3 * c3)
+    r3 = (b3 + sq3) / 2
+    return min(r1, r2, r3)
+def gaussian2D(shape, sigma=1):
+    m, n = [(ss - 1.) / 2. for ss in shape]
+    y = paddle.arange(-m, m + 1, dtype="float32")[:, None]
+    x = paddle.arange(-n, n + 1, dtype="float32")[None, :]
+    # y, x = np.ogrid[-m:m + 1, -n:n + 1]
+    h = paddle.exp(-(x * x + y * y) / (2 * sigma * sigma))
+    h[h < np.finfo(np.float32).eps * h.max()] = 0
+    return h
+def draw_umich_gaussian(heatmap, center, radius, k=1):
+    diameter = 2 * radius + 1
+    gaussian = gaussian2D((diameter, diameter), sigma=diameter / 6)
+    gaussian = paddle.to_tensor(gaussian, dtype=heatmap.dtype)
+    x, y = int(center[0]), int(center[1])
+    radius = int(radius)
+    height, width = heatmap.shape[0:2]
+    left, right = min(x, radius), min(width - x, radius + 1)
+    top, bottom = min(y, radius), min(height - y, radius + 1)
+    masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right]
+    masked_gaussian = gaussian[radius - top:radius + bottom, radius - left:
+                               radius + right]
+    # assert masked_gaussian.equal(1).float().sum() == 1
+    if min(masked_gaussian.shape) > 0 and min(masked_heatmap.shape) > 0:
+        heatmap[y - top:y + bottom, x - left:x + right] = paddle.maximum(
+            masked_heatmap, masked_gaussian * k)
+    return heatmap
+@register
+class PETRHead(nn.Layer):
+    """Head of `End-to-End Multi-Person Pose Estimation with Transformers`.
+    Args:
+        num_classes (int): Number of categories excluding the background.
+        in_channels (int): Number of channels in the input feature map.
+        num_query (int): Number of query in Transformer.
+        num_kpt_fcs (int, optional): Number of fully-connected layers used in
+            `FFN`, which is then used for the keypoint regression head.
+            Default 2.
+        transformer (obj:`mmcv.ConfigDict`|dict): ConfigDict is used for
+            building the Encoder and Decoder. Default: None.
+        sync_cls_avg_factor (bool): Whether to sync the avg_factor of
+            all ranks. Default to False.
+        positional_encoding (obj:`mmcv.ConfigDict`|dict):
+            Config for position encoding.
+        loss_cls (obj:`mmcv.ConfigDict`|dict): Config of the
+            classification loss. Default `CrossEntropyLoss`.
+        loss_kpt (obj:`mmcv.ConfigDict`|dict): Config of the
+            regression loss. Default `L1Loss`.
+        loss_oks (obj:`mmcv.ConfigDict`|dict): Config of the
+            regression oks loss. Default `OKSLoss`.
+        loss_hm (obj:`mmcv.ConfigDict`|dict): Config of the
+            regression heatmap loss. Default `NegLoss`.
+        as_two_stage (bool) : Whether to generate the proposal from
+            the outputs of encoder.
+        with_kpt_refine (bool): Whether to refine the reference points
+            in the decoder. Defaults to True.
+        test_cfg (obj:`mmcv.ConfigDict`|dict): Testing config of
+            transformer head.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
+            Default: None.
+    """
+    __inject__ = [
+        "transformer", "positional_encoding", "assigner", "sampler", "loss_cls",
+        "loss_kpt", "loss_oks", "loss_hm", "loss_kpt_rpn", "loss_kpt_refine",
+        "loss_oks_refine"
+    ]
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 num_query=100,
+                 num_kpt_fcs=2,
+                 num_keypoints=17,
+                 transformer=None,
+                 sync_cls_avg_factor=True,
+                 positional_encoding='SinePositionalEncoding',
+                 loss_cls='FocalLoss',
+                 loss_kpt='L1Loss',
+                 loss_oks='OKSLoss',
+                 loss_hm='CenterFocalLoss',
+                 with_kpt_refine=True,
+                 assigner='PoseHungarianAssigner',
+                 sampler='PseudoSampler',
+                 loss_kpt_rpn='L1Loss',
+                 loss_kpt_refine='L1Loss',
+                 loss_oks_refine='opera.OKSLoss',
+                 test_cfg=dict(max_per_img=100),
+                 init_cfg=None,
+                 **kwargs):
+        # NOTE here use `AnchorFreeHead` instead of `TransformerHead`,
+        # since it brings inconvenience when the initialization of
+        # `AnchorFreeHead` is called.
+        super().__init__()
+        self.bg_cls_weight = 0
+        self.sync_cls_avg_factor = sync_cls_avg_factor
+        self.assigner = assigner
+        self.sampler = sampler
+        self.num_query = num_query
+        self.num_classes = num_classes
+        self.in_channels = in_channels
+        self.num_kpt_fcs = num_kpt_fcs
+        self.test_cfg = test_cfg
+        self.fp16_enabled = False
+        self.as_two_stage = transformer.as_two_stage
+        self.with_kpt_refine = with_kpt_refine
+        self.num_keypoints = num_keypoints
+        self.loss_cls = loss_cls
+        self.loss_kpt = loss_kpt
+        self.loss_kpt_rpn = loss_kpt_rpn
+        self.loss_kpt_refine = loss_kpt_refine
+        self.loss_oks = loss_oks
+        self.loss_oks_refine = loss_oks_refine
+        self.loss_hm = loss_hm
+        if self.loss_cls.use_sigmoid:
+            self.cls_out_channels = num_classes
+        else:
+            self.cls_out_channels = num_classes + 1
+        self.positional_encoding = positional_encoding
+        self.transformer = transformer
+        self.embed_dims = self.transformer.embed_dims
+        # assert 'num_feats' in positional_encoding
+        num_feats = positional_encoding.num_pos_feats
+        assert num_feats * 2 == self.embed_dims, 'embed_dims should' \
+            f' be exactly 2 times of num_feats. Found {self.embed_dims}' \
+            f' and {num_feats}.'
+        self._init_layers()
+        self.init_weights()
+    def _init_layers(self):
+        """Initialize classification branch and keypoint branch of head."""
+        fc_cls = nn.Linear(self.embed_dims, self.cls_out_channels)
+        kpt_branch = []
+        kpt_branch.append(nn.Linear(self.embed_dims, 512))
+        kpt_branch.append(nn.ReLU())
+        for _ in range(self.num_kpt_fcs):
+            kpt_branch.append(nn.Linear(512, 512))
+            kpt_branch.append(nn.ReLU())
+        kpt_branch.append(nn.Linear(512, 2 * self.num_keypoints))
+        kpt_branch = nn.Sequential(*kpt_branch)
+        def _get_clones(module, N):
+            return nn.LayerList([copy.deepcopy(module) for i in range(N)])
+        # last kpt_branch is used to generate proposal from
+        # encode feature map when as_two_stage is True.
+        num_pred = (self.transformer.decoder.num_layers + 1) if \
+            self.as_two_stage else self.transformer.decoder.num_layers
+        if self.with_kpt_refine:
+            self.cls_branches = _get_clones(fc_cls, num_pred)
+            self.kpt_branches = _get_clones(kpt_branch, num_pred)
+        else:
+            self.cls_branches = nn.LayerList([fc_cls for _ in range(num_pred)])
+            self.kpt_branches = nn.LayerList(
+                [kpt_branch for _ in range(num_pred)])
+        self.query_embedding = nn.Embedding(self.num_query, self.embed_dims * 2)
+        refine_kpt_branch = []
+        for _ in range(self.num_kpt_fcs):
+            refine_kpt_branch.append(
+                nn.Linear(self.embed_dims, self.embed_dims))
+            refine_kpt_branch.append(nn.ReLU())
+        refine_kpt_branch.append(nn.Linear(self.embed_dims, 2))
+        refine_kpt_branch = nn.Sequential(*refine_kpt_branch)
+        if self.with_kpt_refine:
+            num_pred = self.transformer.refine_decoder.num_layers
+            self.refine_kpt_branches = _get_clones(refine_kpt_branch, num_pred)
+        self.fc_hm = nn.Linear(self.embed_dims, self.num_keypoints)
+    def init_weights(self):
+        """Initialize weights of the PETR head."""
+        self.transformer.init_weights()
+        if self.loss_cls.use_sigmoid:
+            bias_init = bias_init_with_prob(0.01)
+            for m in self.cls_branches:
+                constant_(m.bias, bias_init)
+        for m in self.kpt_branches:
+            constant_(m[-1].bias, 0)
+        # initialization of keypoint refinement branch
+        if self.with_kpt_refine:
+            for m in self.refine_kpt_branches:
+                constant_(m[-1].bias, 0)
+        # initialize bias for heatmap prediction
+        bias_init = bias_init_with_prob(0.1)
+        normal_(self.fc_hm.weight, std=0.01)
+        constant_(self.fc_hm.bias, bias_init)
+    def forward(self, mlvl_feats, img_metas):
+        """Forward function.
+        Args:
+            mlvl_feats (tuple[Tensor]): Features from the upstream
+                network, each is a 4D-tensor with shape
+                (N, C, H, W).
+            img_metas (list[dict]): List of image information.
+        Returns:
+            outputs_classes (Tensor): Outputs from the classification head,
+                shape [nb_dec, bs, num_query, cls_out_channels]. Note
+                cls_out_channels should include background.
+            outputs_kpts (Tensor): Sigmoid outputs from the regression
+                head with normalized coordinate format (cx, cy, w, h).
+                Shape [nb_dec, bs, num_query, K*2].
+            enc_outputs_class (Tensor): The score of each point on encode
+                feature map, has shape (N, h*w, num_class). Only when
+                as_two_stage is Ture it would be returned, otherwise
+                `None` would be returned.
+            enc_outputs_kpt (Tensor): The proposal generate from the
+                encode feature map, has shape (N, h*w, K*2). Only when
+                as_two_stage is Ture it would be returned, otherwise
+                `None` would be returned.
+        """
+        batch_size = mlvl_feats[0].shape[0]
+        input_img_h, input_img_w = img_metas[0]['batch_input_shape']
+        img_masks = paddle.zeros(
+            (batch_size, input_img_h, input_img_w), dtype=mlvl_feats[0].dtype)
+        for img_id in range(batch_size):
+            img_h, img_w, _ = img_metas[img_id]['img_shape']
+            img_masks[img_id, :img_h, :img_w] = 1
+        mlvl_masks = []
+        mlvl_positional_encodings = []
+        for feat in mlvl_feats:
+            mlvl_masks.append(
+                F.interpolate(
+                    img_masks[None], size=feat.shape[-2:]).squeeze(0))
+            mlvl_positional_encodings.append(
+                self.positional_encoding(mlvl_masks[-1]).transpose(
+                    [0, 3, 1, 2]))
+        query_embeds = self.query_embedding.weight
+        hs, init_reference, inter_references, \
+            enc_outputs_class, enc_outputs_kpt, hm_proto, memory = \
+                self.transformer(
+                    mlvl_feats,
+                    mlvl_masks,
+                    query_embeds,
+                    mlvl_positional_encodings,
+                    kpt_branches=self.kpt_branches \
+                        if self.with_kpt_refine else None,  # noqa:E501
+                    cls_branches=self.cls_branches \
+                        if self.as_two_stage else None  # noqa:E501
+            )
+        outputs_classes = []
+        outputs_kpts = []
+        for lvl in range(hs.shape[0]):
+            if lvl == 0:
+                reference = init_reference
+            else:
+                reference = inter_references[lvl - 1]
+            reference = inverse_sigmoid(reference)
+            outputs_class = self.cls_branches[lvl](hs[lvl])
+            tmp_kpt = self.kpt_branches[lvl](hs[lvl])
+            assert reference.shape[-1] == self.num_keypoints * 2
+            tmp_kpt += reference
+            outputs_kpt = F.sigmoid(tmp_kpt)
+            outputs_classes.append(outputs_class)
+            outputs_kpts.append(outputs_kpt)
+        outputs_classes = paddle.stack(outputs_classes)
+        outputs_kpts = paddle.stack(outputs_kpts)
+        if hm_proto is not None:
+            # get heatmap prediction (training phase)
+            hm_memory, hm_mask = hm_proto
+            hm_pred = self.fc_hm(hm_memory)
+            hm_proto = (hm_pred.transpose((0, 3, 1, 2)), hm_mask)
+        if self.as_two_stage:
+            return outputs_classes, outputs_kpts, \
+                enc_outputs_class, F.sigmoid(enc_outputs_kpt), \
+                hm_proto, memory, mlvl_masks
+        else:
+            raise RuntimeError('only "as_two_stage=True" is supported.')
+    def forward_refine(self, memory, mlvl_masks, refine_targets, losses,
+                       img_metas):
+        """Forward function.
+        Args:
+            mlvl_masks (tuple[Tensor]): The key_padding_mask from
+                different level used for encoder and decoder,
+                each is a 3D-tensor with shape (bs, H, W).
+            losses (dict[str, Tensor]): A dictionary of loss components.
+            img_metas (list[dict]): List of image information.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        kpt_preds, kpt_targets, area_targets, kpt_weights = refine_targets
+        pos_inds = kpt_weights.sum(-1) > 0
+        if not pos_inds.any():
+            pos_kpt_preds = paddle.zeros_like(kpt_preds[:1])
+            pos_img_inds = paddle.zeros([1], dtype="int64")
+        else:
+            pos_kpt_preds = kpt_preds[pos_inds]
+            pos_img_inds = (pos_inds.nonzero() /
+                            self.num_query).squeeze(1).astype("int64")
+        hs, init_reference, inter_references = self.transformer.forward_refine(
+            mlvl_masks,
+            memory,
+            pos_kpt_preds.detach(),
+            pos_img_inds,
+            kpt_branches=self.refine_kpt_branches
+            if self.with_kpt_refine else None,  # noqa:E501
+        )
+        outputs_kpts = []
+        for lvl in range(hs.shape[0]):
+            if lvl == 0:
+                reference = init_reference
+            else:
+                reference = inter_references[lvl - 1]
+            reference = inverse_sigmoid(reference)
+            tmp_kpt = self.refine_kpt_branches[lvl](hs[lvl])
+            assert reference.shape[-1] == 2
+            tmp_kpt += reference
+            outputs_kpt = F.sigmoid(tmp_kpt)
+            outputs_kpts.append(outputs_kpt)
+        outputs_kpts = paddle.stack(outputs_kpts)
+        if not self.training:
+            return outputs_kpts
+        num_valid_kpt = paddle.clip(
+            reduce_mean(kpt_weights.sum()), min=1).item()
+        num_total_pos = paddle.to_tensor(
+            [outputs_kpts.shape[1]], dtype=kpt_weights.dtype)
+        num_total_pos = paddle.clip(reduce_mean(num_total_pos), min=1).item()
+        if not pos_inds.any():
+            for i, kpt_refine_preds in enumerate(outputs_kpts):
+                loss_kpt = loss_oks = kpt_refine_preds.sum() * 0
+                losses[f'd{i}.loss_kpt_refine'] = loss_kpt
+                losses[f'd{i}.loss_oks_refine'] = loss_oks
+                continue
+            return losses
+        batch_size = mlvl_masks[0].shape[0]
+        factors = []
+        for img_id in range(batch_size):
+            img_h, img_w, _ = img_metas[img_id]['img_shape']
+            factor = paddle.to_tensor(
+                [img_w, img_h, img_w, img_h],
+                dtype="float32").squeeze(-1).unsqueeze(0).tile(
+                    (self.num_query, 1))
+            factors.append(factor)
+        factors = paddle.concat(factors, 0)
+        factors = factors[pos_inds][:, :2].tile((1, kpt_preds.shape[-1] // 2))
+        pos_kpt_weights = kpt_weights[pos_inds]
+        pos_kpt_targets = kpt_targets[pos_inds]
+        pos_kpt_targets_scaled = pos_kpt_targets * factors
+        pos_areas = area_targets[pos_inds]
+        pos_valid = kpt_weights[pos_inds][:, 0::2]
+        for i, kpt_refine_preds in enumerate(outputs_kpts):
+            if not pos_inds.any():
+                print("refine kpt and oks skip")
+                loss_kpt = loss_oks = kpt_refine_preds.sum() * 0
+                losses[f'd{i}.loss_kpt_refine'] = loss_kpt
+                losses[f'd{i}.loss_oks_refine'] = loss_oks
+                continue
+            # kpt L1 Loss
+            pos_refine_preds = kpt_refine_preds.reshape(
+                (kpt_refine_preds.shape[0], -1))
+            loss_kpt = self.loss_kpt_refine(
+                pos_refine_preds,
+                pos_kpt_targets,
+                pos_kpt_weights,
+                avg_factor=num_valid_kpt)
+            losses[f'd{i}.loss_kpt_refine'] = loss_kpt
+            # kpt oks loss
+            pos_refine_preds_scaled = pos_refine_preds * factors
+            assert (pos_areas > 0).all()
+            loss_oks = self.loss_oks_refine(
+                pos_refine_preds_scaled,
+                pos_kpt_targets_scaled,
+                pos_valid,
+                pos_areas,
+                avg_factor=num_total_pos)
+            losses[f'd{i}.loss_oks_refine'] = loss_oks
+        return losses
+    # over-write because img_metas are needed as inputs for bbox_head.
+    def forward_train(self,
+                      x,
+                      img_metas,
+                      gt_bboxes,
+                      gt_labels=None,
+                      gt_keypoints=None,
+                      gt_areas=None,
+                      gt_bboxes_ignore=None,
+                      proposal_cfg=None,
+                      **kwargs):
+        """Forward function for training mode.
+        Args:
+            x (list[Tensor]): Features from backbone.
+            img_metas (list[dict]): Meta information of each image, e.g.,
+                image size, scaling factor, etc.
+            gt_bboxes (list[Tensor]): Ground truth bboxes of the image,
+                shape (num_gts, 4).
+            gt_labels (list[Tensor]): Ground truth labels of each box,
+                shape (num_gts,).
+            gt_keypoints (list[Tensor]): Ground truth keypoints of the image,
+                shape (num_gts, K*3).
+            gt_areas (list[Tensor]): Ground truth mask areas of each box,
+                shape (num_gts,).
+            gt_bboxes_ignore (list[Tensor]): Ground truth bboxes to be
+                ignored, shape (num_ignored_gts, 4).
+            proposal_cfg (mmcv.Config): Test / postprocessing configuration,
+                if None, test_cfg would be used.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert proposal_cfg is None, '"proposal_cfg" must be None'
+        outs = self(x, img_metas)
+        memory, mlvl_masks = outs[-2:]
+        outs = outs[:-2]
+        if gt_labels is None:
+            loss_inputs = outs + (gt_bboxes, gt_keypoints, gt_areas, img_metas)
+        else:
+            loss_inputs = outs + (gt_bboxes, gt_labels, gt_keypoints, gt_areas,
+                                  img_metas)
+        losses_and_targets = self.loss(
+            *loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
+        # losses = losses_and_targets
+        losses, refine_targets = losses_and_targets
+        # get pose refinement loss
+        losses = self.forward_refine(memory, mlvl_masks, refine_targets, losses,
+                                     img_metas)
+        return losses
+    def loss(self,
+             all_cls_scores,
+             all_kpt_preds,
+             enc_cls_scores,
+             enc_kpt_preds,
+             enc_hm_proto,
+             gt_bboxes_list,
+             gt_labels_list,
+             gt_keypoints_list,
+             gt_areas_list,
+             img_metas,
+             gt_bboxes_ignore=None):
+        """Loss function.
+        Args:
+            all_cls_scores (Tensor): Classification score of all
+                decoder layers, has shape
+                [nb_dec, bs, num_query, cls_out_channels].
+            all_kpt_preds (Tensor): Sigmoid regression
+                outputs of all decode layers. Each is a 4D-tensor with
+                normalized coordinate format (x_{i}, y_{i}) and shape
+                [nb_dec, bs, num_query, K*2].
+            enc_cls_scores (Tensor): Classification scores of
+                points on encode feature map, has shape
+                (N, h*w, num_classes). Only be passed when as_two_stage is
+                True, otherwise is None.
+            enc_kpt_preds (Tensor): Regression results of each points
+                on the encode feature map, has shape (N, h*w, K*2). Only be
+                passed when as_two_stage is True, otherwise is None.
+            gt_bboxes_list (list[Tensor]): Ground truth bboxes for each image
+                with shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            gt_keypoints_list (list[Tensor]): Ground truth keypoints for each
+                image with shape (num_gts, K*3) in [p^{1}_x, p^{1}_y, p^{1}_v,
+                    ..., p^{K}_x, p^{K}_y, p^{K}_v] format.
+            gt_areas_list (list[Tensor]): Ground truth mask areas for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+            gt_bboxes_ignore (list[Tensor], optional): Bounding boxes
+                which can be ignored for each image. Default None.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components.
+        """
+        assert gt_bboxes_ignore is None, \
+            f'{self.__class__.__name__} only supports ' \
+            f'for gt_bboxes_ignore setting to None.'
+        num_dec_layers = len(all_cls_scores)
+        all_gt_labels_list = [gt_labels_list for _ in range(num_dec_layers)]
+        all_gt_keypoints_list = [
+            gt_keypoints_list for _ in range(num_dec_layers)
+        ]
+        all_gt_areas_list = [gt_areas_list for _ in range(num_dec_layers)]
+        img_metas_list = [img_metas for _ in range(num_dec_layers)]
+        losses_cls, losses_kpt, losses_oks, kpt_preds_list, kpt_targets_list, \
+            area_targets_list, kpt_weights_list = multi_apply(
+                self.loss_single, all_cls_scores, all_kpt_preds,
+                all_gt_labels_list, all_gt_keypoints_list,
+                all_gt_areas_list, img_metas_list)
+        loss_dict = dict()
+        # loss of proposal generated from encode feature map.
+        if enc_cls_scores is not None:
+            binary_labels_list = [
+                paddle.zeros_like(gt_labels_list[i])
+                for i in range(len(img_metas))
+            ]
+            enc_loss_cls, enc_losses_kpt = \
+                self.loss_single_rpn(
+                    enc_cls_scores, enc_kpt_preds, binary_labels_list,
+                    gt_keypoints_list, gt_areas_list, img_metas)
+            loss_dict['enc_loss_cls'] = enc_loss_cls
+            loss_dict['enc_loss_kpt'] = enc_losses_kpt
+        # loss from the last decoder layer
+        loss_dict['loss_cls'] = losses_cls[-1]
+        loss_dict['loss_kpt'] = losses_kpt[-1]
+        loss_dict['loss_oks'] = losses_oks[-1]
+        # loss from other decoder layers
+        num_dec_layer = 0
+        for loss_cls_i, loss_kpt_i, loss_oks_i in zip(
+                losses_cls[:-1], losses_kpt[:-1], losses_oks[:-1]):
+            loss_dict[f'd{num_dec_layer}.loss_cls'] = loss_cls_i
+            loss_dict[f'd{num_dec_layer}.loss_kpt'] = loss_kpt_i
+            loss_dict[f'd{num_dec_layer}.loss_oks'] = loss_oks_i
+            num_dec_layer += 1
+        # losses of heatmap generated from P3 feature map
+        hm_pred, hm_mask = enc_hm_proto
+        loss_hm = self.loss_heatmap(hm_pred, hm_mask, gt_keypoints_list,
+                                    gt_labels_list, gt_bboxes_list)
+        loss_dict['loss_hm'] = loss_hm
+        return loss_dict, (kpt_preds_list[-1], kpt_targets_list[-1],
+                           area_targets_list[-1], kpt_weights_list[-1])
+    def loss_heatmap(self, hm_pred, hm_mask, gt_keypoints, gt_labels,
+                     gt_bboxes):
+        assert hm_pred.shape[-2:] == hm_mask.shape[-2:]
+        num_img, _, h, w = hm_pred.shape
+        # placeholder of heatmap target (Gaussian distribution)
+        hm_target = paddle.zeros(hm_pred.shape, hm_pred.dtype)
+        for i, (gt_label, gt_bbox, gt_keypoint
+                ) in enumerate(zip(gt_labels, gt_bboxes, gt_keypoints)):
+            if gt_label.shape[0] == 0:
+                continue
+            gt_keypoint = gt_keypoint.reshape((gt_keypoint.shape[0], -1,
+                                               3)).clone()
+            gt_keypoint[..., :2] /= 8
+            assert gt_keypoint[..., 0].max() <= w + 0.5  # new coordinate system
+            assert gt_keypoint[..., 1].max() <= h + 0.5  # new coordinate system
+            gt_bbox /= 8
+            gt_w = gt_bbox[:, 2] - gt_bbox[:, 0]
+            gt_h = gt_bbox[:, 3] - gt_bbox[:, 1]
+            for j in range(gt_label.shape[0]):
+                # get heatmap radius
+                kp_radius = paddle.clip(
+                    paddle.floor(
+                        gaussian_radius(
+                            (gt_h[j], gt_w[j]), min_overlap=0.9)),
+                    min=0,
+                    max=3)
+                for k in range(self.num_keypoints):
+                    if gt_keypoint[j, k, 2] > 0:
+                        gt_kp = gt_keypoint[j, k, :2]
+                        gt_kp_int = paddle.floor(gt_kp)
+                        hm_target[i, k] = draw_umich_gaussian(
+                            hm_target[i, k], gt_kp_int, kp_radius)
+        # compute heatmap loss
+        hm_pred = paddle.clip(
+            F.sigmoid(hm_pred), min=1e-4, max=1 - 1e-4)  # refer to CenterNet
+        loss_hm = self.loss_hm(
+            hm_pred,
+            hm_target.detach(),
+            mask=~hm_mask.astype("bool").unsqueeze(1))
+        return loss_hm
+    def loss_single(self, cls_scores, kpt_preds, gt_labels_list,
+                    gt_keypoints_list, gt_areas_list, img_metas):
+        """Loss function for outputs from a single decoder layer of a single
+        feature level.
+        Args:
+            cls_scores (Tensor): Box score logits from a single decoder layer
+                for all images. Shape [bs, num_query, cls_out_channels].
+            kpt_preds (Tensor): Sigmoid outputs from a single decoder layer
+                for all images, with normalized coordinate (x_{i}, y_{i}) and
+                shape [bs, num_query, K*2].
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            gt_keypoints_list (list[Tensor]): Ground truth keypoints for each
+                image with shape (num_gts, K*3) in [p^{1}_x, p^{1}_y, p^{1}_v,
+                ..., p^{K}_x, p^{K}_y, p^{K}_v] format.
+            gt_areas_list (list[Tensor]): Ground truth mask areas for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components for outputs from
+                a single decoder layer.
+        """
+        num_imgs = cls_scores.shape[0]
+        cls_scores_list = [cls_scores[i] for i in range(num_imgs)]
+        kpt_preds_list = [kpt_preds[i] for i in range(num_imgs)]
+        cls_reg_targets = self.get_targets(cls_scores_list, kpt_preds_list,
+                                           gt_labels_list, gt_keypoints_list,
+                                           gt_areas_list, img_metas)
+        (labels_list, label_weights_list, kpt_targets_list, kpt_weights_list,
+         area_targets_list, num_total_pos, num_total_neg) = cls_reg_targets
+        labels = paddle.concat(labels_list, 0)
+        label_weights = paddle.concat(label_weights_list, 0)
+        kpt_targets = paddle.concat(kpt_targets_list, 0)
+        kpt_weights = paddle.concat(kpt_weights_list, 0)
+        area_targets = paddle.concat(area_targets_list, 0)
+        # classification loss
+        cls_scores = cls_scores.reshape((-1, self.cls_out_channels))
+        # construct weighted avg_factor to match with the official DETR repo
+        cls_avg_factor = num_total_pos * 1.0 + \
+            num_total_neg * self.bg_cls_weight
+        if self.sync_cls_avg_factor:
+            cls_avg_factor = reduce_mean(
+                paddle.to_tensor(
+                    [cls_avg_factor], dtype=cls_scores.dtype))
+        cls_avg_factor = max(cls_avg_factor, 1)
+        loss_cls = self.loss_cls(
+            cls_scores, labels, label_weights, avg_factor=cls_avg_factor)
+        # Compute the average number of gt keypoints accross all gpus, for
+        # normalization purposes
+        num_total_pos = paddle.to_tensor([num_total_pos], dtype=loss_cls.dtype)
+        num_total_pos = paddle.clip(reduce_mean(num_total_pos), min=1).item()
+        # construct factors used for rescale keypoints
+        factors = []
+        for img_meta, kpt_pred in zip(img_metas, kpt_preds):
+            img_h, img_w, _ = img_meta['img_shape']
+            factor = paddle.to_tensor(
+                [img_w, img_h, img_w, img_h],
+                dtype=kpt_pred.dtype).squeeze().unsqueeze(0).tile(
+                    (kpt_pred.shape[0], 1))
+            factors.append(factor)
+        factors = paddle.concat(factors, 0)
+        # keypoint regression loss
+        kpt_preds = kpt_preds.reshape((-1, kpt_preds.shape[-1]))
+        num_valid_kpt = paddle.clip(
+            reduce_mean(kpt_weights.sum()), min=1).item()
+        # assert num_valid_kpt == (kpt_targets>0).sum().item()
+        loss_kpt = self.loss_kpt(
+            kpt_preds,
+            kpt_targets.detach(),
+            kpt_weights.detach(),
+            avg_factor=num_valid_kpt)
+        # keypoint oks loss
+        pos_inds = kpt_weights.sum(-1) > 0
+        if not pos_inds.any():
+            loss_oks = kpt_preds.sum() * 0
+        else:
+            factors = factors[pos_inds][:, :2].tile((
+                (1, kpt_preds.shape[-1] // 2)))
+            pos_kpt_preds = kpt_preds[pos_inds] * factors
+            pos_kpt_targets = kpt_targets[pos_inds] * factors
+            pos_areas = area_targets[pos_inds]
+            pos_valid = kpt_weights[pos_inds][..., 0::2]
+            assert (pos_areas > 0).all()
+            loss_oks = self.loss_oks(
+                pos_kpt_preds,
+                pos_kpt_targets,
+                pos_valid,
+                pos_areas,
+                avg_factor=num_total_pos)
+        return loss_cls, loss_kpt, loss_oks, kpt_preds, kpt_targets, \
+            area_targets, kpt_weights
+    def get_targets(self, cls_scores_list, kpt_preds_list, gt_labels_list,
+                    gt_keypoints_list, gt_areas_list, img_metas):
+        """Compute regression and classification targets for a batch image.
+        Outputs from a single decoder layer of a single feature level are used.
+        Args:
+            cls_scores_list (list[Tensor]): Box score logits from a single
+                decoder layer for each image with shape [num_query,
+                cls_out_channels].
+            kpt_preds_list (list[Tensor]): Sigmoid outputs from a single
+                decoder layer for each image, with normalized coordinate
+                (x_{i}, y_{i}) and shape [num_query, K*2].
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            gt_keypoints_list (list[Tensor]): Ground truth keypoints for each
+                image with shape (num_gts, K*3).
+            gt_areas_list (list[Tensor]): Ground truth mask areas for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+        Returns:
+            tuple: a tuple containing the following targets.
+                - labels_list (list[Tensor]): Labels for all images.
+                - label_weights_list (list[Tensor]): Label weights for all
+                    images.
+                - kpt_targets_list (list[Tensor]): Keypoint targets for all
+                    images.
+                - kpt_weights_list (list[Tensor]): Keypoint weights for all
+                    images.
+                - area_targets_list (list[Tensor]): area targets for all
+                    images.
+                - num_total_pos (int): Number of positive samples in all
+                    images.
+                - num_total_neg (int): Number of negative samples in all
+                    images.
+        """
+        (labels_list, label_weights_list, kpt_targets_list, kpt_weights_list,
+         area_targets_list, pos_inds_list, neg_inds_list) = multi_apply(
+             self._get_target_single, cls_scores_list, kpt_preds_list,
+             gt_labels_list, gt_keypoints_list, gt_areas_list, img_metas)
+        num_total_pos = sum((inds.numel() for inds in pos_inds_list))
+        num_total_neg = sum((inds.numel() for inds in neg_inds_list))
+        return (labels_list, label_weights_list, kpt_targets_list,
+                kpt_weights_list, area_targets_list, num_total_pos,
+                num_total_neg)
+    def _get_target_single(self, cls_score, kpt_pred, gt_labels, gt_keypoints,
+                           gt_areas, img_meta):
+        """Compute regression and classification targets for one image.
+        Outputs from a single decoder layer of a single feature level are used.
+        Args:
+            cls_score (Tensor): Box score logits from a single decoder layer
+                for one image. Shape [num_query, cls_out_channels].
+            kpt_pred (Tensor): Sigmoid outputs from a single decoder layer
+                for one image, with normalized coordinate (x_{i}, y_{i}) and
+                shape [num_query, K*2].
+            gt_labels (Tensor): Ground truth class indices for one image
+                with shape (num_gts, ).
+            gt_keypoints (Tensor): Ground truth keypoints for one image with
+                shape (num_gts, K*3) in [p^{1}_x, p^{1}_y, p^{1}_v, ..., \
+                    p^{K}_x, p^{K}_y, p^{K}_v] format.
+            gt_areas (Tensor): Ground truth mask areas for one image
+                with shape (num_gts, ).
+            img_meta (dict): Meta information for one image.
+        Returns:
+            tuple[Tensor]: a tuple containing the following for one image.
+                - labels (Tensor): Labels of each image.
+                - label_weights (Tensor): Label weights of each image.
+                - kpt_targets (Tensor): Keypoint targets of each image.
+                - kpt_weights (Tensor): Keypoint weights of each image.
+                - area_targets (Tensor): Area targets of each image.
+                - pos_inds (Tensor): Sampled positive indices for each image.
+                - neg_inds (Tensor): Sampled negative indices for each image.
+        """
+        num_bboxes = kpt_pred.shape[0]
+        # assigner and sampler
+        assign_result = self.assigner.assign(cls_score, kpt_pred, gt_labels,
+                                             gt_keypoints, gt_areas, img_meta)
+        sampling_result = self.sampler.sample(assign_result, kpt_pred,
+                                              gt_keypoints)
+        pos_inds = sampling_result.pos_inds
+        neg_inds = sampling_result.neg_inds
+        # label targets
+        labels = paddle.full((num_bboxes, ), self.num_classes, dtype="int64")
+        label_weights = paddle.ones((num_bboxes, ), dtype=gt_labels.dtype)
+        kpt_targets = paddle.zeros_like(kpt_pred)
+        kpt_weights = paddle.zeros_like(kpt_pred)
+        area_targets = paddle.zeros((kpt_pred.shape[0], ), dtype=kpt_pred.dtype)
+        if pos_inds.size == 0:
+            return (labels, label_weights, kpt_targets, kpt_weights,
+                    area_targets, pos_inds, neg_inds)
+        labels[pos_inds] = gt_labels[sampling_result.pos_assigned_gt_inds][
+            ..., 0].astype("int64")
+        img_h, img_w, _ = img_meta['img_shape']
+        # keypoint targets
+        pos_gt_kpts = gt_keypoints[sampling_result.pos_assigned_gt_inds]
+        pos_gt_kpts = pos_gt_kpts.reshape(
+            (len(sampling_result.pos_assigned_gt_inds), -1, 3))
+        valid_idx = pos_gt_kpts[:, :, 2] > 0
+        pos_kpt_weights = kpt_weights[pos_inds].reshape(
+            (pos_gt_kpts.shape[0], kpt_weights.shape[-1] // 2, 2))
+        # pos_kpt_weights[valid_idx][...] = 1.0
+        pos_kpt_weights = masked_fill(pos_kpt_weights,
+                                      valid_idx.unsqueeze(-1), 1.0)
+        kpt_weights[pos_inds] = pos_kpt_weights.reshape(
+            (pos_kpt_weights.shape[0], kpt_pred.shape[-1]))
+        factor = paddle.to_tensor(
+            [img_w, img_h], dtype=kpt_pred.dtype).squeeze().unsqueeze(0)
+        pos_gt_kpts_normalized = pos_gt_kpts[..., :2]
+        pos_gt_kpts_normalized[..., 0] = pos_gt_kpts_normalized[..., 0] / \
+            factor[:, 0:1]
+        pos_gt_kpts_normalized[..., 1] = pos_gt_kpts_normalized[..., 1] / \
+            factor[:, 1:2]
+        kpt_targets[pos_inds] = pos_gt_kpts_normalized.reshape(
+            (pos_gt_kpts.shape[0], kpt_pred.shape[-1]))
+        pos_gt_areas = gt_areas[sampling_result.pos_assigned_gt_inds][..., 0]
+        area_targets[pos_inds] = pos_gt_areas
+        return (labels, label_weights, kpt_targets, kpt_weights, area_targets,
+                pos_inds, neg_inds)
+    def loss_single_rpn(self, cls_scores, kpt_preds, gt_labels_list,
+                        gt_keypoints_list, gt_areas_list, img_metas):
+        """Loss function for outputs from a single decoder layer of a single
+        feature level.
+        Args:
+            cls_scores (Tensor): Box score logits from a single decoder layer
+                for all images. Shape [bs, num_query, cls_out_channels].
+            kpt_preds (Tensor): Sigmoid outputs from a single decoder layer
+                for all images, with normalized coordinate (x_{i}, y_{i}) and
+                shape [bs, num_query, K*2].
+            gt_labels_list (list[Tensor]): Ground truth class indices for each
+                image with shape (num_gts, ).
+            gt_keypoints_list (list[Tensor]): Ground truth keypoints for each
+                image with shape (num_gts, K*3) in [p^{1}_x, p^{1}_y, p^{1}_v,
+                ..., p^{K}_x, p^{K}_y, p^{K}_v] format.
+            gt_areas_list (list[Tensor]): Ground truth mask areas for each
+                image with shape (num_gts, ).
+            img_metas (list[dict]): List of image meta information.
+        Returns:
+            dict[str, Tensor]: A dictionary of loss components for outputs from
+                a single decoder layer.
+        """
+        num_imgs = cls_scores.shape[0]
+        cls_scores_list = [cls_scores[i] for i in range(num_imgs)]
+        kpt_preds_list = [kpt_preds[i] for i in range(num_imgs)]
+        cls_reg_targets = self.get_targets(cls_scores_list, kpt_preds_list,
+                                           gt_labels_list, gt_keypoints_list,
+                                           gt_areas_list, img_metas)
+        (labels_list, label_weights_list, kpt_targets_list, kpt_weights_list,
+         area_targets_list, num_total_pos, num_total_neg) = cls_reg_targets
+        labels = paddle.concat(labels_list, 0)
+        label_weights = paddle.concat(label_weights_list, 0)
+        kpt_targets = paddle.concat(kpt_targets_list, 0)
+        kpt_weights = paddle.concat(kpt_weights_list, 0)
+        # classification loss
+        cls_scores = cls_scores.reshape((-1, self.cls_out_channels))
+        # construct weighted avg_factor to match with the official DETR repo
+        cls_avg_factor = num_total_pos * 1.0 + \
+            num_total_neg * self.bg_cls_weight
+        if self.sync_cls_avg_factor:
+            cls_avg_factor = reduce_mean(
+                paddle.to_tensor(
+                    [cls_avg_factor], dtype=cls_scores.dtype))
+        cls_avg_factor = max(cls_avg_factor, 1)
+        cls_avg_factor = max(cls_avg_factor, 1)
+        loss_cls = self.loss_cls(
+            cls_scores, labels, label_weights, avg_factor=cls_avg_factor)
+        # Compute the average number of gt keypoints accross all gpus, for
+        # normalization purposes
+        # num_total_pos = loss_cls.to_tensor([num_total_pos])
+        # num_total_pos = paddle.clip(reduce_mean(num_total_pos), min=1).item()
+        # keypoint regression loss
+        kpt_preds = kpt_preds.reshape((-1, kpt_preds.shape[-1]))
+        num_valid_kpt = paddle.clip(
+            reduce_mean(kpt_weights.sum()), min=1).item()
+        # assert num_valid_kpt == (kpt_targets>0).sum().item()
+        loss_kpt = self.loss_kpt_rpn(
+            kpt_preds, kpt_targets, kpt_weights, avg_factor=num_valid_kpt)
+        return loss_cls, loss_kpt
+    def get_bboxes(self,
+                   all_cls_scores,
+                   all_kpt_preds,
+                   enc_cls_scores,
+                   enc_kpt_preds,
+                   hm_proto,
+                   memory,
+                   mlvl_masks,
+                   img_metas,
+                   rescale=False):
+        """Transform network outputs for a batch into bbox predictions.
+        Args:
+            all_cls_scores (Tensor): Classification score of all
+                decoder layers, has shape
+                [nb_dec, bs, num_query, cls_out_channels].
+            all_kpt_preds (Tensor): Sigmoid regression
+                outputs of all decode layers. Each is a 4D-tensor with
+                normalized coordinate format (x_{i}, y_{i}) and shape
+                [nb_dec, bs, num_query, K*2].
+            enc_cls_scores (Tensor): Classification scores of points on
+                encode feature map, has shape (N, h*w, num_classes).
+                Only be passed when as_two_stage is True, otherwise is None.
+            enc_kpt_preds (Tensor): Regression results of each points
+                on the encode feature map, has shape (N, h*w, K*2). Only be
+                passed when as_two_stage is True, otherwise is None.
+            img_metas (list[dict]): Meta information of each image.
+            rescale (bool, optional): If True, return boxes in original
+                image space. Defalut False.
+        Returns:
+            list[list[Tensor, Tensor]]: Each item in result_list is 3-tuple.
+                The first item is an (n, 5) tensor, where the first 4 columns
+                are bounding box positions (tl_x, tl_y, br_x, br_y) and the
+                5-th column is a score between 0 and 1. The second item is a
+                (n,) tensor where each item is the predicted class label of
+                the corresponding box. The third item is an (n, K, 3) tensor
+                with [p^{1}_x, p^{1}_y, p^{1}_v, ..., p^{K}_x, p^{K}_y,
+                p^{K}_v] format.
+        """
+        cls_scores = all_cls_scores[-1]
+        kpt_preds = all_kpt_preds[-1]
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score = cls_scores[img_id]
+            kpt_pred = kpt_preds[img_id]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            # TODO: only support single image test
+            # memory_i = memory[:, img_id, :]
+            # mlvl_mask = mlvl_masks[img_id]
+            proposals = self._get_bboxes_single(cls_score, kpt_pred, img_shape,
+                                                scale_factor, memory,
+                                                mlvl_masks, rescale)
+            result_list.append(proposals)
+        return result_list
+    def _get_bboxes_single(self,
+                           cls_score,
+                           kpt_pred,
+                           img_shape,
+                           scale_factor,
+                           memory,
+                           mlvl_masks,
+                           rescale=False):
+        """Transform outputs from the last decoder layer into bbox predictions
+        for each image.
+        Args:
+            cls_score (Tensor): Box score logits from the last decoder layer
+                for each image. Shape [num_query, cls_out_channels].
+            kpt_pred (Tensor): Sigmoid outputs from the last decoder layer
+                for each image, with coordinate format (x_{i}, y_{i}) and
+                shape [num_query, K*2].
+            img_shape (tuple[int]): Shape of input image, (height, width, 3).
+            scale_factor (ndarray, optional): Scale factor of the image arange
+                as (w_scale, h_scale, w_scale, h_scale).
+            rescale (bool, optional): If True, return boxes in original image
+                space. Default False.
+        Returns:
+            tuple[Tensor]: Results of detected bboxes and labels.
+                - det_bboxes: Predicted bboxes with shape [num_query, 5],
+                    where the first 4 columns are bounding box positions
+                    (tl_x, tl_y, br_x, br_y) and the 5-th column are scores
+                    between 0 and 1.
+                - det_labels: Predicted labels of the corresponding box with
+                    shape [num_query].
+                - det_kpts: Predicted keypoints with shape [num_query, K, 3].
+        """
+        assert len(cls_score) == len(kpt_pred)
+        max_per_img = self.test_cfg.get('max_per_img', self.num_query)
+        # exclude background
+        if self.loss_cls.use_sigmoid:
+            cls_score = F.sigmoid(cls_score)
+            scores, indexs = cls_score.reshape([-1]).topk(max_per_img)
+            det_labels = indexs % self.num_classes
+            bbox_index = indexs // self.num_classes
+            kpt_pred = kpt_pred[bbox_index]
+        else:
+            scores, det_labels = F.softmax(cls_score, axis=-1)[..., :-1].max(-1)
+            scores, bbox_index = scores.topk(max_per_img)
+            kpt_pred = kpt_pred[bbox_index]
+            det_labels = det_labels[bbox_index]
+        # ----- results after pose decoder -----
+        # det_kpts = kpt_pred.reshape((kpt_pred.shape[0], -1, 2))
+        # ----- results after joint decoder (default) -----
+        # import time
+        # start = time.time()
+        refine_targets = (kpt_pred, None, None, paddle.ones_like(kpt_pred))
+        refine_outputs = self.forward_refine(memory, mlvl_masks, refine_targets,
+                                             None, None)
+        # end = time.time()
+        # print(f'refine time: {end - start:.6f}')
+        det_kpts = refine_outputs[-1]
+        det_kpts[..., 0] = det_kpts[..., 0] * img_shape[1]
+        det_kpts[..., 1] = det_kpts[..., 1] * img_shape[0]
+        det_kpts[..., 0].clip_(min=0, max=img_shape[1])
+        det_kpts[..., 1].clip_(min=0, max=img_shape[0])
+        if rescale:
+            det_kpts /= paddle.to_tensor(
+                scale_factor[:2],
+                dtype=det_kpts.dtype).unsqueeze(0).unsqueeze(0)
+        # use circumscribed rectangle box of keypoints as det bboxes
+        x1 = det_kpts[..., 0].min(axis=1, keepdim=True)
+        y1 = det_kpts[..., 1].min(axis=1, keepdim=True)
+        x2 = det_kpts[..., 0].max(axis=1, keepdim=True)
+        y2 = det_kpts[..., 1].max(axis=1, keepdim=True)
+        det_bboxes = paddle.concat([x1, y1, x2, y2], axis=1)
+        det_bboxes = paddle.concat((det_bboxes, scores.unsqueeze(1)), -1)
+        det_kpts = paddle.concat(
+            (det_kpts, paddle.ones(
+                det_kpts[..., :1].shape, dtype=det_kpts.dtype)),
+            axis=2)
+        return det_bboxes, det_labels, det_kpts
+    def simple_test(self, feats, img_metas, rescale=False):
+        """Test det bboxes without test-time augmentation.
+        Args:
+            feats (tuple[paddle.Tensor]): Multi-level features from the
+                upstream network, each is a 4D-tensor.
+            img_metas (list[dict]): List of image information.
+            rescale (bool, optional): Whether to rescale the results.
+                Defaults to False.
+        Returns:
+            list[tuple[Tensor, Tensor, Tensor]]: Each item in result_list is
+                3-tuple. The first item is ``bboxes`` with shape (n, 5),
+                where 5 represent (tl_x, tl_y, br_x, br_y, score).
+                The shape of the second tensor in the tuple is ``labels``
+                with shape (n,). The third item is ``kpts`` with shape
+                (n, K, 3), in [p^{1}_x, p^{1}_y, p^{1}_v, p^{K}_x, p^{K}_y,
+                p^{K}_v] format.
+        """
+        # forward of this head requires img_metas
+        outs = self.forward(feats, img_metas)
+        results_list = self.get_bboxes(*outs, img_metas, rescale=rescale)
+        return results_list
+    def get_loss(self, boxes, scores, gt_bbox, gt_class, prior_boxes):
+        return self.loss(boxes, scores, gt_bbox, gt_class, prior_boxes)
--- a/ppdet/modeling/layers.py
+++ b/ppdet/modeling/layers.py
@@ -1135,7 +1135,7 @@ def _convert_attention_mask(attn_mask, dtype):
    """
    return nn.layer.transformer._convert_attention_mask(attn_mask, dtype)
+@register
 class MultiHeadAttention(nn.Layer):
    """
    Attention mapps queries and a set of key-value pairs to outputs, and

--- a/ppdet/modeling/losses/focal_loss.py
+++ b/ppdet/modeling/losses/focal_loss.py
@@ -21,7 +21,7 @@ import paddle.nn.functional as F
 import paddle.nn as nn
 from ppdet.core.workspace import register
-__all__ = ['FocalLoss']
+__all__ = ['FocalLoss', 'Weighted_FocalLoss']
 @register
 class FocalLoss(nn.Layer):
@@ -59,3 +59,80 @@ class FocalLoss(nn.Layer):
            pred, target, alpha=self.alpha, gamma=self.gamma,
            reduction=reduction)
        return loss * self.loss_weight
+@register
+class Weighted_FocalLoss(FocalLoss):
+    """A wrapper around paddle.nn.functional.sigmoid_focal_loss.
+    Args:
+        use_sigmoid (bool): currently only support use_sigmoid=True
+        alpha (float): parameter alpha in Focal Loss
+        gamma (float): parameter gamma in Focal Loss
+        loss_weight (float): final loss will be multiplied by this
+    """
+    def __init__(self,
+                 use_sigmoid=True,
+                 alpha=0.25,
+                 gamma=2.0,
+                 loss_weight=1.0,
+                 reduction="mean"):
+        super(FocalLoss, self).__init__()
+        assert use_sigmoid == True, \
+            'Focal Loss only supports sigmoid at the moment'
+        self.use_sigmoid = use_sigmoid
+        self.alpha = alpha
+        self.gamma = gamma
+        self.loss_weight = loss_weight
+        self.reduction = reduction
+    def forward(self, pred, target, weight=None, avg_factor=None, reduction_override=None):
+        """forward function.
+        Args:
+            pred (Tensor): logits of class prediction, of shape (N, num_classes)
+            target (Tensor): target class label, of shape (N, )
+            reduction (str): the way to reduce loss, one of (none, sum, mean)
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        num_classes = pred.shape[1]
+        target = F.one_hot(target, num_classes + 1).astype(pred.dtype)
+        target = target[:, :-1].detach()
+        loss = F.sigmoid_focal_loss(
+            pred, target, alpha=self.alpha, gamma=self.gamma,
+            reduction='none')
+        if weight is not None:
+            if weight.shape != loss.shape:
+                if weight.shape[0] == loss.shape[0]:
+                    # For most cases, weight is of shape (num_priors, ),
+                    #  which means it does not have the second axis num_class
+                    weight = weight.reshape((-1, 1))
+                else:
+                    # Sometimes, weight per anchor per class is also needed. e.g.
+                    #  in FSAF. But it may be flattened of shape
+                    #  (num_priors x num_class, ), while loss is still of shape
+                    #  (num_priors, num_class).
+                    assert weight.numel() == loss.numel()
+                    weight = weight.reshape((loss.shape[0], -1))
+            assert weight.ndim == loss.ndim
+            loss = loss * weight
+        # if avg_factor is not specified, just reduce the loss
+        if avg_factor is None:
+            if reduction == 'mean':
+                loss = loss.mean()
+            elif reduction == 'sum':
+                loss = loss.sum()
+        else:
+            # if reduction is mean, then average the loss by avg_factor
+            if reduction == 'mean':
+                # Avoid causing ZeroDivisionError when avg_factor is 0.0,
+                # i.e., all labels of an image belong to ignore index.
+                eps = 1e-10
+                loss = loss.sum() / (avg_factor + eps)
+            # if reduction is 'none', then do nothing, otherwise raise an error
+            elif reduction != 'none':
+                raise ValueError('avg_factor can not be used with reduction="sum"')
+        return loss * self.loss_weight
--- a/ppdet/modeling/losses/keypoint_loss.py
+++ b/ppdet/modeling/losses/keypoint_loss.py
@@ -18,12 +18,13 @@ from __future__ import print_function
 from itertools import cycle, islice
 from collections import abc
+import numpy as np
 import paddle
 import paddle.nn as nn
 from ppdet.core.workspace import register, serializable
-__all__ = ['HrHRNetLoss', 'KeyPointMSELoss']
+__all__ = ['HrHRNetLoss', 'KeyPointMSELoss', 'OKSLoss', 'CenterFocalLoss', 'L1Loss']
 @register
@@ -226,3 +227,406 @@ def recursive_sum(inputs):
    if isinstance(inputs, abc.Sequence):
        return sum([recursive_sum(x) for x in inputs])
    return inputs
+def oks_overlaps(kpt_preds, kpt_gts, kpt_valids, kpt_areas, sigmas):
+    if not kpt_gts.astype('bool').any():
+        return kpt_preds.sum()*0
+    sigmas = paddle.to_tensor(sigmas, dtype=kpt_preds.dtype)
+    variances = (sigmas * 2)**2
+    assert kpt_preds.shape[0] == kpt_gts.shape[0]
+    kpt_preds = kpt_preds.reshape((-1, kpt_preds.shape[-1] // 2, 2))
+    kpt_gts = kpt_gts.reshape((-1, kpt_gts.shape[-1] // 2, 2))
+    squared_distance = (kpt_preds[:, :, 0] - kpt_gts[:, :, 0]) ** 2 + \
+        (kpt_preds[:, :, 1] - kpt_gts[:, :, 1]) ** 2
+    assert (kpt_valids.sum(-1) > 0).all()
+    squared_distance0 = squared_distance / (
+        kpt_areas[:, None] * variances[None, :] * 2)
+    squared_distance1 = paddle.exp(-squared_distance0)
+    squared_distance1 = squared_distance1 * kpt_valids
+    oks = squared_distance1.sum(axis=1) / kpt_valids.sum(axis=1)
+    return oks
+def oks_loss(pred,
+             target,
+             weight,
+             valid=None,
+             area=None,
+             linear=False,
+             sigmas=None,
+             eps=1e-6,
+             avg_factor=None, 
+             reduction=None):
+    """Oks loss.
+    Computing the oks loss between a set of predicted poses and target poses.
+    The loss is calculated as negative log of oks.
+    Args:
+        pred (Tensor): Predicted poses of format (x1, y1, x2, y2, ...),
+            shape (n, K*2).
+        target (Tensor): Corresponding gt poses, shape (n, K*2).
+        linear (bool, optional): If True, use linear scale of loss instead of
+            log scale. Default: False.
+        eps (float): Eps to avoid log(0).
+    Returns:
+        Tensor: Loss tensor.
+    """
+    oks = oks_overlaps(pred, target, valid, area, sigmas).clip(min=eps)
+    if linear:
+        loss = 1 - oks
+    else:
+        loss = -oks.log()
+    if weight is not None:
+        if weight.shape != loss.shape:
+            if weight.shape[0] == loss.shape[0]:
+                # For most cases, weight is of shape (num_priors, ),
+                #  which means it does not have the second axis num_class
+                weight = weight.reshape((-1, 1))
+            else:
+                # Sometimes, weight per anchor per class is also needed. e.g.
+                #  in FSAF. But it may be flattened of shape
+                #  (num_priors x num_class, ), while loss is still of shape
+                #  (num_priors, num_class).
+                assert weight.numel() == loss.numel()
+                weight = weight.reshape((loss.shape[0], -1))
+        assert weight.ndim == loss.ndim
+        loss = loss * weight
+    # if avg_factor is not specified, just reduce the loss
+    if avg_factor is None:
+        if reduction == 'mean':
+            loss = loss.mean()
+        elif reduction == 'sum':
+            loss = loss.sum()
+    else:
+        # if reduction is mean, then average the loss by avg_factor
+        if reduction == 'mean':
+            # Avoid causing ZeroDivisionError when avg_factor is 0.0,
+            # i.e., all labels of an image belong to ignore index.
+            eps = 1e-10
+            loss = loss.sum() / (avg_factor + eps)
+        # if reduction is 'none', then do nothing, otherwise raise an error
+        elif reduction != 'none':
+            raise ValueError('avg_factor can not be used with reduction="sum"')
+    return loss
+@register
+@serializable
+class OKSLoss(nn.Layer):
+    """OKSLoss.
+    Computing the oks loss between a set of predicted poses and target poses.
+    Args:
+        linear (bool): If True, use linear scale of loss instead of log scale.
+            Default: False.
+        eps (float): Eps to avoid log(0).
+        reduction (str): Options are "none", "mean" and "sum".
+        loss_weight (float): Weight of loss.
+    """
+    def __init__(self,
+                 linear=False,
+                 num_keypoints=17,
+                 eps=1e-6,
+                 reduction='mean',
+                 loss_weight=1.0):
+        super(OKSLoss, self).__init__()
+        self.linear = linear
+        self.eps = eps
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+        if num_keypoints == 17:
+            self.sigmas = np.array([
+                .26, .25, .25, .35, .35, .79, .79, .72, .72, .62, .62, 1.07,
+                1.07, .87, .87, .89, .89
+            ], dtype=np.float32) / 10.0
+        elif num_keypoints == 14:
+            self.sigmas = np.array([
+                .79, .79, .72, .72, .62, .62, 1.07, 1.07, .87, .87, .89, .89,
+                .79, .79
+            ]) / 10.0
+        else:
+            raise ValueError(f'Unsupported keypoints number {num_keypoints}')
+    def forward(self,
+                pred,
+                target,
+                valid,
+                area,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                **kwargs):
+        """Forward function.
+        Args:
+            pred (Tensor): The prediction.
+            target (Tensor): The learning target of the prediction.
+            valid (Tensor): The visible flag of the target pose.
+            area (Tensor): The area of the target pose.
+            weight (Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None. Options are "none", "mean" and "sum".
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if (weight is not None) and (not paddle.any(weight > 0)) and (
+                reduction != 'none'):
+            if pred.dim() == weight.dim() + 1:
+                weight = weight.unsqueeze(1)
+            return (pred * weight).sum()  # 0
+        if weight is not None and weight.dim() > 1:
+            # TODO: remove this in the future
+            # reduce the weight of shape (n, 4) to (n,) to match the
+            # iou_loss of shape (n,)
+            assert weight.shape == pred.shape
+            weight = weight.mean(-1)
+        loss = self.loss_weight * oks_loss(
+            pred,
+            target,
+            weight,
+            valid=valid,
+            area=area,
+            linear=self.linear,
+            sigmas=self.sigmas,
+            eps=self.eps,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            **kwargs)
+        return loss
+def center_focal_loss(pred, gt, weight=None, mask=None, avg_factor=None, reduction=None):
+    """Modified focal loss. Exactly the same as CornerNet.
+    Runs faster and costs a little bit more memory.
+    Args:
+        pred (Tensor): The prediction with shape [bs, c, h, w].
+        gt (Tensor): The learning target of the prediction in gaussian
+            distribution, with shape [bs, c, h, w].
+        mask (Tensor): The valid mask. Defaults to None.
+    """
+    if not gt.astype('bool').any():
+        return pred.sum()*0
+    pos_inds = gt.equal(1).astype('float32')
+    if mask is None:
+        neg_inds = gt.less_than(paddle.to_tensor([1], dtype='float32')).astype('float32')
+    else:
+        neg_inds = gt.less_than(paddle.to_tensor([1], dtype='float32')).astype('float32') * mask.equal(0).astype('float32')
+    neg_weights = paddle.pow(1 - gt, 4)
+    loss = 0
+    pos_loss = paddle.log(pred) * paddle.pow(1 - pred, 2) * pos_inds
+    neg_loss = paddle.log(1 - pred) * paddle.pow(pred, 2) * neg_weights * \
+        neg_inds
+    num_pos = pos_inds.astype('float32').sum()
+    pos_loss = pos_loss.sum()
+    neg_loss = neg_loss.sum()
+    if num_pos == 0:
+        loss = loss - neg_loss
+    else:
+        loss = loss - (pos_loss + neg_loss) / num_pos
+    if weight is not None:
+        if weight.shape != loss.shape:
+            if weight.shape[0] == loss.shape[0]:
+                # For most cases, weight is of shape (num_priors, ),
+                #  which means it does not have the second axis num_class
+                weight = weight.reshape((-1, 1))
+            else:
+                # Sometimes, weight per anchor per class is also needed. e.g.
+                #  in FSAF. But it may be flattened of shape
+                #  (num_priors x num_class, ), while loss is still of shape
+                #  (num_priors, num_class).
+                assert weight.numel() == loss.numel()
+                weight = weight.reshape((loss.shape[0], -1))
+        assert weight.ndim == loss.ndim
+        loss = loss * weight
+    # if avg_factor is not specified, just reduce the loss
+    if avg_factor is None:
+        if reduction == 'mean':
+            loss = loss.mean()
+        elif reduction == 'sum':
+            loss = loss.sum()
+    else:
+        # if reduction is mean, then average the loss by avg_factor
+        if reduction == 'mean':
+            # Avoid causing ZeroDivisionError when avg_factor is 0.0,
+            # i.e., all labels of an image belong to ignore index.
+            eps = 1e-10
+            loss = loss.sum() / (avg_factor + eps)
+        # if reduction is 'none', then do nothing, otherwise raise an error
+        elif reduction != 'none':
+            raise ValueError('avg_factor can not be used with reduction="sum"')
+    return loss
+@register
+@serializable
+class CenterFocalLoss(nn.Layer):
+    """CenterFocalLoss is a variant of focal loss.
+    More details can be found in the `paper
+    <https://arxiv.org/abs/1808.01244>`_
+    Args:
+        reduction (str): Options are "none", "mean" and "sum".
+        loss_weight (float): Loss weight of current loss.
+    """
+    def __init__(self,
+                 reduction='none',
+                 loss_weight=1.0):
+        super(CenterFocalLoss, self).__init__()
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                mask=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+        Args:
+            pred (Tensor): The prediction.
+            target (Tensor): The learning target of the prediction in gaussian
+                distribution.
+            weight (Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            mask (Tensor): The valid mask. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_reg = self.loss_weight * center_focal_loss(
+            pred,
+            target,
+            weight,
+            mask=mask,
+            reduction=reduction,
+            avg_factor=avg_factor)
+        return loss_reg
+def l1_loss(pred, target, weight=None, reduction='mean', avg_factor=None):
+    """L1 loss.
+    Args:
+        pred (Tensor): The prediction.
+        target (Tensor): The learning target of the prediction.
+    Returns:
+        Tensor: Calculated loss
+    """
+    if not target.astype('bool').any():
+        return pred.sum() * 0
+    assert pred.shape == target.shape
+    loss = paddle.abs(pred - target)
+    if weight is not None:
+        if weight.shape != loss.shape:
+            if weight.shape[0] == loss.shape[0]:
+                # For most cases, weight is of shape (num_priors, ),
+                #  which means it does not have the second axis num_class
+                weight = weight.reshape((-1, 1))
+            else:
+                # Sometimes, weight per anchor per class is also needed. e.g.
+                #  in FSAF. But it may be flattened of shape
+                #  (num_priors x num_class, ), while loss is still of shape
+                #  (num_priors, num_class).
+                assert weight.numel() == loss.numel()
+                weight = weight.reshape((loss.shape[0], -1))
+        assert weight.ndim == loss.ndim
+        loss = loss * weight
+    # if avg_factor is not specified, just reduce the loss
+    if avg_factor is None:
+        if reduction == 'mean':
+            loss = loss.mean()
+        elif reduction == 'sum':
+            loss = loss.sum()
+    else:
+        # if reduction is mean, then average the loss by avg_factor
+        if reduction == 'mean':
+            # Avoid causing ZeroDivisionError when avg_factor is 0.0,
+            # i.e., all labels of an image belong to ignore index.
+            eps = 1e-10
+            loss = loss.sum() / (avg_factor + eps)
+        # if reduction is 'none', then do nothing, otherwise raise an error
+        elif reduction != 'none':
+            raise ValueError('avg_factor can not be used with reduction="sum"')
+    return loss
+@register
+@serializable
+class L1Loss(nn.Layer):
+    """L1 loss.
+    Args:
+        reduction (str, optional): The method to reduce the loss.
+            Options are "none", "mean" and "sum".
+        loss_weight (float, optional): The weight of loss.
+    """
+    def __init__(self, reduction='mean', loss_weight=1.0):
+        super(L1Loss, self).__init__()
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None):
+        """Forward function.
+        Args:
+            pred (Tensor): The prediction.
+            target (Tensor): The learning target of the prediction.
+            weight (Tensor, optional): The weight of loss for each
+                prediction. Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Defaults to None.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        loss_bbox = self.loss_weight * l1_loss(
+            pred, target, weight, reduction=reduction, avg_factor=avg_factor)
+        return loss_bbox
--- a/ppdet/modeling/necks/__init__.py
+++ b/ppdet/modeling/necks/__init__.py
@@ -36,3 +36,4 @@ from .es_pan import *
 from .lc_pan import *
 from .custom_pan import *
 from .dilated_encoder import *
+from .channel_mapper import *
--- a/ppdet/modeling/necks/channel_mapper.py
+++ b/ppdet/modeling/necks/channel_mapper.py
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. 
+#   
+# Licensed under the Apache License, Version 2.0 (the "License");   
+# you may not use this file except in compliance with the License.  
+# You may obtain a copy of the License at   
+#   
+#     http://www.apache.org/licenses/LICENSE-2.0    
+#   
+# Unless required by applicable law or agreed to in writing, software   
+# distributed under the License is distributed on an "AS IS" BASIS, 
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.  
+# See the License for the specific language governing permissions and   
+# limitations under the License.
+"""
+this code is base on mmdet: git@github.com:open-mmlab/mmdetection.git
+"""
+import paddle.nn as nn
+from ppdet.core.workspace import register, serializable
+from ..backbones.hrnet import ConvNormLayer
+from ..shape_spec import ShapeSpec
+from ..initializer import xavier_uniform_, constant_
+__all__ = ['ChannelMapper']
+@register
+@serializable
+class ChannelMapper(nn.Layer):
+    """Channel Mapper to reduce/increase channels of backbone features.
+    This is used to reduce/increase channels of backbone features.
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale).
+        kernel_size (int, optional): kernel_size for reducing channels (used
+            at each scale). Default: 3.
+        conv_cfg (dict, optional): Config dict for convolution layer.
+            Default: None.
+        norm_cfg (dict, optional): Config dict for normalization layer.
+            Default: None.
+        act_cfg (dict, optional): Config dict for activation layer in
+            ConvModule. Default: dict(type='ReLU').
+        num_outs (int, optional): Number of output feature maps. There
+            would be extra_convs when num_outs larger than the length
+            of in_channels.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
+    """
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 norm_type="gn",
+                 norm_groups=32,
+                 act='relu',
+                 num_outs=None,
+                 init_cfg=dict(
+                     type='Xavier', layer='Conv2d', distribution='uniform')):
+        super(ChannelMapper, self).__init__()
+        assert isinstance(in_channels, list)
+        self.extra_convs = None
+        if num_outs is None:
+            num_outs = len(in_channels)
+        self.convs = nn.LayerList()
+        for in_channel in in_channels:
+            self.convs.append(
+                ConvNormLayer(
+                    ch_in=in_channel,
+                    ch_out=out_channels,
+                    filter_size=kernel_size,
+                    norm_type='gn',
+                    norm_groups=32,
+                    act=act))
+        if num_outs > len(in_channels):
+            self.extra_convs = nn.LayerList()
+            for i in range(len(in_channels), num_outs):
+                if i == len(in_channels):
+                    in_channel = in_channels[-1]
+                else:
+                    in_channel = out_channels
+                self.extra_convs.append(
+                    ConvNormLayer(
+                        ch_in=in_channel,
+                        ch_out=out_channels,
+                        filter_size=3,
+                        stride=2,
+                        norm_type='gn',
+                        norm_groups=32,
+                        act=act))
+        self.init_weights()
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.convs)
+        outs = [self.convs[i](inputs[i]) for i in range(len(inputs))]
+        if self.extra_convs:
+            for i in range(len(self.extra_convs)):
+                if i == 0:
+                    outs.append(self.extra_convs[0](inputs[-1]))
+                else:
+                    outs.append(self.extra_convs[i](outs[-1]))
+        return tuple(outs)
+    @property
+    def out_shape(self):
+        return [
+            ShapeSpec(
+                channels=self.out_channel, stride=1. / s)
+            for s in self.spatial_scales
+        ]
+    def init_weights(self):
+        """Initialize the transformer weights."""
+        for p in self.parameters():
+            if p.rank() > 1:
+                xavier_uniform_(p)
+                if hasattr(p, 'bias') and p.bias is not None:
+                    constant_(p.bais)
--- a/ppdet/modeling/transformers/__init__.py
+++ b/ppdet/modeling/transformers/__init__.py
@@ -25,3 +25,4 @@ from .matchers import *
 from .position_encoding import *
 from .deformable_transformer import *
 from .dino_transformer import *
+from .petr_transformer import *
--- a/ppdet/modeling/transformers/petr_transformer.py
+++ b/ppdet/modeling/transformers/petr_transformer.py
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+this code is base on https://github.com/hikvision-research/opera/blob/main/opera/models/utils/transformer.py
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import math
+import numpy as np
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+from ppdet.core.workspace import register
+from ..layers import MultiHeadAttention, _convert_attention_mask
+from .utils import _get_clones
+from ..initializer import linear_init_, normal_, constant_, xavier_uniform_
+__all__ = [
+    'PETRTransformer', 'MultiScaleDeformablePoseAttention',
+    'PETR_TransformerDecoderLayer', 'PETR_TransformerDecoder',
+    'PETR_DeformableDetrTransformerDecoder',
+    'PETR_DeformableTransformerDecoder', 'TransformerEncoderLayer',
+    'TransformerEncoder', 'MSDeformableAttention'
+]
+def masked_fill(x, mask, value):
+    y = paddle.full(x.shape, value, x.dtype)
+    return paddle.where(mask, y, x)
+def inverse_sigmoid(x, eps=1e-5):
+    """Inverse function of sigmoid.
+    Args:
+        x (Tensor): The tensor to do the
+            inverse.
+        eps (float): EPS avoid numerical
+            overflow. Defaults 1e-5.
+    Returns:
+        Tensor: The x has passed the inverse
+            function of sigmoid, has same
+            shape with input.
+    """
+    x = x.clip(min=0, max=1)
+    x1 = x.clip(min=eps)
+    x2 = (1 - x).clip(min=eps)
+    return paddle.log(x1 / x2)
+@register
+class TransformerEncoderLayer(nn.Layer):
+    __inject__ = ['attn']
+    def __init__(self,
+                 d_model,
+                 attn=None,
+                 nhead=8,
+                 dim_feedforward=2048,
+                 dropout=0.1,
+                 activation="relu",
+                 attn_dropout=None,
+                 act_dropout=None,
+                 normalize_before=False):
+        super(TransformerEncoderLayer, self).__init__()
+        attn_dropout = dropout if attn_dropout is None else attn_dropout
+        act_dropout = dropout if act_dropout is None else act_dropout
+        self.normalize_before = normalize_before
+        self.embed_dims = d_model
+        if attn is None:
+            self.self_attn = MultiHeadAttention(d_model, nhead, attn_dropout)
+        else:
+            self.self_attn = attn
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(act_dropout, mode="upscale_in_train")
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout, mode="upscale_in_train")
+        self.dropout2 = nn.Dropout(dropout, mode="upscale_in_train")
+        self.activation = getattr(F, activation)
+        self._reset_parameters()
+    def _reset_parameters(self):
+        linear_init_(self.linear1)
+        linear_init_(self.linear2)
+    @staticmethod
+    def with_pos_embed(tensor, pos_embed):
+        return tensor if pos_embed is None else tensor + pos_embed
+    def forward(self, src, src_mask=None, pos_embed=None, **kwargs):
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        q = k = self.with_pos_embed(src, pos_embed)
+        src = self.self_attn(q, k, value=src, attn_mask=src_mask, **kwargs)
+        src = residual + self.dropout1(src)
+        if not self.normalize_before:
+            src = self.norm1(src)
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+@register
+class TransformerEncoder(nn.Layer):
+    __inject__ = ['encoder_layer']
+    def __init__(self, encoder_layer, num_layers, norm=None):
+        super(TransformerEncoder, self).__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+        self.embed_dims = encoder_layer.embed_dims
+    def forward(self, src, src_mask=None, pos_embed=None, **kwargs):
+        output = src
+        for layer in self.layers:
+            output = layer(
+                output, src_mask=src_mask, pos_embed=pos_embed, **kwargs)
+        if self.norm is not None:
+            output = self.norm(output)
+        return output
+@register
+class MSDeformableAttention(nn.Layer):
+    def __init__(self,
+                 embed_dim=256,
+                 num_heads=8,
+                 num_levels=4,
+                 num_points=4,
+                 lr_mult=0.1):
+        """
+        Multi-Scale Deformable Attention Module
+        """
+        super(MSDeformableAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.num_levels = num_levels
+        self.num_points = num_points
+        self.total_points = num_heads * num_levels * num_points
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
+        self.sampling_offsets = nn.Linear(
+            embed_dim,
+            self.total_points * 2,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
+        self.attention_weights = nn.Linear(embed_dim, self.total_points)
+        self.value_proj = nn.Linear(embed_dim, embed_dim)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+        try:
+            # use cuda op
+            print("use deformable_detr_ops in ms_deformable_attn")
+            from deformable_detr_ops import ms_deformable_attn
+        except:
+            # use paddle func
+            from .utils import deformable_attention_core_func as ms_deformable_attn
+        self.ms_deformable_attn_core = ms_deformable_attn
+        self._reset_parameters()
+    def _reset_parameters(self):
+        # sampling_offsets
+        constant_(self.sampling_offsets.weight)
+        thetas = paddle.arange(
+            self.num_heads,
+            dtype=paddle.float32) * (2.0 * math.pi / self.num_heads)
+        grid_init = paddle.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = grid_init / grid_init.abs().max(-1, keepdim=True)
+        grid_init = grid_init.reshape([self.num_heads, 1, 1, 2]).tile(
+            [1, self.num_levels, self.num_points, 1])
+        scaling = paddle.arange(
+            1, self.num_points + 1,
+            dtype=paddle.float32).reshape([1, 1, -1, 1])
+        grid_init *= scaling
+        self.sampling_offsets.bias.set_value(grid_init.flatten())
+        # attention_weights
+        constant_(self.attention_weights.weight)
+        constant_(self.attention_weights.bias)
+        # proj
+        xavier_uniform_(self.value_proj.weight)
+        constant_(self.value_proj.bias)
+        xavier_uniform_(self.output_proj.weight)
+        constant_(self.output_proj.bias)
+    def forward(self,
+                query,
+                key,
+                value,
+                reference_points,
+                value_spatial_shapes,
+                value_level_start_index,
+                attn_mask=None,
+                **kwargs):
+        """
+        Args:
+            query (Tensor): [bs, query_length, C]
+            reference_points (Tensor): [bs, query_length, n_levels, 2], range in [0, 1], top-left (0,0),
+                bottom-right (1, 1), including padding area
+            value (Tensor): [bs, value_length, C]
+            value_spatial_shapes (Tensor): [n_levels, 2], [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
+            value_level_start_index (Tensor(int64)): [n_levels], [0, H_0*W_0, H_0*W_0+H_1*W_1, ...]
+            attn_mask (Tensor): [bs, value_length], True for non-padding elements, False for padding elements
+        Returns:
+            output (Tensor): [bs, Length_{query}, C]
+        """
+        bs, Len_q = query.shape[:2]
+        Len_v = value.shape[1]
+        assert int(value_spatial_shapes.prod(1).sum()) == Len_v
+        value = self.value_proj(value)
+        if attn_mask is not None:
+            attn_mask = attn_mask.astype(value.dtype).unsqueeze(-1)
+            value *= attn_mask
+        value = value.reshape([bs, Len_v, self.num_heads, self.head_dim])
+        sampling_offsets = self.sampling_offsets(query).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels, self.num_points, 2])
+        attention_weights = self.attention_weights(query).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels * self.num_points])
+        attention_weights = F.softmax(attention_weights).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels, self.num_points])
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = value_spatial_shapes.flip([1]).reshape(
+                [1, 1, 1, self.num_levels, 1, 2])
+            sampling_locations = reference_points.reshape([
+                bs, Len_q, 1, self.num_levels, 1, 2
+            ]) + sampling_offsets / offset_normalizer
+        elif reference_points.shape[-1] == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2] + sampling_offsets /
+                self.num_points * reference_points[:, :, None, :, None, 2:] *
+                0.5)
+        else:
+            raise ValueError(
+                "Last dim of reference_points must be 2 or 4, but get {} instead.".
+                format(reference_points.shape[-1]))
+        output = self.ms_deformable_attn_core(
+            value, value_spatial_shapes, value_level_start_index,
+            sampling_locations, attention_weights)
+        output = self.output_proj(output)
+        return output
+@register
+class MultiScaleDeformablePoseAttention(nn.Layer):
+    """An attention module used in PETR. `End-to-End Multi-Person
+    Pose Estimation with Transformers`.
+    Args:
+        embed_dims (int): The embedding dimension of Attention.
+            Default: 256.
+        num_heads (int): Parallel attention heads. Default: 8.
+        num_levels (int): The number of feature map used in
+            Attention. Default: 4.
+        num_points (int): The number of sampling points for
+            each query in each head. Default: 17.
+        im2col_step (int): The step used in image_to_column.
+            Default: 64.
+        dropout (float): A Dropout layer on `inp_residual`.
+            Default: 0.1.
+        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
+            Default: None.
+    """
+    def __init__(self,
+                 embed_dims=256,
+                 num_heads=8,
+                 num_levels=4,
+                 num_points=17,
+                 im2col_step=64,
+                 dropout=0.1,
+                 norm_cfg=None,
+                 init_cfg=None,
+                 batch_first=False,
+                 lr_mult=0.1):
+        super().__init__()
+        if embed_dims % num_heads != 0:
+            raise ValueError(f'embed_dims must be divisible by num_heads, '
+                             f'but got {embed_dims} and {num_heads}')
+        dim_per_head = embed_dims // num_heads
+        self.norm_cfg = norm_cfg
+        self.init_cfg = init_cfg
+        self.dropout = nn.Dropout(dropout)
+        self.batch_first = batch_first
+        # you'd better set dim_per_head to a power of 2
+        # which is more efficient in the CUDA implementation
+        def _is_power_of_2(n):
+            if (not isinstance(n, int)) or (n < 0):
+                raise ValueError(
+                    'invalid input for _is_power_of_2: {} (type: {})'.format(
+                        n, type(n)))
+            return (n & (n - 1) == 0) and n != 0
+        if not _is_power_of_2(dim_per_head):
+            warnings.warn("You'd better set embed_dims in "
+                          'MultiScaleDeformAttention to make '
+                          'the dimension of each attention head a power of 2 '
+                          'which is more efficient in our CUDA implementation.')
+        self.im2col_step = im2col_step
+        self.embed_dims = embed_dims
+        self.num_levels = num_levels
+        self.num_heads = num_heads
+        self.num_points = num_points
+        self.sampling_offsets = nn.Linear(
+            embed_dims,
+            num_heads * num_levels * num_points * 2,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
+        self.attention_weights = nn.Linear(embed_dims,
+                                           num_heads * num_levels * num_points)
+        self.value_proj = nn.Linear(embed_dims, embed_dims)
+        self.output_proj = nn.Linear(embed_dims, embed_dims)
+        try:
+            # use cuda op
+            from deformable_detr_ops import ms_deformable_attn
+        except:
+            # use paddle func
+            from .utils import deformable_attention_core_func as ms_deformable_attn
+        self.ms_deformable_attn_core = ms_deformable_attn
+        self.init_weights()
+    def init_weights(self):
+        """Default initialization for Parameters of Module."""
+        constant_(self.sampling_offsets.weight)
+        constant_(self.sampling_offsets.bias)
+        constant_(self.attention_weights.weight)
+        constant_(self.attention_weights.bias)
+        xavier_uniform_(self.value_proj.weight)
+        constant_(self.value_proj.bias)
+        xavier_uniform_(self.output_proj.weight)
+        constant_(self.output_proj.bias)
+    def forward(self,
+                query,
+                key,
+                value,
+                residual=None,
+                attn_mask=None,
+                reference_points=None,
+                value_spatial_shapes=None,
+                value_level_start_index=None,
+                **kwargs):
+        """Forward Function of MultiScaleDeformAttention.
+        Args:
+            query (Tensor): Query of Transformer with shape
+                (num_query, bs, embed_dims).
+            key (Tensor): The key tensor with shape (num_key, bs, embed_dims).
+            value (Tensor): The value tensor with shape
+                (num_key, bs, embed_dims).
+            residual (Tensor): The tensor used for addition, with the
+                same shape as `x`. Default None. If None, `x` will be used.
+            reference_points (Tensor):  The normalized reference points with
+                shape (bs, num_query, num_levels, K*2), all elements is range
+                in [0, 1], top-left (0,0), bottom-right (1, 1), including
+                padding area.
+            attn_mask (Tensor): ByteTensor for `query`, with
+                shape [bs, num_key].
+            value_spatial_shapes (Tensor): Spatial shape of features in
+                different level. With shape  (num_levels, 2),
+                last dimension represent (h, w).
+            value_level_start_index (Tensor): The start index of each level.
+                A tensor has shape (num_levels) and can be represented
+                as [0, h_0*w_0, h_0*w_0+h_1*w_1, ...].
+        Returns:
+            Tensor: forwarded results with shape [num_query, bs, embed_dims].
+        """
+        if key is None:
+            key = query
+        if value is None:
+            value = key
+        bs, num_query, _ = query.shape
+        bs, num_key, _ = value.shape
+        assert (value_spatial_shapes[:, 0].numpy() *
+                value_spatial_shapes[:, 1].numpy()).sum() == num_key
+        value = self.value_proj(value)
+        if attn_mask is not None:
+            # value = value.masked_fill(attn_mask[..., None], 0.0)
+            value *= attn_mask.unsqueeze(-1)
+        value = value.reshape([bs, num_key, self.num_heads, -1])
+        sampling_offsets = self.sampling_offsets(query).reshape([
+            bs, num_query, self.num_heads, self.num_levels, self.num_points, 2
+        ])
+        attention_weights = self.attention_weights(query).reshape(
+            [bs, num_query, self.num_heads, self.num_levels * self.num_points])
+        attention_weights = F.softmax(attention_weights, axis=-1)
+        attention_weights = attention_weights.reshape(
+            [bs, num_query, self.num_heads, self.num_levels, self.num_points])
+        if reference_points.shape[-1] == self.num_points * 2:
+            reference_points_reshape = reference_points.reshape(
+                (bs, num_query, self.num_levels, -1, 2)).unsqueeze(2)
+            x1 = reference_points[:, :, :, 0::2].min(axis=-1, keepdim=True)
+            y1 = reference_points[:, :, :, 1::2].min(axis=-1, keepdim=True)
+            x2 = reference_points[:, :, :, 0::2].max(axis=-1, keepdim=True)
+            y2 = reference_points[:, :, :, 1::2].max(axis=-1, keepdim=True)
+            w = paddle.clip(x2 - x1, min=1e-4)
+            h = paddle.clip(y2 - y1, min=1e-4)
+            wh = paddle.concat([w, h], axis=-1)[:, :, None, :, None, :]
+            sampling_locations = reference_points_reshape \
+                                 + sampling_offsets * wh * 0.5
+        else:
+            raise ValueError(
+                f'Last dim of reference_points must be'
+                f' 2K, but get {reference_points.shape[-1]} instead.')
+        output = self.ms_deformable_attn_core(
+            value, value_spatial_shapes, value_level_start_index,
+            sampling_locations, attention_weights)
+        output = self.output_proj(output)
+        return output
+@register
+class PETR_TransformerDecoderLayer(nn.Layer):
+    __inject__ = ['self_attn', 'cross_attn']
+    def __init__(self,
+                 d_model,
+                 nhead=8,
+                 self_attn=None,
+                 cross_attn=None,
+                 dim_feedforward=2048,
+                 dropout=0.1,
+                 activation="relu",
+                 attn_dropout=None,
+                 act_dropout=None,
+                 normalize_before=False):
+        super(PETR_TransformerDecoderLayer, self).__init__()
+        attn_dropout = dropout if attn_dropout is None else attn_dropout
+        act_dropout = dropout if act_dropout is None else act_dropout
+        self.normalize_before = normalize_before
+        if self_attn is None:
+            self.self_attn = MultiHeadAttention(d_model, nhead, attn_dropout)
+        else:
+            self.self_attn = self_attn
+        if cross_attn is None:
+            self.cross_attn = MultiHeadAttention(d_model, nhead, attn_dropout)
+        else:
+            self.cross_attn = cross_attn
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(act_dropout, mode="upscale_in_train")
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout, mode="upscale_in_train")
+        self.dropout2 = nn.Dropout(dropout, mode="upscale_in_train")
+        self.dropout3 = nn.Dropout(dropout, mode="upscale_in_train")
+        self.activation = getattr(F, activation)
+        self._reset_parameters()
+    def _reset_parameters(self):
+        linear_init_(self.linear1)
+        linear_init_(self.linear2)
+    @staticmethod
+    def with_pos_embed(tensor, pos_embed):
+        return tensor if pos_embed is None else tensor + pos_embed
+    def forward(self,
+                tgt,
+                memory,
+                tgt_mask=None,
+                memory_mask=None,
+                pos_embed=None,
+                query_pos_embed=None,
+                **kwargs):
+        tgt_mask = _convert_attention_mask(tgt_mask, tgt.dtype)
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        q = k = self.with_pos_embed(tgt, query_pos_embed)
+        tgt = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask)
+        tgt = residual + self.dropout1(tgt)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        q = self.with_pos_embed(tgt, query_pos_embed)
+        key_tmp = tgt
+        # k = self.with_pos_embed(memory, pos_embed)
+        tgt = self.cross_attn(
+            q, key=key_tmp, value=memory, attn_mask=memory_mask, **kwargs)
+        tgt = residual + self.dropout2(tgt)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+@register
+class PETR_TransformerDecoder(nn.Layer):
+    """Implements the decoder in PETR transformer.
+    Args:
+        return_intermediate (bool): Whether to return intermediate outputs.
+        coder_norm_cfg (dict): Config of last normalization layer. Default：
+            `LN`.
+    """
+    __inject__ = ['decoder_layer']
+    def __init__(self,
+                 decoder_layer,
+                 num_layers,
+                 norm=None,
+                 return_intermediate=False,
+                 num_keypoints=17,
+                 **kwargs):
+        super(PETR_TransformerDecoder, self).__init__()
+        self.layers = _get_clones(decoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+        self.return_intermediate = return_intermediate
+        self.num_keypoints = num_keypoints
+    def forward(self,
+                query,
+                *args,
+                reference_points=None,
+                valid_ratios=None,
+                kpt_branches=None,
+                **kwargs):
+        """Forward function for `TransformerDecoder`.
+        Args:
+            query (Tensor): Input query with shape (num_query, bs, embed_dims).
+            reference_points (Tensor): The reference points of offset,
+                has shape (bs, num_query, K*2).
+            valid_ratios (Tensor): The radios of valid points on the feature
+                map, has shape (bs, num_levels, 2).
+            kpt_branches: (obj:`nn.LayerList`): Used for refining the
+                regression results. Only would be passed when `with_box_refine`
+                is True, otherwise would be passed a `None`.
+        Returns:
+            tuple (Tensor): Results with shape [1, num_query, bs, embed_dims] when
+                return_intermediate is `False`, otherwise it has shape
+                [num_layers, num_query, bs, embed_dims] and
+                [num_layers, bs, num_query, K*2].
+        """
+        output = query
+        intermediate = []
+        intermediate_reference_points = []
+        for lid, layer in enumerate(self.layers):
+            if reference_points.shape[-1] == self.num_keypoints * 2:
+                reference_points_input = \
+                    reference_points[:, :, None] * \
+                    valid_ratios.tile((1, 1, self.num_keypoints))[:, None]
+            else:
+                assert reference_points.shape[-1] == 2
+                reference_points_input = reference_points[:, :, None] * \
+                                         valid_ratios[:, None]
+            output = layer(
+                output,
+                *args,
+                reference_points=reference_points_input,
+                **kwargs)
+            if kpt_branches is not None:
+                tmp = kpt_branches[lid](output)
+                if reference_points.shape[-1] == self.num_keypoints * 2:
+                    new_reference_points = tmp + inverse_sigmoid(
+                        reference_points)
+                    new_reference_points = F.sigmoid(new_reference_points)
+                else:
+                    raise NotImplementedError
+                reference_points = new_reference_points.detach()
+            if self.return_intermediate:
+                intermediate.append(output)
+                intermediate_reference_points.append(reference_points)
+        if self.return_intermediate:
+            return paddle.stack(intermediate), paddle.stack(
+                intermediate_reference_points)
+        return output, reference_points
+@register
+class PETR_DeformableTransformerDecoder(nn.Layer):
+    __inject__ = ['decoder_layer']
+    def __init__(self, decoder_layer, num_layers, return_intermediate=False):
+        super(PETR_DeformableTransformerDecoder, self).__init__()
+        self.layers = _get_clones(decoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.return_intermediate = return_intermediate
+    def forward(self,
+                tgt,
+                reference_points,
+                memory,
+                memory_spatial_shapes,
+                memory_mask=None,
+                query_pos_embed=None):
+        output = tgt
+        intermediate = []
+        for lid, layer in enumerate(self.layers):
+            output = layer(output, reference_points, memory,
+                           memory_spatial_shapes, memory_mask, query_pos_embed)
+            if self.return_intermediate:
+                intermediate.append(output)
+        if self.return_intermediate:
+            return paddle.stack(intermediate)
+        return output.unsqueeze(0)
+@register
+class PETR_DeformableDetrTransformerDecoder(PETR_DeformableTransformerDecoder):
+    """Implements the decoder in DETR transformer.
+    Args:
+        return_intermediate (bool): Whether to return intermediate outputs.
+        coder_norm_cfg (dict): Config of last normalization layer. Default：
+            `LN`.
+    """
+    def __init__(self, *args, return_intermediate=False, **kwargs):
+        super(PETR_DeformableDetrTransformerDecoder, self).__init__(*args,
+                                                                    **kwargs)
+        self.return_intermediate = return_intermediate
+    def forward(self,
+                query,
+                *args,
+                reference_points=None,
+                valid_ratios=None,
+                reg_branches=None,
+                **kwargs):
+        """Forward function for `TransformerDecoder`.
+        Args:
+            query (Tensor): Input query with shape
+                `(num_query, bs, embed_dims)`.
+            reference_points (Tensor): The reference
+                points of offset. has shape
+                (bs, num_query, 4) when as_two_stage,
+                otherwise has shape ((bs, num_query, 2).
+            valid_ratios (Tensor): The radios of valid
+                points on the feature map, has shape
+                (bs, num_levels, 2)
+            reg_branch: (obj:`nn.LayerList`): Used for
+                refining the regression results. Only would
+                be passed when with_box_refine is True,
+                otherwise would be passed a `None`.
+        Returns:
+            Tensor: Results with shape [1, num_query, bs, embed_dims] when
+                return_intermediate is `False`, otherwise it has shape
+                [num_layers, num_query, bs, embed_dims].
+        """
+        output = query
+        intermediate = []
+        intermediate_reference_points = []
+        for lid, layer in enumerate(self.layers):
+            if reference_points.shape[-1] == 4:
+                reference_points_input = reference_points[:, :, None] * \
+                    paddle.concat([valid_ratios, valid_ratios], -1)[:, None]
+            else:
+                assert reference_points.shape[-1] == 2
+                reference_points_input = reference_points[:, :, None] * \
+                    valid_ratios[:, None]
+            output = layer(
+                output,
+                *args,
+                reference_points=reference_points_input,
+                **kwargs)
+            if reg_branches is not None:
+                tmp = reg_branches[lid](output)
+                if reference_points.shape[-1] == 4:
+                    new_reference_points = tmp + inverse_sigmoid(
+                        reference_points)
+                    new_reference_points = F.sigmoid(new_reference_points)
+                else:
+                    assert reference_points.shape[-1] == 2
+                    new_reference_points = tmp
+                    new_reference_points[..., :2] = tmp[
+                        ..., :2] + inverse_sigmoid(reference_points)
+                    new_reference_points = F.sigmoid(new_reference_points)
+                reference_points = new_reference_points.detach()
+            if self.return_intermediate:
+                intermediate.append(output)
+                intermediate_reference_points.append(reference_points)
+        if self.return_intermediate:
+            return paddle.stack(intermediate), paddle.stack(
+                intermediate_reference_points)
+        return output, reference_points
+@register
+class PETRTransformer(nn.Layer):
+    """Implements the PETR transformer.
+    Args:
+        as_two_stage (bool): Generate query from encoder features.
+            Default: False.
+        num_feature_levels (int): Number of feature maps from FPN:
+            Default: 4.
+        two_stage_num_proposals (int): Number of proposals when set
+            `as_two_stage` as True. Default: 300.
+    """
+    __inject__ = ["encoder", "decoder", "hm_encoder", "refine_decoder"]
+    def __init__(self,
+                 encoder="",
+                 decoder="",
+                 hm_encoder="",
+                 refine_decoder="",
+                 as_two_stage=True,
+                 num_feature_levels=4,
+                 two_stage_num_proposals=300,
+                 num_keypoints=17,
+                 **kwargs):
+        super(PETRTransformer, self).__init__(**kwargs)
+        self.as_two_stage = as_two_stage
+        self.num_feature_levels = num_feature_levels
+        self.two_stage_num_proposals = two_stage_num_proposals
+        self.num_keypoints = num_keypoints
+        self.encoder = encoder
+        self.decoder = decoder
+        self.embed_dims = self.encoder.embed_dims
+        self.hm_encoder = hm_encoder
+        self.refine_decoder = refine_decoder
+        self.init_layers()
+        self.init_weights()
+    def init_layers(self):
+        """Initialize layers of the DeformableDetrTransformer."""
+        #paddle.create_parameter
+        self.level_embeds = paddle.create_parameter(
+            (self.num_feature_levels, self.embed_dims), dtype="float32")
+        if self.as_two_stage:
+            self.enc_output = nn.Linear(self.embed_dims, self.embed_dims)
+            self.enc_output_norm = nn.LayerNorm(self.embed_dims)
+            self.refine_query_embedding = nn.Embedding(self.num_keypoints,
+                                                       self.embed_dims * 2)
+        else:
+            self.reference_points = nn.Linear(self.embed_dims,
+                                              2 * self.num_keypoints)
+    def init_weights(self):
+        """Initialize the transformer weights."""
+        for p in self.parameters():
+            if p.rank() > 1:
+                xavier_uniform_(p)
+                if hasattr(p, 'bias') and p.bias is not None:
+                    constant_(p.bais)
+        for m in self.sublayers():
+            if isinstance(m, MSDeformableAttention):
+                m._reset_parameters()
+        for m in self.sublayers():
+            if isinstance(m, MultiScaleDeformablePoseAttention):
+                m.init_weights()
+        if not self.as_two_stage:
+            xavier_uniform_(self.reference_points.weight)
+            constant_(self.reference_points.bias)
+        normal_(self.level_embeds)
+        normal_(self.refine_query_embedding.weight)
+    def gen_encoder_output_proposals(self, memory, memory_padding_mask,
+                                     spatial_shapes):
+        """Generate proposals from encoded memory.
+        Args:
+            memory (Tensor): The output of encoder, has shape
+                (bs, num_key, embed_dim). num_key is equal the number of points
+                on feature map from all level.
+            memory_padding_mask (Tensor): Padding mask for memory.
+                has shape (bs, num_key).
+            spatial_shapes (Tensor): The shape of all feature maps.
+                has shape (num_level, 2).
+        Returns:
+            tuple: A tuple of feature map and bbox prediction.
+                - output_memory (Tensor): The input of decoder, has shape
+                    (bs, num_key, embed_dim). num_key is equal the number of
+                    points on feature map from all levels.
+                - output_proposals (Tensor): The normalized proposal
+                    after a inverse sigmoid, has shape (bs, num_keys, 4).
+        """
+        N, S, C = memory.shape
+        proposals = []
+        _cur = 0
+        for lvl, (H, W) in enumerate(spatial_shapes):
+            mask_flatten_ = memory_padding_mask[:, _cur:(_cur + H * W)].reshape(
+                [N, H, W, 1])
+            valid_H = paddle.sum(mask_flatten_[:, :, 0, 0], 1)
+            valid_W = paddle.sum(mask_flatten_[:, 0, :, 0], 1)
+            grid_y, grid_x = paddle.meshgrid(
+                paddle.linspace(
+                    0, H - 1, H, dtype="float32"),
+                paddle.linspace(
+                    0, W - 1, W, dtype="float32"))
+            grid = paddle.concat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)],
+                                 -1)
+            scale = paddle.concat(
+                [valid_W.unsqueeze(-1),
+                 valid_H.unsqueeze(-1)], 1).reshape([N, 1, 1, 2])
+            grid = (grid.unsqueeze(0).expand((N, -1, -1, -1)) + 0.5) / scale
+            proposal = grid.reshape([N, -1, 2])
+            proposals.append(proposal)
+            _cur += (H * W)
+        output_proposals = paddle.concat(proposals, 1)
+        output_proposals_valid = ((output_proposals > 0.01) &
+                                  (output_proposals < 0.99)).all(
+                                      -1, keepdim=True).astype("bool")
+        output_proposals = paddle.log(output_proposals / (1 - output_proposals))
+        output_proposals = masked_fill(
+            output_proposals, ~memory_padding_mask.astype("bool").unsqueeze(-1),
+            float('inf'))
+        output_proposals = masked_fill(output_proposals,
+                                       ~output_proposals_valid, float('inf'))
+        output_memory = memory
+        output_memory = masked_fill(
+            output_memory, ~memory_padding_mask.astype("bool").unsqueeze(-1),
+            float(0))
+        output_memory = masked_fill(output_memory, ~output_proposals_valid,
+                                    float(0))
+        output_memory = self.enc_output_norm(self.enc_output(output_memory))
+        return output_memory, output_proposals
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios):
+        """Get the reference points used in decoder.
+        Args:
+            spatial_shapes (Tensor): The shape of all feature maps,
+                has shape (num_level, 2).
+            valid_ratios (Tensor): The radios of valid points on the
+                feature map, has shape (bs, num_levels, 2).
+        Returns:
+            Tensor: reference points used in decoder, has \
+                shape (bs, num_keys, num_levels, 2).
+        """
+        reference_points_list = []
+        for lvl, (H, W) in enumerate(spatial_shapes):
+            ref_y, ref_x = paddle.meshgrid(
+                paddle.linspace(
+                    0.5, H - 0.5, H, dtype="float32"),
+                paddle.linspace(
+                    0.5, W - 0.5, W, dtype="float32"))
+            ref_y = ref_y.reshape(
+                (-1, ))[None] / (valid_ratios[:, None, lvl, 1] * H)
+            ref_x = ref_x.reshape(
+                (-1, ))[None] / (valid_ratios[:, None, lvl, 0] * W)
+            ref = paddle.stack((ref_x, ref_y), -1)
+            reference_points_list.append(ref)
+        reference_points = paddle.concat(reference_points_list, 1)
+        reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+        return reference_points
+    def get_valid_ratio(self, mask):
+        """Get the valid radios of feature maps of all level."""
+        _, H, W = mask.shape
+        valid_H = paddle.sum(mask[:, :, 0].astype('float'), 1)
+        valid_W = paddle.sum(mask[:, 0, :].astype('float'), 1)
+        valid_ratio_h = valid_H.astype('float') / H
+        valid_ratio_w = valid_W.astype('float') / W
+        valid_ratio = paddle.stack([valid_ratio_w, valid_ratio_h], -1)
+        return valid_ratio
+    def get_proposal_pos_embed(self,
+                               proposals,
+                               num_pos_feats=128,
+                               temperature=10000):
+        """Get the position embedding of proposal."""
+        scale = 2 * math.pi
+        dim_t = paddle.arange(num_pos_feats, dtype="float32")
+        dim_t = temperature**(2 * (dim_t // 2) / num_pos_feats)
+        # N, L, 4
+        proposals = F.sigmoid(proposals) * scale
+        # N, L, 4, 128
+        pos = proposals[:, :, :, None] / dim_t
+        # N, L, 4, 64, 2
+        pos = paddle.stack(
+            (pos[:, :, :, 0::2].sin(), pos[:, :, :, 1::2].cos()),
+            axis=4).flatten(2)
+        return pos
+    def forward(self,
+                mlvl_feats,
+                mlvl_masks,
+                query_embed,
+                mlvl_pos_embeds,
+                kpt_branches=None,
+                cls_branches=None):
+        """Forward function for `Transformer`.
+        Args:
+            mlvl_feats (list(Tensor)): Input queries from different level.
+                Each element has shape [bs, embed_dims, h, w].
+            mlvl_masks (list(Tensor)): The key_padding_mask from different
+                level used for encoder and decoder, each element has shape
+                [bs, h, w].
+            query_embed (Tensor): The query embedding for decoder,
+                with shape [num_query, c].
+            mlvl_pos_embeds (list(Tensor)): The positional encoding
+                of feats from different level, has the shape
+                 [bs, embed_dims, h, w].
+            kpt_branches (obj:`nn.LayerList`): Keypoint Regression heads for
+                feature maps from each decoder layer. Only would be passed when
+                `with_box_refine` is Ture. Default to None.
+            cls_branches (obj:`nn.LayerList`): Classification heads for
+                feature maps from each decoder layer. Only would be passed when
+                `as_two_stage` is Ture. Default to None.
+        Returns:
+            tuple[Tensor]: results of decoder containing the following tensor.
+                - inter_states: Outputs from decoder. If
+                    `return_intermediate_dec` is True output has shape \
+                    (num_dec_layers, bs, num_query, embed_dims), else has \
+                    shape (1, bs, num_query, embed_dims).
+                - init_reference_out: The initial value of reference \
+                    points, has shape (bs, num_queries, 4).
+                - inter_references_out: The internal value of reference \
+                    points in decoder, has shape \
+                    (num_dec_layers, bs,num_query, embed_dims)
+                - enc_outputs_class: The classification score of proposals \
+                    generated from encoder's feature maps, has shape \
+                    (batch, h*w, num_classes). \
+                    Only would be returned when `as_two_stage` is True, \
+                    otherwise None.
+                - enc_outputs_kpt_unact: The regression results generated from \
+                    encoder's feature maps., has shape (batch, h*w, K*2).
+                    Only would be returned when `as_two_stage` is True, \
+                    otherwise None.
+        """
+        assert self.as_two_stage or query_embed is not None
+        feat_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        for lvl, (feat, mask, pos_embed
+                  ) in enumerate(zip(mlvl_feats, mlvl_masks, mlvl_pos_embeds)):
+            bs, c, h, w = feat.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+            feat = feat.flatten(2).transpose((0, 2, 1))
+            mask = mask.flatten(1)
+            pos_embed = pos_embed.flatten(2).transpose((0, 2, 1))
+            lvl_pos_embed = pos_embed + self.level_embeds[lvl].reshape(
+                [1, 1, -1])
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            feat_flatten.append(feat)
+            mask_flatten.append(mask)
+        feat_flatten = paddle.concat(feat_flatten, 1)
+        mask_flatten = paddle.concat(mask_flatten, 1)
+        lvl_pos_embed_flatten = paddle.concat(lvl_pos_embed_flatten, 1)
+        spatial_shapes_cumsum = paddle.to_tensor(
+            np.array(spatial_shapes).prod(1).cumsum(0))
+        spatial_shapes = paddle.to_tensor(spatial_shapes, dtype="int64")
+        level_start_index = paddle.concat((paddle.zeros(
+            (1, ), dtype=spatial_shapes.dtype), spatial_shapes_cumsum[:-1]))
+        valid_ratios = paddle.stack(
+            [self.get_valid_ratio(m) for m in mlvl_masks], 1)
+        reference_points = \
+            self.get_reference_points(spatial_shapes,
+                                      valid_ratios)
+        memory = self.encoder(
+            src=feat_flatten,
+            pos_embed=lvl_pos_embed_flatten,
+            src_mask=mask_flatten,
+            value_spatial_shapes=spatial_shapes,
+            reference_points=reference_points,
+            value_level_start_index=level_start_index,
+            valid_ratios=valid_ratios)
+        bs, _, c = memory.shape
+        hm_proto = None
+        if self.training:
+            hm_memory = paddle.slice(
+                memory,
+                starts=level_start_index[0],
+                ends=level_start_index[1],
+                axes=[1])
+            hm_pos_embed = paddle.slice(
+                lvl_pos_embed_flatten,
+                starts=level_start_index[0],
+                ends=level_start_index[1],
+                axes=[1])
+            hm_mask = paddle.slice(
+                mask_flatten,
+                starts=level_start_index[0],
+                ends=level_start_index[1],
+                axes=[1])
+            hm_reference_points = paddle.slice(
+                reference_points,
+                starts=level_start_index[0],
+                ends=level_start_index[1],
+                axes=[1])[:, :, :1, :]
+            # official code make a mistake of pos_embed to pose_embed, which disable pos_embed
+            hm_memory = self.hm_encoder(
+                src=hm_memory,
+                pose_embed=hm_pos_embed,
+                src_mask=hm_mask,
+                value_spatial_shapes=spatial_shapes[[0]],
+                reference_points=hm_reference_points,
+                value_level_start_index=level_start_index[0],
+                valid_ratios=valid_ratios[:, :1, :])
+            hm_memory = hm_memory.reshape((bs, spatial_shapes[0, 0],
+                                           spatial_shapes[0, 1], -1))
+            hm_proto = (hm_memory, mlvl_masks[0])
+        if self.as_two_stage:
+            output_memory, output_proposals = \
+                self.gen_encoder_output_proposals(
+                    memory, mask_flatten, spatial_shapes)
+            enc_outputs_class = cls_branches[self.decoder.num_layers](
+                output_memory)
+            enc_outputs_kpt_unact = \
+                kpt_branches[self.decoder.num_layers](output_memory)
+            enc_outputs_kpt_unact[..., 0::2] += output_proposals[..., 0:1]
+            enc_outputs_kpt_unact[..., 1::2] += output_proposals[..., 1:2]
+            topk = self.two_stage_num_proposals
+            topk_proposals = paddle.topk(
+                enc_outputs_class[..., 0], topk, axis=1)[1].unsqueeze(-1)
+            #paddle.take_along_axis 对应torch.gather
+            topk_kpts_unact = paddle.take_along_axis(enc_outputs_kpt_unact,
+                                                     topk_proposals, 1)
+            topk_kpts_unact = topk_kpts_unact.detach()
+            reference_points = F.sigmoid(topk_kpts_unact)
+            init_reference_out = reference_points
+            # learnable query and query_pos
+            query_pos, query = paddle.split(
+                query_embed, query_embed.shape[1] // c, axis=1)
+            query_pos = query_pos.unsqueeze(0).expand((bs, -1, -1))
+            query = query.unsqueeze(0).expand((bs, -1, -1))
+        else:
+            query_pos, query = paddle.split(
+                query_embed, query_embed.shape[1] // c, axis=1)
+            query_pos = query_pos.unsqueeze(0).expand((bs, -1, -1))
+            query = query.unsqueeze(0).expand((bs, -1, -1))
+            reference_points = F.sigmoid(self.reference_points(query_pos))
+            init_reference_out = reference_points
+        # decoder
+        inter_states, inter_references = self.decoder(
+            query=query,
+            memory=memory,
+            query_pos_embed=query_pos,
+            memory_mask=mask_flatten,
+            reference_points=reference_points,
+            value_spatial_shapes=spatial_shapes,
+            value_level_start_index=level_start_index,
+            valid_ratios=valid_ratios,
+            kpt_branches=kpt_branches)
+        inter_references_out = inter_references
+        if self.as_two_stage:
+            return inter_states, init_reference_out, \
+                   inter_references_out, enc_outputs_class, \
+                   enc_outputs_kpt_unact, hm_proto, memory
+        return inter_states, init_reference_out, \
+               inter_references_out, None, None, None, None, None, hm_proto
+    def forward_refine(self,
+                       mlvl_masks,
+                       memory,
+                       reference_points_pose,
+                       img_inds,
+                       kpt_branches=None,
+                       **kwargs):
+        mask_flatten = []
+        spatial_shapes = []
+        for lvl, mask in enumerate(mlvl_masks):
+            bs, h, w = mask.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+            mask = mask.flatten(1)
+            mask_flatten.append(mask)
+        mask_flatten = paddle.concat(mask_flatten, 1)
+        spatial_shapes_cumsum = paddle.to_tensor(
+            np.array(
+                spatial_shapes, dtype='int64').prod(1).cumsum(0))
+        spatial_shapes = paddle.to_tensor(spatial_shapes, dtype="int64")
+        level_start_index = paddle.concat((paddle.zeros(
+            (1, ), dtype=spatial_shapes.dtype), spatial_shapes_cumsum[:-1]))
+        valid_ratios = paddle.stack(
+            [self.get_valid_ratio(m) for m in mlvl_masks], 1)
+        # pose refinement (17 queries corresponding to 17 keypoints)
+        # learnable query and query_pos
+        refine_query_embedding = self.refine_query_embedding.weight
+        query_pos, query = paddle.split(refine_query_embedding, 2, axis=1)
+        pos_num = reference_points_pose.shape[0]
+        query_pos = query_pos.unsqueeze(0).expand((pos_num, -1, -1))
+        query = query.unsqueeze(0).expand((pos_num, -1, -1))
+        reference_points = reference_points_pose.reshape(
+            (pos_num, reference_points_pose.shape[1] // 2, 2))
+        pos_memory = memory[img_inds]
+        mask_flatten = mask_flatten[img_inds]
+        valid_ratios = valid_ratios[img_inds]
+        if img_inds.size == 1:
+            pos_memory = pos_memory.unsqueeze(0)
+            mask_flatten = mask_flatten.unsqueeze(0)
+            valid_ratios = valid_ratios.unsqueeze(0)
+        inter_states, inter_references = self.refine_decoder(
+            query=query,
+            memory=pos_memory,
+            query_pos_embed=query_pos,
+            memory_mask=mask_flatten,
+            reference_points=reference_points,
+            value_spatial_shapes=spatial_shapes,
+            value_level_start_index=level_start_index,
+            valid_ratios=valid_ratios,
+            reg_branches=kpt_branches,
+            **kwargs)
+        # [num_decoder, num_query, bs, embed_dim]
+        init_reference_out = reference_points
+        return inter_states, init_reference_out, inter_references
--- a/ppdet/utils/visualizer.py
+++ b/ppdet/utils/visualizer.py
@@ -238,7 +238,7 @@ def draw_pose(image,
                     'for example: `pip install matplotlib`.')
        raise e
-    skeletons = np.array([item['keypoints'] for item in results])
+    skeletons = np.array([item['keypoints'] for item in results]).reshape((-1, 51))
    kpt_nums = 17
    if len(skeletons) > 0:
        kpt_nums = int(skeletons.shape[1] / 3)