mmdet/models/dense_heads/corner_head.py

# Copyright (c) OpenMMLab. All rights reserved.
from logging import warning
from math import ceil, log
from typing import List, Optional, Sequence, Tuple

import torch
import torch.nn as nn
from mmcv.cnn import ConvModule
from mmcv.ops import CornerPool, batched_nms
from mmengine.config import ConfigDict
from mmengine.model import BaseModule, bias_init_with_prob
from mmengine.structures import InstanceData
from torch import Tensor

from mmdet.registry import MODELS
from mmdet.utils import (ConfigType, InstanceList, OptConfigType,
                         OptInstanceList, OptMultiConfig)
from ..utils import (gather_feat, gaussian_radius, gen_gaussian_target,
                     get_local_maximum, get_topk_from_heatmap, multi_apply,
                     transpose_and_gather_feat)
from .base_dense_head import BaseDenseHead


class BiCornerPool(BaseModule):
    """Bidirectional Corner Pooling Module (TopLeft, BottomRight, etc.)

    Args:
        in_channels (int): Input channels of module.
        directions (list[str]): Directions of two CornerPools.
        out_channels (int): Output channels of module.
        feat_channels (int): Feature channels of module.
        norm_cfg (:obj:`ConfigDict` or dict): Dictionary to construct
            and config norm layer.
        init_cfg (:obj:`ConfigDict` or dict, optional): the config to
            control the initialization.
    """

    def __init__(self,
                 in_channels: int,
                 directions: List[int],
                 feat_channels: int = 128,
                 out_channels: int = 128,
                 norm_cfg: ConfigType = dict(type='BN', requires_grad=True),
                 init_cfg: OptMultiConfig = None) -> None:
        super().__init__(init_cfg)
        self.direction1_conv = ConvModule(
            in_channels, feat_channels, 3, padding=1, norm_cfg=norm_cfg)
        self.direction2_conv = ConvModule(
            in_channels, feat_channels, 3, padding=1, norm_cfg=norm_cfg)

        self.aftpool_conv = ConvModule(
            feat_channels,
            out_channels,
            3,
            padding=1,
            norm_cfg=norm_cfg,
            act_cfg=None)

        self.conv1 = ConvModule(
            in_channels, out_channels, 1, norm_cfg=norm_cfg, act_cfg=None)
        self.conv2 = ConvModule(
            in_channels, out_channels, 3, padding=1, norm_cfg=norm_cfg)

        self.direction1_pool = CornerPool(directions[0])
        self.direction2_pool = CornerPool(directions[1])
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x: Tensor) -> Tensor:
        """Forward features from the upstream network.

        Args:
            x (tensor): Input feature of BiCornerPool.

        Returns:
            conv2 (tensor): Output feature of BiCornerPool.
        """
        direction1_conv = self.direction1_conv(x)
        direction2_conv = self.direction2_conv(x)
        direction1_feat = self.direction1_pool(direction1_conv)
        direction2_feat = self.direction2_pool(direction2_conv)
        aftpool_conv = self.aftpool_conv(direction1_feat + direction2_feat)
        conv1 = self.conv1(x)
        relu = self.relu(aftpool_conv + conv1)
        conv2 = self.conv2(relu)
        return conv2


@MODELS.register_module()
class CornerHead(BaseDenseHead):
    """Head of CornerNet: Detecting Objects as Paired Keypoints.

    Code is modified from the `official github repo
    <https://github.com/princeton-vl/CornerNet/blob/master/models/py_utils/
    kp.py#L73>`_ .

    More details can be found in the `paper
    <https://arxiv.org/abs/1808.01244>`_ .

    Args:
        num_classes (int): Number of categories excluding the background
            category.
        in_channels (int): Number of channels in the input feature map.
        num_feat_levels (int): Levels of feature from the previous module.
            2 for HourglassNet-104 and 1 for HourglassNet-52. Because
            HourglassNet-104 outputs the final feature and intermediate
            supervision feature and HourglassNet-52 only outputs the final
            feature. Defaults to 2.
        corner_emb_channels (int): Channel of embedding vector. Defaults to 1.
        train_cfg (:obj:`ConfigDict` or dict, optional): Training config.
            Useless in CornerHead, but we keep this variable for
            SingleStageDetector.
        test_cfg (:obj:`ConfigDict` or dict, optional): Testing config of
            CornerHead.
        loss_heatmap (:obj:`ConfigDict` or dict): Config of corner heatmap
            loss. Defaults to GaussianFocalLoss.
        loss_embedding (:obj:`ConfigDict` or dict): Config of corner embedding
            loss. Defaults to AssociativeEmbeddingLoss.
        loss_offset (:obj:`ConfigDict` or dict): Config of corner offset loss.
            Defaults to SmoothL1Loss.
        init_cfg (:obj:`ConfigDict` or dict, optional): the config to control
            the initialization.
    """

    def __init__(self,
                 num_classes: int,
                 in_channels: int,
                 num_feat_levels: int = 2,
                 corner_emb_channels: int = 1,
                 train_cfg: OptConfigType = None,
                 test_cfg: OptConfigType = None,
                 loss_heatmap: ConfigType = dict(
                     type='GaussianFocalLoss',
                     alpha=2.0,
                     gamma=4.0,
                     loss_weight=1),
                 loss_embedding: ConfigType = dict(
                     type='AssociativeEmbeddingLoss',
                     pull_weight=0.25,
                     push_weight=0.25),
                 loss_offset: ConfigType = dict(
                     type='SmoothL1Loss', beta=1.0, loss_weight=1),
                 init_cfg: OptMultiConfig = None) -> None:
        assert init_cfg is None, 'To prevent abnormal initialization ' \
                                 'behavior, init_cfg is not allowed to be set'
        super().__init__(init_cfg=init_cfg)
        self.num_classes = num_classes
        self.in_channels = in_channels
        self.corner_emb_channels = corner_emb_channels
        self.with_corner_emb = self.corner_emb_channels > 0
        self.corner_offset_channels = 2
        self.num_feat_levels = num_feat_levels
        self.loss_heatmap = MODELS.build(
            loss_heatmap) if loss_heatmap is not None else None
        self.loss_embedding = MODELS.build(
            loss_embedding) if loss_embedding is not None else None
        self.loss_offset = MODELS.build(
            loss_offset) if loss_offset is not None else None
        self.train_cfg = train_cfg
        self.test_cfg = test_cfg

        self._init_layers()

    def _make_layers(self,
                     out_channels: int,
                     in_channels: int = 256,
                     feat_channels: int = 256) -> nn.Sequential:
        """Initialize conv sequential for CornerHead."""
        return nn.Sequential(
            ConvModule(in_channels, feat_channels, 3, padding=1),
            ConvModule(
                feat_channels, out_channels, 1, norm_cfg=None, act_cfg=None))

    def _init_corner_kpt_layers(self) -> None:
        """Initialize corner keypoint layers.

        Including corner heatmap branch and corner offset branch. Each branch
        has two parts: prefix `tl_` for top-left and `br_` for bottom-right.
        """
        self.tl_pool, self.br_pool = nn.ModuleList(), nn.ModuleList()
        self.tl_heat, self.br_heat = nn.ModuleList(), nn.ModuleList()
        self.tl_off, self.br_off = nn.ModuleList(), nn.ModuleList()

        for _ in range(self.num_feat_levels):
            self.tl_pool.append(
                BiCornerPool(
                    self.in_channels, ['top', 'left'],
                    out_channels=self.in_channels))
            self.br_pool.append(
                BiCornerPool(
                    self.in_channels, ['bottom', 'right'],
                    out_channels=self.in_channels))

            self.tl_heat.append(
                self._make_layers(
                    out_channels=self.num_classes,
                    in_channels=self.in_channels))
            self.br_heat.append(
                self._make_layers(
                    out_channels=self.num_classes,
                    in_channels=self.in_channels))

            self.tl_off.append(
                self._make_layers(
                    out_channels=self.corner_offset_channels,
                    in_channels=self.in_channels))
            self.br_off.append(
                self._make_layers(
                    out_channels=self.corner_offset_channels,
                    in_channels=self.in_channels))

    def _init_corner_emb_layers(self) -> None:
        """Initialize corner embedding layers.

        Only include corner embedding branch with two parts: prefix `tl_` for
        top-left and `br_` for bottom-right.
        """
        self.tl_emb, self.br_emb = nn.ModuleList(), nn.ModuleList()

        for _ in range(self.num_feat_levels):
            self.tl_emb.append(
                self._make_layers(
                    out_channels=self.corner_emb_channels,
                    in_channels=self.in_channels))
            self.br_emb.append(
                self._make_layers(
                    out_channels=self.corner_emb_channels,
                    in_channels=self.in_channels))

    def _init_layers(self) -> None:
        """Initialize layers for CornerHead.

        Including two parts: corner keypoint layers and corner embedding layers
        """
        self._init_corner_kpt_layers()
        if self.with_corner_emb:
            self._init_corner_emb_layers()

    def init_weights(self) -> None:
        super().init_weights()
        bias_init = bias_init_with_prob(0.1)
        for i in range(self.num_feat_levels):
            # The initialization of parameters are different between
            # nn.Conv2d and ConvModule. Our experiments show that
            # using the original initialization of nn.Conv2d increases
            # the final mAP by about 0.2%
            self.tl_heat[i][-1].conv.reset_parameters()
            self.tl_heat[i][-1].conv.bias.data.fill_(bias_init)
            self.br_heat[i][-1].conv.reset_parameters()
            self.br_heat[i][-1].conv.bias.data.fill_(bias_init)
            self.tl_off[i][-1].conv.reset_parameters()
            self.br_off[i][-1].conv.reset_parameters()
            if self.with_corner_emb:
                self.tl_emb[i][-1].conv.reset_parameters()
                self.br_emb[i][-1].conv.reset_parameters()

    def forward(self, feats: Tuple[Tensor]) -> tuple:
        """Forward features from the upstream network.

        Args:
            feats (tuple[Tensor]): Features from the upstream network, each is
                a 4D-tensor.

        Returns:
            tuple: Usually a tuple of corner heatmaps, offset heatmaps and
            embedding heatmaps.
                - tl_heats (list[Tensor]): Top-left corner heatmaps for all
                  levels, each is a 4D-tensor, the channels number is
                  num_classes.
                - br_heats (list[Tensor]): Bottom-right corner heatmaps for all
                  levels, each is a 4D-tensor, the channels number is
                  num_classes.
                - tl_embs (list[Tensor] | list[None]): Top-left embedding
                  heatmaps for all levels, each is a 4D-tensor or None.
                  If not None, the channels number is corner_emb_channels.
                - br_embs (list[Tensor] | list[None]): Bottom-right embedding
                  heatmaps for all levels, each is a 4D-tensor or None.
                  If not None, the channels number is corner_emb_channels.
                - tl_offs (list[Tensor]): Top-left offset heatmaps for all
                  levels, each is a 4D-tensor. The channels number is
                  corner_offset_channels.
                - br_offs (list[Tensor]): Bottom-right offset heatmaps for all
                  levels, each is a 4D-tensor. The channels number is
                  corner_offset_channels.
        """
        lvl_ind = list(range(self.num_feat_levels))
        return multi_apply(self.forward_single, feats, lvl_ind)

    def forward_single(self,
                       x: Tensor,
                       lvl_ind: int,
                       return_pool: bool = False) -> List[Tensor]:
        """Forward feature of a single level.

        Args:
            x (Tensor): Feature of a single level.
            lvl_ind (int): Level index of current feature.
            return_pool (bool): Return corner pool feature or not.
                Defaults to False.

        Returns:
            tuple[Tensor]: A tuple of CornerHead's output for current feature
            level. Containing the following Tensors:

                - tl_heat (Tensor): Predicted top-left corner heatmap.
                - br_heat (Tensor): Predicted bottom-right corner heatmap.
                - tl_emb (Tensor | None): Predicted top-left embedding heatmap.
                  None for `self.with_corner_emb == False`.
                - br_emb (Tensor | None): Predicted bottom-right embedding
                  heatmap. None for `self.with_corner_emb == False`.
                - tl_off (Tensor): Predicted top-left offset heatmap.
                - br_off (Tensor): Predicted bottom-right offset heatmap.
                - tl_pool (Tensor): Top-left corner pool feature. Not must
                  have.
                - br_pool (Tensor): Bottom-right corner pool feature. Not must
                  have.
        """
        tl_pool = self.tl_pool[lvl_ind](x)
        tl_heat = self.tl_heat[lvl_ind](tl_pool)
        br_pool = self.br_pool[lvl_ind](x)
        br_heat = self.br_heat[lvl_ind](br_pool)

        tl_emb, br_emb = None, None
        if self.with_corner_emb:
            tl_emb = self.tl_emb[lvl_ind](tl_pool)
            br_emb = self.br_emb[lvl_ind](br_pool)

        tl_off = self.tl_off[lvl_ind](tl_pool)
        br_off = self.br_off[lvl_ind](br_pool)

        result_list = [tl_heat, br_heat, tl_emb, br_emb, tl_off, br_off]
        if return_pool:
            result_list.append(tl_pool)
            result_list.append(br_pool)

        return result_list

    def get_targets(self,
                    gt_bboxes: List[Tensor],
                    gt_labels: List[Tensor],
                    feat_shape: Sequence[int],
                    img_shape: Sequence[int],
                    with_corner_emb: bool = False,
                    with_guiding_shift: bool = False,
                    with_centripetal_shift: bool = False) -> dict:
        """Generate corner targets.

        Including corner heatmap, corner offset.

        Optional: corner embedding, corner guiding shift, centripetal shift.

        For CornerNet, we generate corner heatmap, corner offset and corner
        embedding from this function.

        For CentripetalNet, we generate corner heatmap, corner offset, guiding
        shift and centripetal shift from this function.

        Args:
            gt_bboxes (list[Tensor]): Ground truth bboxes of each image, each
                has shape (num_gt, 4).
            gt_labels (list[Tensor]): Ground truth labels of each box, each has
                shape (num_gt, ).
            feat_shape (Sequence[int]): Shape of output feature,
                [batch, channel, height, width].
            img_shape (Sequence[int]): Shape of input image,
                [height, width, channel].
            with_corner_emb (bool): Generate corner embedding target or not.
                Defaults to False.
            with_guiding_shift (bool): Generate guiding shift target or not.
                Defaults to False.
            with_centripetal_shift (bool): Generate centripetal shift target or
                not. Defaults to False.

        Returns:
            dict: Ground truth of corner heatmap, corner offset, corner
            embedding, guiding shift and centripetal shift. Containing the
            following keys:

                - topleft_heatmap (Tensor): Ground truth top-left corner
                  heatmap.
                - bottomright_heatmap (Tensor): Ground truth bottom-right
                  corner heatmap.
                - topleft_offset (Tensor): Ground truth top-left corner offset.
                - bottomright_offset (Tensor): Ground truth bottom-right corner
                  offset.
                - corner_embedding (list[list[list[int]]]): Ground truth corner
                  embedding. Not must have.
                - topleft_guiding_shift (Tensor): Ground truth top-left corner
                  guiding shift. Not must have.
                - bottomright_guiding_shift (Tensor): Ground truth bottom-right
                  corner guiding shift. Not must have.
                - topleft_centripetal_shift (Tensor): Ground truth top-left
                  corner centripetal shift. Not must have.
                - bottomright_centripetal_shift (Tensor): Ground truth
                  bottom-right corner centripetal shift. Not must have.
        """
        batch_size, _, height, width = feat_shape
        img_h, img_w = img_shape[:2]

        width_ratio = float(width / img_w)
        height_ratio = float(height / img_h)

        gt_tl_heatmap = gt_bboxes[-1].new_zeros(
            [batch_size, self.num_classes, height, width])
        gt_br_heatmap = gt_bboxes[-1].new_zeros(
            [batch_size, self.num_classes, height, width])
        gt_tl_offset = gt_bboxes[-1].new_zeros([batch_size, 2, height, width])
        gt_br_offset = gt_bboxes[-1].new_zeros([batch_size, 2, height, width])

        if with_corner_emb:
            match = []

        # Guiding shift is a kind of offset, from center to corner
        if with_guiding_shift:
            gt_tl_guiding_shift = gt_bboxes[-1].new_zeros(
                [batch_size, 2, height, width])
            gt_br_guiding_shift = gt_bboxes[-1].new_zeros(
                [batch_size, 2, height, width])
        # Centripetal shift is also a kind of offset, from center to corner
        # and normalized by log.
        if with_centripetal_shift:
            gt_tl_centripetal_shift = gt_bboxes[-1].new_zeros(
                [batch_size, 2, height, width])
            gt_br_centripetal_shift = gt_bboxes[-1].new_zeros(
                [batch_size, 2, height, width])

        for batch_id in range(batch_size):
            # Ground truth of corner embedding per image is a list of coord set
            corner_match = []
            for box_id in range(len(gt_labels[batch_id])):
                left, top, right, bottom = gt_bboxes[batch_id][box_id]
                center_x = (left + right) / 2.0
                center_y = (top + bottom) / 2.0
                label = gt_labels[batch_id][box_id]

                # Use coords in the feature level to generate ground truth
                scale_left = left * width_ratio
                scale_right = right * width_ratio
                scale_top = top * height_ratio
                scale_bottom = bottom * height_ratio
                scale_center_x = center_x * width_ratio
                scale_center_y = center_y * height_ratio

                # Int coords on feature map/ground truth tensor
                left_idx = int(min(scale_left, width - 1))
                right_idx = int(min(scale_right, width - 1))
                top_idx = int(min(scale_top, height - 1))
                bottom_idx = int(min(scale_bottom, height - 1))

                # Generate gaussian heatmap
                scale_box_width = ceil(scale_right - scale_left)
                scale_box_height = ceil(scale_bottom - scale_top)
                radius = gaussian_radius((scale_box_height, scale_box_width),
                                         min_overlap=0.3)
                radius = max(0, int(radius))
                gt_tl_heatmap[batch_id, label] = gen_gaussian_target(
                    gt_tl_heatmap[batch_id, label], [left_idx, top_idx],
                    radius)
                gt_br_heatmap[batch_id, label] = gen_gaussian_target(
                    gt_br_heatmap[batch_id, label], [right_idx, bottom_idx],
                    radius)

                # Generate corner offset
                left_offset = scale_left - left_idx
                top_offset = scale_top - top_idx
                right_offset = scale_right - right_idx
                bottom_offset = scale_bottom - bottom_idx
                gt_tl_offset[batch_id, 0, top_idx, left_idx] = left_offset
                gt_tl_offset[batch_id, 1, top_idx, left_idx] = top_offset
                gt_br_offset[batch_id, 0, bottom_idx, right_idx] = right_offset
                gt_br_offset[batch_id, 1, bottom_idx,
                             right_idx] = bottom_offset

                # Generate corner embedding
                if with_corner_emb:
                    corner_match.append([[top_idx, left_idx],
                                         [bottom_idx, right_idx]])
                # Generate guiding shift
                if with_guiding_shift:
                    gt_tl_guiding_shift[batch_id, 0, top_idx,
                                        left_idx] = scale_center_x - left_idx
                    gt_tl_guiding_shift[batch_id, 1, top_idx,
                                        left_idx] = scale_center_y - top_idx
                    gt_br_guiding_shift[batch_id, 0, bottom_idx,
                                        right_idx] = right_idx - scale_center_x
                    gt_br_guiding_shift[
                        batch_id, 1, bottom_idx,
                        right_idx] = bottom_idx - scale_center_y
                # Generate centripetal shift
                if with_centripetal_shift:
                    gt_tl_centripetal_shift[batch_id, 0, top_idx,
                                            left_idx] = log(scale_center_x -
                                                            scale_left)
                    gt_tl_centripetal_shift[batch_id, 1, top_idx,
                                            left_idx] = log(scale_center_y -
                                                            scale_top)
                    gt_br_centripetal_shift[batch_id, 0, bottom_idx,
                                            right_idx] = log(scale_right -
                                                             scale_center_x)
                    gt_br_centripetal_shift[batch_id, 1, bottom_idx,
                                            right_idx] = log(scale_bottom -
                                                             scale_center_y)

            if with_corner_emb:
                match.append(corner_match)

        target_result = dict(
            topleft_heatmap=gt_tl_heatmap,
            topleft_offset=gt_tl_offset,
            bottomright_heatmap=gt_br_heatmap,
            bottomright_offset=gt_br_offset)

        if with_corner_emb:
            target_result.update(corner_embedding=match)
        if with_guiding_shift:
            target_result.update(
                topleft_guiding_shift=gt_tl_guiding_shift,
                bottomright_guiding_shift=gt_br_guiding_shift)
        if with_centripetal_shift:
            target_result.update(
                topleft_centripetal_shift=gt_tl_centripetal_shift,
                bottomright_centripetal_shift=gt_br_centripetal_shift)

        return target_result

    def loss_by_feat(
            self,
            tl_heats: List[Tensor],
            br_heats: List[Tensor],
            tl_embs: List[Tensor],
            br_embs: List[Tensor],
            tl_offs: List[Tensor],
            br_offs: List[Tensor],
            batch_gt_instances: InstanceList,
            batch_img_metas: List[dict],
            batch_gt_instances_ignore: OptInstanceList = None) -> dict:
        """Calculate the loss based on the features extracted by the detection
        head.

        Args:
            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
                with shape (N, num_classes, H, W).
            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
                level with shape (N, num_classes, H, W).
            tl_embs (list[Tensor]): Top-left corner embeddings for each level
                with shape (N, corner_emb_channels, H, W).
            br_embs (list[Tensor]): Bottom-right corner embeddings for each
                level with shape (N, corner_emb_channels, H, W).
            tl_offs (list[Tensor]): Top-left corner offsets for each level
                with shape (N, corner_offset_channels, H, W).
            br_offs (list[Tensor]): Bottom-right corner offsets for each level
                with shape (N, corner_offset_channels, H, W).
            batch_gt_instances (list[:obj:`InstanceData`]): Batch of
                gt_instance. It usually includes ``bboxes`` and ``labels``
                attributes.
            batch_img_metas (list[dict]): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            batch_gt_instances_ignore (list[:obj:`InstanceData`], optional):
                Specify which bounding boxes can be ignored when computing
                the loss.

        Returns:
            dict[str, Tensor]: A dictionary of loss components. Containing the
            following losses:

                - det_loss (list[Tensor]): Corner keypoint losses of all
                  feature levels.
                - pull_loss (list[Tensor]): Part one of AssociativeEmbedding
                  losses of all feature levels.
                - push_loss (list[Tensor]): Part two of AssociativeEmbedding
                  losses of all feature levels.
                - off_loss (list[Tensor]): Corner offset losses of all feature
                  levels.
        """
        gt_bboxes = [
            gt_instances.bboxes for gt_instances in batch_gt_instances
        ]
        gt_labels = [
            gt_instances.labels for gt_instances in batch_gt_instances
        ]

        targets = self.get_targets(
            gt_bboxes,
            gt_labels,
            tl_heats[-1].shape,
            batch_img_metas[0]['batch_input_shape'],
            with_corner_emb=self.with_corner_emb)
        mlvl_targets = [targets for _ in range(self.num_feat_levels)]
        det_losses, pull_losses, push_losses, off_losses = multi_apply(
            self.loss_by_feat_single, tl_heats, br_heats, tl_embs, br_embs,
            tl_offs, br_offs, mlvl_targets)
        loss_dict = dict(det_loss=det_losses, off_loss=off_losses)
        if self.with_corner_emb:
            loss_dict.update(pull_loss=pull_losses, push_loss=push_losses)
        return loss_dict

    def loss_by_feat_single(self, tl_hmp: Tensor, br_hmp: Tensor,
                            tl_emb: Optional[Tensor], br_emb: Optional[Tensor],
                            tl_off: Tensor, br_off: Tensor,
                            targets: dict) -> Tuple[Tensor, ...]:
        """Calculate the loss of a single scale level based on the features
        extracted by the detection head.

        Args:
            tl_hmp (Tensor): Top-left corner heatmap for current level with
                shape (N, num_classes, H, W).
            br_hmp (Tensor): Bottom-right corner heatmap for current level with
                shape (N, num_classes, H, W).
            tl_emb (Tensor, optional): Top-left corner embedding for current
                level with shape (N, corner_emb_channels, H, W).
            br_emb (Tensor, optional): Bottom-right corner embedding for
                current level with shape (N, corner_emb_channels, H, W).
            tl_off (Tensor): Top-left corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            br_off (Tensor): Bottom-right corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            targets (dict): Corner target generated by `get_targets`.

        Returns:
            tuple[torch.Tensor]: Losses of the head's different branches
            containing the following losses:

                - det_loss (Tensor): Corner keypoint loss.
                - pull_loss (Tensor): Part one of AssociativeEmbedding loss.
                - push_loss (Tensor): Part two of AssociativeEmbedding loss.
                - off_loss (Tensor): Corner offset loss.
        """
        gt_tl_hmp = targets['topleft_heatmap']
        gt_br_hmp = targets['bottomright_heatmap']
        gt_tl_off = targets['topleft_offset']
        gt_br_off = targets['bottomright_offset']
        gt_embedding = targets['corner_embedding']

        # Detection loss
        tl_det_loss = self.loss_heatmap(
            tl_hmp.sigmoid(),
            gt_tl_hmp,
            avg_factor=max(1, gt_tl_hmp.eq(1).sum()))
        br_det_loss = self.loss_heatmap(
            br_hmp.sigmoid(),
            gt_br_hmp,
            avg_factor=max(1, gt_br_hmp.eq(1).sum()))
        det_loss = (tl_det_loss + br_det_loss) / 2.0

        # AssociativeEmbedding loss
        if self.with_corner_emb and self.loss_embedding is not None:
            pull_loss, push_loss = self.loss_embedding(tl_emb, br_emb,
                                                       gt_embedding)
        else:
            pull_loss, push_loss = None, None

        # Offset loss
        # We only compute the offset loss at the real corner position.
        # The value of real corner would be 1 in heatmap ground truth.
        # The mask is computed in class agnostic mode and its shape is
        # batch * 1 * width * height.
        tl_off_mask = gt_tl_hmp.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
            gt_tl_hmp)
        br_off_mask = gt_br_hmp.eq(1).sum(1).gt(0).unsqueeze(1).type_as(
            gt_br_hmp)
        tl_off_loss = self.loss_offset(
            tl_off,
            gt_tl_off,
            tl_off_mask,
            avg_factor=max(1, tl_off_mask.sum()))
        br_off_loss = self.loss_offset(
            br_off,
            gt_br_off,
            br_off_mask,
            avg_factor=max(1, br_off_mask.sum()))

        off_loss = (tl_off_loss + br_off_loss) / 2.0

        return det_loss, pull_loss, push_loss, off_loss

    def predict_by_feat(self,
                        tl_heats: List[Tensor],
                        br_heats: List[Tensor],
                        tl_embs: List[Tensor],
                        br_embs: List[Tensor],
                        tl_offs: List[Tensor],
                        br_offs: List[Tensor],
                        batch_img_metas: Optional[List[dict]] = None,
                        rescale: bool = False,
                        with_nms: bool = True) -> InstanceList:
        """Transform a batch of output features extracted from the head into
        bbox results.

        Args:
            tl_heats (list[Tensor]): Top-left corner heatmaps for each level
                with shape (N, num_classes, H, W).
            br_heats (list[Tensor]): Bottom-right corner heatmaps for each
                level with shape (N, num_classes, H, W).
            tl_embs (list[Tensor]): Top-left corner embeddings for each level
                with shape (N, corner_emb_channels, H, W).
            br_embs (list[Tensor]): Bottom-right corner embeddings for each
                level with shape (N, corner_emb_channels, H, W).
            tl_offs (list[Tensor]): Top-left corner offsets for each level
                with shape (N, corner_offset_channels, H, W).
            br_offs (list[Tensor]): Bottom-right corner offsets for each level
                with shape (N, corner_offset_channels, H, W).
            batch_img_metas (list[dict], optional): Batch image meta info.
                Defaults to None.
            rescale (bool): If True, return boxes in original image space.
                Defaults to False.
            with_nms (bool): If True, do nms before return boxes.
                Defaults to True.

        Returns:
            list[:obj:`InstanceData`]: Object detection results of each image
            after the post process. Each item usually contains following keys.

                - scores (Tensor): Classification scores, has a shape
                  (num_instance, )
                - labels (Tensor): Labels of bboxes, has a shape
                  (num_instances, ).
                - bboxes (Tensor): Has a shape (num_instances, 4),
                  the last dimension 4 arrange as (x1, y1, x2, y2).
        """
        assert tl_heats[-1].shape[0] == br_heats[-1].shape[0] == len(
            batch_img_metas)
        result_list = []
        for img_id in range(len(batch_img_metas)):
            result_list.append(
                self._predict_by_feat_single(
                    tl_heats[-1][img_id:img_id + 1, :],
                    br_heats[-1][img_id:img_id + 1, :],
                    tl_offs[-1][img_id:img_id + 1, :],
                    br_offs[-1][img_id:img_id + 1, :],
                    batch_img_metas[img_id],
                    tl_emb=tl_embs[-1][img_id:img_id + 1, :],
                    br_emb=br_embs[-1][img_id:img_id + 1, :],
                    rescale=rescale,
                    with_nms=with_nms))

        return result_list

    def _predict_by_feat_single(self,
                                tl_heat: Tensor,
                                br_heat: Tensor,
                                tl_off: Tensor,
                                br_off: Tensor,
                                img_meta: dict,
                                tl_emb: Optional[Tensor] = None,
                                br_emb: Optional[Tensor] = None,
                                tl_centripetal_shift: Optional[Tensor] = None,
                                br_centripetal_shift: Optional[Tensor] = None,
                                rescale: bool = False,
                                with_nms: bool = True) -> InstanceData:
        """Transform a single image's features extracted from the head into
        bbox results.

        Args:
            tl_heat (Tensor): Top-left corner heatmap for current level with
                shape (N, num_classes, H, W).
            br_heat (Tensor): Bottom-right corner heatmap for current level
                with shape (N, num_classes, H, W).
            tl_off (Tensor): Top-left corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            br_off (Tensor): Bottom-right corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            img_meta (dict): Meta information of current image, e.g.,
                image size, scaling factor, etc.
            tl_emb (Tensor): Top-left corner embedding for current level with
                shape (N, corner_emb_channels, H, W).
            br_emb (Tensor): Bottom-right corner embedding for current level
                with shape (N, corner_emb_channels, H, W).
            tl_centripetal_shift: Top-left corner's centripetal shift for
                current level with shape (N, 2, H, W).
            br_centripetal_shift: Bottom-right corner's centripetal shift for
                current level with shape (N, 2, H, W).
            rescale (bool): If True, return boxes in original image space.
                Defaults to False.
            with_nms (bool): If True, do nms before return boxes.
                Defaults to True.

        Returns:
            :obj:`InstanceData`: Detection results of each image
            after the post process.
            Each item usually contains following keys.

                - scores (Tensor): Classification scores, has a shape
                  (num_instance, )
                - labels (Tensor): Labels of bboxes, has a shape
                  (num_instances, ).
                - bboxes (Tensor): Has a shape (num_instances, 4),
                  the last dimension 4 arrange as (x1, y1, x2, y2).
        """
        if isinstance(img_meta, (list, tuple)):
            img_meta = img_meta[0]

        batch_bboxes, batch_scores, batch_clses = self._decode_heatmap(
            tl_heat=tl_heat.sigmoid(),
            br_heat=br_heat.sigmoid(),
            tl_off=tl_off,
            br_off=br_off,
            tl_emb=tl_emb,
            br_emb=br_emb,
            tl_centripetal_shift=tl_centripetal_shift,
            br_centripetal_shift=br_centripetal_shift,
            img_meta=img_meta,
            k=self.test_cfg.corner_topk,
            kernel=self.test_cfg.local_maximum_kernel,
            distance_threshold=self.test_cfg.distance_threshold)

        if rescale and 'scale_factor' in img_meta:
            batch_bboxes /= batch_bboxes.new_tensor(
                img_meta['scale_factor']).repeat((1, 2))

        bboxes = batch_bboxes.view([-1, 4])
        scores = batch_scores.view(-1)
        clses = batch_clses.view(-1)

        det_bboxes = torch.cat([bboxes, scores.unsqueeze(-1)], -1)
        keepinds = (det_bboxes[:, -1] > -0.1)
        det_bboxes = det_bboxes[keepinds]
        det_labels = clses[keepinds]

        if with_nms:
            det_bboxes, det_labels = self._bboxes_nms(det_bboxes, det_labels,
                                                      self.test_cfg)

        results = InstanceData()
        results.bboxes = det_bboxes[..., :4]
        results.scores = det_bboxes[..., 4]
        results.labels = det_labels
        return results

    def _bboxes_nms(self, bboxes: Tensor, labels: Tensor,
                    cfg: ConfigDict) -> Tuple[Tensor, Tensor]:
        """bboxes nms."""
        if 'nms_cfg' in cfg:
            warning.warn('nms_cfg in test_cfg will be deprecated. '
                         'Please rename it as nms')
        if 'nms' not in cfg:
            cfg.nms = cfg.nms_cfg

        if labels.numel() > 0:
            max_num = cfg.max_per_img
            bboxes, keep = batched_nms(bboxes[:, :4], bboxes[:,
                                                             -1].contiguous(),
                                       labels, cfg.nms)
            if max_num > 0:
                bboxes = bboxes[:max_num]
                labels = labels[keep][:max_num]

        return bboxes, labels

    def _decode_heatmap(self,
                        tl_heat: Tensor,
                        br_heat: Tensor,
                        tl_off: Tensor,
                        br_off: Tensor,
                        tl_emb: Optional[Tensor] = None,
                        br_emb: Optional[Tensor] = None,
                        tl_centripetal_shift: Optional[Tensor] = None,
                        br_centripetal_shift: Optional[Tensor] = None,
                        img_meta: Optional[dict] = None,
                        k: int = 100,
                        kernel: int = 3,
                        distance_threshold: float = 0.5,
                        num_dets: int = 1000) -> Tuple[Tensor, Tensor, Tensor]:
        """Transform outputs into detections raw bbox prediction.

        Args:
            tl_heat (Tensor): Top-left corner heatmap for current level with
                shape (N, num_classes, H, W).
            br_heat (Tensor): Bottom-right corner heatmap for current level
                with shape (N, num_classes, H, W).
            tl_off (Tensor): Top-left corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            br_off (Tensor): Bottom-right corner offset for current level with
                shape (N, corner_offset_channels, H, W).
            tl_emb (Tensor, Optional): Top-left corner embedding for current
                level with shape (N, corner_emb_channels, H, W).
            br_emb (Tensor, Optional): Bottom-right corner embedding for
                current level with shape (N, corner_emb_channels, H, W).
            tl_centripetal_shift (Tensor, Optional): Top-left centripetal shift
                for current level with shape (N, 2, H, W).
            br_centripetal_shift (Tensor, Optional): Bottom-right centripetal
                shift for current level with shape (N, 2, H, W).
            img_meta (dict): Meta information of current image, e.g.,
                image size, scaling factor, etc.
            k (int): Get top k corner keypoints from heatmap.
            kernel (int): Max pooling kernel for extract local maximum pixels.
            distance_threshold (float): Distance threshold. Top-left and
                bottom-right corner keypoints with feature distance less than
                the threshold will be regarded as keypoints from same object.
            num_dets (int): Num of raw boxes before doing nms.

        Returns:
            tuple[torch.Tensor]: Decoded output of CornerHead, containing the
            following Tensors:

            - bboxes (Tensor): Coords of each box.
            - scores (Tensor): Scores of each box.
            - clses (Tensor): Categories of each box.
        """
        with_embedding = tl_emb is not None and br_emb is not None
        with_centripetal_shift = (
            tl_centripetal_shift is not None
            and br_centripetal_shift is not None)
        assert with_embedding + with_centripetal_shift == 1
        batch, _, height, width = tl_heat.size()
        if torch.onnx.is_in_onnx_export():
            inp_h, inp_w = img_meta['pad_shape_for_onnx'][:2]
        else:
            inp_h, inp_w = img_meta['batch_input_shape'][:2]

        # perform nms on heatmaps
        tl_heat = get_local_maximum(tl_heat, kernel=kernel)
        br_heat = get_local_maximum(br_heat, kernel=kernel)

        tl_scores, tl_inds, tl_clses, tl_ys, tl_xs = get_topk_from_heatmap(
            tl_heat, k=k)
        br_scores, br_inds, br_clses, br_ys, br_xs = get_topk_from_heatmap(
            br_heat, k=k)

        # We use repeat instead of expand here because expand is a
        # shallow-copy function. Thus it could cause unexpected testing result
        # sometimes. Using expand will decrease about 10% mAP during testing
        # compared to repeat.
        tl_ys = tl_ys.view(batch, k, 1).repeat(1, 1, k)
        tl_xs = tl_xs.view(batch, k, 1).repeat(1, 1, k)
        br_ys = br_ys.view(batch, 1, k).repeat(1, k, 1)
        br_xs = br_xs.view(batch, 1, k).repeat(1, k, 1)

        tl_off = transpose_and_gather_feat(tl_off, tl_inds)
        tl_off = tl_off.view(batch, k, 1, 2)
        br_off = transpose_and_gather_feat(br_off, br_inds)
        br_off = br_off.view(batch, 1, k, 2)

        tl_xs = tl_xs + tl_off[..., 0]
        tl_ys = tl_ys + tl_off[..., 1]
        br_xs = br_xs + br_off[..., 0]
        br_ys = br_ys + br_off[..., 1]

        if with_centripetal_shift:
            tl_centripetal_shift = transpose_and_gather_feat(
                tl_centripetal_shift, tl_inds).view(batch, k, 1, 2).exp()
            br_centripetal_shift = transpose_and_gather_feat(
                br_centripetal_shift, br_inds).view(batch, 1, k, 2).exp()

            tl_ctxs = tl_xs + tl_centripetal_shift[..., 0]
            tl_ctys = tl_ys + tl_centripetal_shift[..., 1]
            br_ctxs = br_xs - br_centripetal_shift[..., 0]
            br_ctys = br_ys - br_centripetal_shift[..., 1]

        # all possible boxes based on top k corners (ignoring class)
        tl_xs *= (inp_w / width)
        tl_ys *= (inp_h / height)
        br_xs *= (inp_w / width)
        br_ys *= (inp_h / height)

        if with_centripetal_shift:
            tl_ctxs *= (inp_w / width)
            tl_ctys *= (inp_h / height)
            br_ctxs *= (inp_w / width)
            br_ctys *= (inp_h / height)

        x_off, y_off = 0, 0  # no crop
        if not torch.onnx.is_in_onnx_export():
            # since `RandomCenterCropPad` is done on CPU with numpy and it's
            # not dynamic traceable when exporting to ONNX, thus 'border'
            # does not appears as key in 'img_meta'. As a tmp solution,
            # we move this 'border' handle part to the postprocess after
            # finished exporting to ONNX, which is handle in
            # `mmdet/core/export/model_wrappers.py`. Though difference between
            # pytorch and exported onnx model, it might be ignored since
            # comparable performance is achieved between them (e.g. 40.4 vs
            # 40.6 on COCO val2017, for CornerNet without test-time flip)
            if 'border' in img_meta:
                x_off = img_meta['border'][2]
                y_off = img_meta['border'][0]

        tl_xs -= x_off
        tl_ys -= y_off
        br_xs -= x_off
        br_ys -= y_off

        zeros = tl_xs.new_zeros(*tl_xs.size())
        tl_xs = torch.where(tl_xs > 0.0, tl_xs, zeros)
        tl_ys = torch.where(tl_ys > 0.0, tl_ys, zeros)
        br_xs = torch.where(br_xs > 0.0, br_xs, zeros)
        br_ys = torch.where(br_ys > 0.0, br_ys, zeros)

        bboxes = torch.stack((tl_xs, tl_ys, br_xs, br_ys), dim=3)
        area_bboxes = ((br_xs - tl_xs) * (br_ys - tl_ys)).abs()

        if with_centripetal_shift:
            tl_ctxs -= x_off
            tl_ctys -= y_off
            br_ctxs -= x_off
            br_ctys -= y_off

            tl_ctxs *= tl_ctxs.gt(0.0).type_as(tl_ctxs)
            tl_ctys *= tl_ctys.gt(0.0).type_as(tl_ctys)
            br_ctxs *= br_ctxs.gt(0.0).type_as(br_ctxs)
            br_ctys *= br_ctys.gt(0.0).type_as(br_ctys)

            ct_bboxes = torch.stack((tl_ctxs, tl_ctys, br_ctxs, br_ctys),
                                    dim=3)
            area_ct_bboxes = ((br_ctxs - tl_ctxs) * (br_ctys - tl_ctys)).abs()

            rcentral = torch.zeros_like(ct_bboxes)
            # magic nums from paper section 4.1
            mu = torch.ones_like(area_bboxes) / 2.4
            mu[area_bboxes > 3500] = 1 / 2.1  # large bbox have smaller mu

            bboxes_center_x = (bboxes[..., 0] + bboxes[..., 2]) / 2
            bboxes_center_y = (bboxes[..., 1] + bboxes[..., 3]) / 2
            rcentral[..., 0] = bboxes_center_x - mu * (bboxes[..., 2] -
                                                       bboxes[..., 0]) / 2
            rcentral[..., 1] = bboxes_center_y - mu * (bboxes[..., 3] -
                                                       bboxes[..., 1]) / 2
            rcentral[..., 2] = bboxes_center_x + mu * (bboxes[..., 2] -
                                                       bboxes[..., 0]) / 2
            rcentral[..., 3] = bboxes_center_y + mu * (bboxes[..., 3] -
                                                       bboxes[..., 1]) / 2
            area_rcentral = ((rcentral[..., 2] - rcentral[..., 0]) *
                             (rcentral[..., 3] - rcentral[..., 1])).abs()
            dists = area_ct_bboxes / area_rcentral

            tl_ctx_inds = (ct_bboxes[..., 0] <= rcentral[..., 0]) | (
                ct_bboxes[..., 0] >= rcentral[..., 2])
            tl_cty_inds = (ct_bboxes[..., 1] <= rcentral[..., 1]) | (
                ct_bboxes[..., 1] >= rcentral[..., 3])
            br_ctx_inds = (ct_bboxes[..., 2] <= rcentral[..., 0]) | (
                ct_bboxes[..., 2] >= rcentral[..., 2])
            br_cty_inds = (ct_bboxes[..., 3] <= rcentral[..., 1]) | (
                ct_bboxes[..., 3] >= rcentral[..., 3])

        if with_embedding:
            tl_emb = transpose_and_gather_feat(tl_emb, tl_inds)
            tl_emb = tl_emb.view(batch, k, 1)
            br_emb = transpose_and_gather_feat(br_emb, br_inds)
            br_emb = br_emb.view(batch, 1, k)
            dists = torch.abs(tl_emb - br_emb)

        tl_scores = tl_scores.view(batch, k, 1).repeat(1, 1, k)
        br_scores = br_scores.view(batch, 1, k).repeat(1, k, 1)

        scores = (tl_scores + br_scores) / 2  # scores for all possible boxes

        # tl and br should have same class
        tl_clses = tl_clses.view(batch, k, 1).repeat(1, 1, k)
        br_clses = br_clses.view(batch, 1, k).repeat(1, k, 1)
        cls_inds = (tl_clses != br_clses)

        # reject boxes based on distances
        dist_inds = dists > distance_threshold

        # reject boxes based on widths and heights
        width_inds = (br_xs <= tl_xs)
        height_inds = (br_ys <= tl_ys)

        # No use `scores[cls_inds]`, instead we use `torch.where` here.
        # Since only 1-D indices with type 'tensor(bool)' are supported
        # when exporting to ONNX, any other bool indices with more dimensions
        # (e.g. 2-D bool tensor) as input parameter in node is invalid
        negative_scores = -1 * torch.ones_like(scores)
        scores = torch.where(cls_inds, negative_scores, scores)
        scores = torch.where(width_inds, negative_scores, scores)
        scores = torch.where(height_inds, negative_scores, scores)
        scores = torch.where(dist_inds, negative_scores, scores)

        if with_centripetal_shift:
            scores[tl_ctx_inds] = -1
            scores[tl_cty_inds] = -1
            scores[br_ctx_inds] = -1
            scores[br_cty_inds] = -1

        scores = scores.view(batch, -1)
        scores, inds = torch.topk(scores, num_dets)
        scores = scores.unsqueeze(2)

        bboxes = bboxes.view(batch, -1, 4)
        bboxes = gather_feat(bboxes, inds)

        clses = tl_clses.contiguous().view(batch, -1, 1)
        clses = gather_feat(clses, inds)

        return bboxes, scores, clses