mmdet/models/dense_heads/guided_anchor_head.py

# Copyright (c) OpenMMLab. All rights reserved.
import warnings

import torch
import torch.nn as nn
from mmcv.ops import DeformConv2d, MaskedConv2d
from mmcv.runner import BaseModule, force_fp32

from mmdet.core import (anchor_inside_flags, build_assigner, build_bbox_coder,
                        build_prior_generator, build_sampler, calc_region,
                        images_to_levels, multi_apply, multiclass_nms, unmap)
from ..builder import HEADS, build_loss
from .anchor_head import AnchorHead


class FeatureAdaption(BaseModule):
    """Feature Adaption Module.

    Feature Adaption Module is implemented based on DCN v1.
    It uses anchor shape prediction rather than feature map to
    predict offsets of deform conv layer.

    Args:
        in_channels (int): Number of channels in the input feature map.
        out_channels (int): Number of channels in the output feature map.
        kernel_size (int): Deformable conv kernel size.
        deform_groups (int): Deformable conv group size.
        init_cfg (dict or list[dict], optional): Initialization config dict.
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size=3,
                 deform_groups=4,
                 init_cfg=dict(
                     type='Normal',
                     layer='Conv2d',
                     std=0.1,
                     override=dict(
                         type='Normal', name='conv_adaption', std=0.01))):
        super(FeatureAdaption, self).__init__(init_cfg)
        offset_channels = kernel_size * kernel_size * 2
        self.conv_offset = nn.Conv2d(
            2, deform_groups * offset_channels, 1, bias=False)
        self.conv_adaption = DeformConv2d(
            in_channels,
            out_channels,
            kernel_size=kernel_size,
            padding=(kernel_size - 1) // 2,
            deform_groups=deform_groups)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x, shape):
        offset = self.conv_offset(shape.detach())
        x = self.relu(self.conv_adaption(x, offset))
        return x


@HEADS.register_module()
class GuidedAnchorHead(AnchorHead):
    """Guided-Anchor-based head (GA-RPN, GA-RetinaNet, etc.).

    This GuidedAnchorHead will predict high-quality feature guided
    anchors and locations where anchors will be kept in inference.
    There are mainly 3 categories of bounding-boxes.

    - Sampled 9 pairs for target assignment. (approxes)
    - The square boxes where the predicted anchors are based on. (squares)
    - Guided anchors.

    Please refer to https://arxiv.org/abs/1901.03278 for more details.

    Args:
        num_classes (int): Number of classes.
        in_channels (int): Number of channels in the input feature map.
        feat_channels (int): Number of hidden channels.
        approx_anchor_generator (dict): Config dict for approx generator
        square_anchor_generator (dict): Config dict for square generator
        anchor_coder (dict): Config dict for anchor coder
        bbox_coder (dict): Config dict for bbox coder
        reg_decoded_bbox (bool): If true, the regression loss would be
            applied directly on decoded bounding boxes, converting both
            the predicted boxes and regression targets to absolute
            coordinates format. Default False. It should be `True` when
            using `IoULoss`, `GIoULoss`, or `DIoULoss` in the bbox head.
        deform_groups: (int): Group number of DCN in
            FeatureAdaption module.
        loc_filter_thr (float): Threshold to filter out unconcerned regions.
        loss_loc (dict): Config of location loss.
        loss_shape (dict): Config of anchor shape loss.
        loss_cls (dict): Config of classification loss.
        loss_bbox (dict): Config of bbox regression loss.
        init_cfg (dict or list[dict], optional): Initialization config dict.
    """

    def __init__(
            self,
            num_classes,
            in_channels,
            feat_channels=256,
            approx_anchor_generator=dict(
                type='AnchorGenerator',
                octave_base_scale=8,
                scales_per_octave=3,
                ratios=[0.5, 1.0, 2.0],
                strides=[4, 8, 16, 32, 64]),
            square_anchor_generator=dict(
                type='AnchorGenerator',
                ratios=[1.0],
                scales=[8],
                strides=[4, 8, 16, 32, 64]),
            anchor_coder=dict(
                type='DeltaXYWHBBoxCoder',
                target_means=[.0, .0, .0, .0],
                target_stds=[1.0, 1.0, 1.0, 1.0]
            ),
            bbox_coder=dict(
                type='DeltaXYWHBBoxCoder',
                target_means=[.0, .0, .0, .0],
                target_stds=[1.0, 1.0, 1.0, 1.0]
            ),
            reg_decoded_bbox=False,
            deform_groups=4,
            loc_filter_thr=0.01,
            train_cfg=None,
            test_cfg=None,
            loss_loc=dict(
                type='FocalLoss',
                use_sigmoid=True,
                gamma=2.0,
                alpha=0.25,
                loss_weight=1.0),
            loss_shape=dict(type='BoundedIoULoss', beta=0.2, loss_weight=1.0),
            loss_cls=dict(
                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
            loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
                           loss_weight=1.0),
            init_cfg=dict(type='Normal', layer='Conv2d', std=0.01,
                          override=dict(type='Normal',
                                        name='conv_loc',
                                        std=0.01,
                                        bias_prob=0.01))):  # yapf: disable
        super(AnchorHead, self).__init__(init_cfg)
        self.in_channels = in_channels
        self.num_classes = num_classes
        self.feat_channels = feat_channels
        self.deform_groups = deform_groups
        self.loc_filter_thr = loc_filter_thr

        # build approx_anchor_generator and square_anchor_generator
        assert (approx_anchor_generator['octave_base_scale'] ==
                square_anchor_generator['scales'][0])
        assert (approx_anchor_generator['strides'] ==
                square_anchor_generator['strides'])
        self.approx_anchor_generator = build_prior_generator(
            approx_anchor_generator)
        self.square_anchor_generator = build_prior_generator(
            square_anchor_generator)
        self.approxs_per_octave = self.approx_anchor_generator \
            .num_base_priors[0]

        self.reg_decoded_bbox = reg_decoded_bbox

        # one anchor per location
        self.num_base_priors = self.square_anchor_generator.num_base_priors[0]

        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
        self.loc_focal_loss = loss_loc['type'] in ['FocalLoss']
        self.sampling = loss_cls['type'] not in ['FocalLoss']
        self.ga_sampling = train_cfg is not None and hasattr(
            train_cfg, 'ga_sampler')
        if self.use_sigmoid_cls:
            self.cls_out_channels = self.num_classes
        else:
            self.cls_out_channels = self.num_classes + 1

        # build bbox_coder
        self.anchor_coder = build_bbox_coder(anchor_coder)
        self.bbox_coder = build_bbox_coder(bbox_coder)

        # build losses
        self.loss_loc = build_loss(loss_loc)
        self.loss_shape = build_loss(loss_shape)
        self.loss_cls = build_loss(loss_cls)
        self.loss_bbox = build_loss(loss_bbox)

        self.train_cfg = train_cfg
        self.test_cfg = test_cfg

        if self.train_cfg:
            self.assigner = build_assigner(self.train_cfg.assigner)
            # use PseudoSampler when sampling is False
            if self.sampling and hasattr(self.train_cfg, 'sampler'):
                sampler_cfg = self.train_cfg.sampler
            else:
                sampler_cfg = dict(type='PseudoSampler')
            self.sampler = build_sampler(sampler_cfg, context=self)

            self.ga_assigner = build_assigner(self.train_cfg.ga_assigner)
            if self.ga_sampling:
                ga_sampler_cfg = self.train_cfg.ga_sampler
            else:
                ga_sampler_cfg = dict(type='PseudoSampler')
            self.ga_sampler = build_sampler(ga_sampler_cfg, context=self)

        self.fp16_enabled = False

        self._init_layers()

    @property
    def num_anchors(self):
        warnings.warn('DeprecationWarning: `num_anchors` is deprecated, '
                      'please use "num_base_priors" instead')
        return self.square_anchor_generator.num_base_priors[0]

    def _init_layers(self):
        self.relu = nn.ReLU(inplace=True)
        self.conv_loc = nn.Conv2d(self.in_channels, 1, 1)
        self.conv_shape = nn.Conv2d(self.in_channels, self.num_base_priors * 2,
                                    1)
        self.feature_adaption = FeatureAdaption(
            self.in_channels,
            self.feat_channels,
            kernel_size=3,
            deform_groups=self.deform_groups)
        self.conv_cls = MaskedConv2d(
            self.feat_channels, self.num_base_priors * self.cls_out_channels,
            1)
        self.conv_reg = MaskedConv2d(self.feat_channels,
                                     self.num_base_priors * 4, 1)

    def forward_single(self, x):
        loc_pred = self.conv_loc(x)
        shape_pred = self.conv_shape(x)
        x = self.feature_adaption(x, shape_pred)
        # masked conv is only used during inference for speed-up
        if not self.training:
            mask = loc_pred.sigmoid()[0] >= self.loc_filter_thr
        else:
            mask = None
        cls_score = self.conv_cls(x, mask)
        bbox_pred = self.conv_reg(x, mask)
        return cls_score, bbox_pred, shape_pred, loc_pred

    def forward(self, feats):
        return multi_apply(self.forward_single, feats)

    def get_sampled_approxs(self, featmap_sizes, img_metas, device='cuda'):
        """Get sampled approxs and inside flags according to feature map sizes.

        Args:
            featmap_sizes (list[tuple]): Multi-level feature map sizes.
            img_metas (list[dict]): Image meta info.
            device (torch.device | str): device for returned tensors

        Returns:
            tuple: approxes of each image, inside flags of each image
        """
        num_imgs = len(img_metas)

        # since feature map sizes of all images are the same, we only compute
        # approxes for one time
        multi_level_approxs = self.approx_anchor_generator.grid_priors(
            featmap_sizes, device=device)
        approxs_list = [multi_level_approxs for _ in range(num_imgs)]

        # for each image, we compute inside flags of multi level approxes
        inside_flag_list = []
        for img_id, img_meta in enumerate(img_metas):
            multi_level_flags = []
            multi_level_approxs = approxs_list[img_id]

            # obtain valid flags for each approx first
            multi_level_approx_flags = self.approx_anchor_generator \
                .valid_flags(featmap_sizes,
                             img_meta['pad_shape'],
                             device=device)

            for i, flags in enumerate(multi_level_approx_flags):
                approxs = multi_level_approxs[i]
                inside_flags_list = []
                for i in range(self.approxs_per_octave):
                    split_valid_flags = flags[i::self.approxs_per_octave]
                    split_approxs = approxs[i::self.approxs_per_octave, :]
                    inside_flags = anchor_inside_flags(
                        split_approxs, split_valid_flags,
                        img_meta['img_shape'][:2],
                        self.train_cfg.allowed_border)
                    inside_flags_list.append(inside_flags)
                # inside_flag for a position is true if any anchor in this
                # position is true
                inside_flags = (
                    torch.stack(inside_flags_list, 0).sum(dim=0) > 0)
                multi_level_flags.append(inside_flags)
            inside_flag_list.append(multi_level_flags)
        return approxs_list, inside_flag_list

    def get_anchors(self,
                    featmap_sizes,
                    shape_preds,
                    loc_preds,
                    img_metas,
                    use_loc_filter=False,
                    device='cuda'):
        """Get squares according to feature map sizes and guided anchors.

        Args:
            featmap_sizes (list[tuple]): Multi-level feature map sizes.
            shape_preds (list[tensor]): Multi-level shape predictions.
            loc_preds (list[tensor]): Multi-level location predictions.
            img_metas (list[dict]): Image meta info.
            use_loc_filter (bool): Use loc filter or not.
            device (torch.device | str): device for returned tensors

        Returns:
            tuple: square approxs of each image, guided anchors of each image,
                loc masks of each image
        """
        num_imgs = len(img_metas)
        num_levels = len(featmap_sizes)

        # since feature map sizes of all images are the same, we only compute
        # squares for one time
        multi_level_squares = self.square_anchor_generator.grid_priors(
            featmap_sizes, device=device)
        squares_list = [multi_level_squares for _ in range(num_imgs)]

        # for each image, we compute multi level guided anchors
        guided_anchors_list = []
        loc_mask_list = []
        for img_id, img_meta in enumerate(img_metas):
            multi_level_guided_anchors = []
            multi_level_loc_mask = []
            for i in range(num_levels):
                squares = squares_list[img_id][i]
                shape_pred = shape_preds[i][img_id]
                loc_pred = loc_preds[i][img_id]
                guided_anchors, loc_mask = self._get_guided_anchors_single(
                    squares,
                    shape_pred,
                    loc_pred,
                    use_loc_filter=use_loc_filter)
                multi_level_guided_anchors.append(guided_anchors)
                multi_level_loc_mask.append(loc_mask)
            guided_anchors_list.append(multi_level_guided_anchors)
            loc_mask_list.append(multi_level_loc_mask)
        return squares_list, guided_anchors_list, loc_mask_list

    def _get_guided_anchors_single(self,
                                   squares,
                                   shape_pred,
                                   loc_pred,
                                   use_loc_filter=False):
        """Get guided anchors and loc masks for a single level.

        Args:
            square (tensor): Squares of a single level.
            shape_pred (tensor): Shape predictions of a single level.
            loc_pred (tensor): Loc predictions of a single level.
            use_loc_filter (list[tensor]): Use loc filter or not.

        Returns:
            tuple: guided anchors, location masks
        """
        # calculate location filtering mask
        loc_pred = loc_pred.sigmoid().detach()
        if use_loc_filter:
            loc_mask = loc_pred >= self.loc_filter_thr
        else:
            loc_mask = loc_pred >= 0.0
        mask = loc_mask.permute(1, 2, 0).expand(-1, -1, self.num_base_priors)
        mask = mask.contiguous().view(-1)
        # calculate guided anchors
        squares = squares[mask]
        anchor_deltas = shape_pred.permute(1, 2, 0).contiguous().view(
            -1, 2).detach()[mask]
        bbox_deltas = anchor_deltas.new_full(squares.size(), 0)
        bbox_deltas[:, 2:] = anchor_deltas
        guided_anchors = self.anchor_coder.decode(
            squares, bbox_deltas, wh_ratio_clip=1e-6)
        return guided_anchors, mask

    def ga_loc_targets(self, gt_bboxes_list, featmap_sizes):
        """Compute location targets for guided anchoring.

        Each feature map is divided into positive, negative and ignore regions.
        - positive regions: target 1, weight 1
        - ignore regions: target 0, weight 0
        - negative regions: target 0, weight 0.1

        Args:
            gt_bboxes_list (list[Tensor]): Gt bboxes of each image.
            featmap_sizes (list[tuple]): Multi level sizes of each feature
                maps.

        Returns:
            tuple
        """
        anchor_scale = self.approx_anchor_generator.octave_base_scale
        anchor_strides = self.approx_anchor_generator.strides
        # Currently only supports same stride in x and y direction.
        for stride in anchor_strides:
            assert (stride[0] == stride[1])
        anchor_strides = [stride[0] for stride in anchor_strides]

        center_ratio = self.train_cfg.center_ratio
        ignore_ratio = self.train_cfg.ignore_ratio
        img_per_gpu = len(gt_bboxes_list)
        num_lvls = len(featmap_sizes)
        r1 = (1 - center_ratio) / 2
        r2 = (1 - ignore_ratio) / 2
        all_loc_targets = []
        all_loc_weights = []
        all_ignore_map = []
        for lvl_id in range(num_lvls):
            h, w = featmap_sizes[lvl_id]
            loc_targets = torch.zeros(
                img_per_gpu,
                1,
                h,
                w,
                device=gt_bboxes_list[0].device,
                dtype=torch.float32)
            loc_weights = torch.full_like(loc_targets, -1)
            ignore_map = torch.zeros_like(loc_targets)
            all_loc_targets.append(loc_targets)
            all_loc_weights.append(loc_weights)
            all_ignore_map.append(ignore_map)
        for img_id in range(img_per_gpu):
            gt_bboxes = gt_bboxes_list[img_id]
            scale = torch.sqrt((gt_bboxes[:, 2] - gt_bboxes[:, 0]) *
                               (gt_bboxes[:, 3] - gt_bboxes[:, 1]))
            min_anchor_size = scale.new_full(
                (1, ), float(anchor_scale * anchor_strides[0]))
            # assign gt bboxes to different feature levels w.r.t. their scales
            target_lvls = torch.floor(
                torch.log2(scale) - torch.log2(min_anchor_size) + 0.5)
            target_lvls = target_lvls.clamp(min=0, max=num_lvls - 1).long()
            for gt_id in range(gt_bboxes.size(0)):
                lvl = target_lvls[gt_id].item()
                # rescaled to corresponding feature map
                gt_ = gt_bboxes[gt_id, :4] / anchor_strides[lvl]
                # calculate ignore regions
                ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
                    gt_, r2, featmap_sizes[lvl])
                # calculate positive (center) regions
                ctr_x1, ctr_y1, ctr_x2, ctr_y2 = calc_region(
                    gt_, r1, featmap_sizes[lvl])
                all_loc_targets[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
                                     ctr_x1:ctr_x2 + 1] = 1
                all_loc_weights[lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
                                     ignore_x1:ignore_x2 + 1] = 0
                all_loc_weights[lvl][img_id, 0, ctr_y1:ctr_y2 + 1,
                                     ctr_x1:ctr_x2 + 1] = 1
                # calculate ignore map on nearby low level feature
                if lvl > 0:
                    d_lvl = lvl - 1
                    # rescaled to corresponding feature map
                    gt_ = gt_bboxes[gt_id, :4] / anchor_strides[d_lvl]
                    ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
                        gt_, r2, featmap_sizes[d_lvl])
                    all_ignore_map[d_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
                                          ignore_x1:ignore_x2 + 1] = 1
                # calculate ignore map on nearby high level feature
                if lvl < num_lvls - 1:
                    u_lvl = lvl + 1
                    # rescaled to corresponding feature map
                    gt_ = gt_bboxes[gt_id, :4] / anchor_strides[u_lvl]
                    ignore_x1, ignore_y1, ignore_x2, ignore_y2 = calc_region(
                        gt_, r2, featmap_sizes[u_lvl])
                    all_ignore_map[u_lvl][img_id, 0, ignore_y1:ignore_y2 + 1,
                                          ignore_x1:ignore_x2 + 1] = 1
        for lvl_id in range(num_lvls):
            # ignore negative regions w.r.t. ignore map
            all_loc_weights[lvl_id][(all_loc_weights[lvl_id] < 0)
                                    & (all_ignore_map[lvl_id] > 0)] = 0
            # set negative regions with weight 0.1
            all_loc_weights[lvl_id][all_loc_weights[lvl_id] < 0] = 0.1
        # loc average factor to balance loss
        loc_avg_factor = sum(
            [t.size(0) * t.size(-1) * t.size(-2)
             for t in all_loc_targets]) / 200
        return all_loc_targets, all_loc_weights, loc_avg_factor

    def _ga_shape_target_single(self,
                                flat_approxs,
                                inside_flags,
                                flat_squares,
                                gt_bboxes,
                                gt_bboxes_ignore,
                                img_meta,
                                unmap_outputs=True):
        """Compute guided anchoring targets.

        This function returns sampled anchors and gt bboxes directly
        rather than calculates regression targets.

        Args:
            flat_approxs (Tensor): flat approxs of a single image,
                shape (n, 4)
            inside_flags (Tensor): inside flags of a single image,
                shape (n, ).
            flat_squares (Tensor): flat squares of a single image,
                shape (approxs_per_octave * n, 4)
            gt_bboxes (Tensor): Ground truth bboxes of a single image.
            img_meta (dict): Meta info of a single image.
            approxs_per_octave (int): number of approxs per octave
            cfg (dict): RPN train configs.
            unmap_outputs (bool): unmap outputs or not.

        Returns:
            tuple
        """
        if not inside_flags.any():
            return (None, ) * 5
        # assign gt and sample anchors
        expand_inside_flags = inside_flags[:, None].expand(
            -1, self.approxs_per_octave).reshape(-1)
        approxs = flat_approxs[expand_inside_flags, :]
        squares = flat_squares[inside_flags, :]

        assign_result = self.ga_assigner.assign(approxs, squares,
                                                self.approxs_per_octave,
                                                gt_bboxes, gt_bboxes_ignore)
        sampling_result = self.ga_sampler.sample(assign_result, squares,
                                                 gt_bboxes)

        bbox_anchors = torch.zeros_like(squares)
        bbox_gts = torch.zeros_like(squares)
        bbox_weights = torch.zeros_like(squares)

        pos_inds = sampling_result.pos_inds
        neg_inds = sampling_result.neg_inds
        if len(pos_inds) > 0:
            bbox_anchors[pos_inds, :] = sampling_result.pos_bboxes
            bbox_gts[pos_inds, :] = sampling_result.pos_gt_bboxes
            bbox_weights[pos_inds, :] = 1.0

        # map up to original set of anchors
        if unmap_outputs:
            num_total_anchors = flat_squares.size(0)
            bbox_anchors = unmap(bbox_anchors, num_total_anchors, inside_flags)
            bbox_gts = unmap(bbox_gts, num_total_anchors, inside_flags)
            bbox_weights = unmap(bbox_weights, num_total_anchors, inside_flags)

        return (bbox_anchors, bbox_gts, bbox_weights, pos_inds, neg_inds)

    def ga_shape_targets(self,
                         approx_list,
                         inside_flag_list,
                         square_list,
                         gt_bboxes_list,
                         img_metas,
                         gt_bboxes_ignore_list=None,
                         unmap_outputs=True):
        """Compute guided anchoring targets.

        Args:
            approx_list (list[list]): Multi level approxs of each image.
            inside_flag_list (list[list]): Multi level inside flags of each
                image.
            square_list (list[list]): Multi level squares of each image.
            gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.
            img_metas (list[dict]): Meta info of each image.
            gt_bboxes_ignore_list (list[Tensor]): ignore list of gt bboxes.
            unmap_outputs (bool): unmap outputs or not.

        Returns:
            tuple
        """
        num_imgs = len(img_metas)
        assert len(approx_list) == len(inside_flag_list) == len(
            square_list) == num_imgs
        # anchor number of multi levels
        num_level_squares = [squares.size(0) for squares in square_list[0]]
        # concat all level anchors and flags to a single tensor
        inside_flag_flat_list = []
        approx_flat_list = []
        square_flat_list = []
        for i in range(num_imgs):
            assert len(square_list[i]) == len(inside_flag_list[i])
            inside_flag_flat_list.append(torch.cat(inside_flag_list[i]))
            approx_flat_list.append(torch.cat(approx_list[i]))
            square_flat_list.append(torch.cat(square_list[i]))

        # compute targets for each image
        if gt_bboxes_ignore_list is None:
            gt_bboxes_ignore_list = [None for _ in range(num_imgs)]
        (all_bbox_anchors, all_bbox_gts, all_bbox_weights, pos_inds_list,
         neg_inds_list) = multi_apply(
             self._ga_shape_target_single,
             approx_flat_list,
             inside_flag_flat_list,
             square_flat_list,
             gt_bboxes_list,
             gt_bboxes_ignore_list,
             img_metas,
             unmap_outputs=unmap_outputs)
        # no valid anchors
        if any([bbox_anchors is None for bbox_anchors in all_bbox_anchors]):
            return None
        # sampled anchors of all images
        num_total_pos = sum([max(inds.numel(), 1) for inds in pos_inds_list])
        num_total_neg = sum([max(inds.numel(), 1) for inds in neg_inds_list])
        # split targets to a list w.r.t. multiple levels
        bbox_anchors_list = images_to_levels(all_bbox_anchors,
                                             num_level_squares)
        bbox_gts_list = images_to_levels(all_bbox_gts, num_level_squares)
        bbox_weights_list = images_to_levels(all_bbox_weights,
                                             num_level_squares)
        return (bbox_anchors_list, bbox_gts_list, bbox_weights_list,
                num_total_pos, num_total_neg)

    def loss_shape_single(self, shape_pred, bbox_anchors, bbox_gts,
                          anchor_weights, anchor_total_num):
        shape_pred = shape_pred.permute(0, 2, 3, 1).contiguous().view(-1, 2)
        bbox_anchors = bbox_anchors.contiguous().view(-1, 4)
        bbox_gts = bbox_gts.contiguous().view(-1, 4)
        anchor_weights = anchor_weights.contiguous().view(-1, 4)
        bbox_deltas = bbox_anchors.new_full(bbox_anchors.size(), 0)
        bbox_deltas[:, 2:] += shape_pred
        # filter out negative samples to speed-up weighted_bounded_iou_loss
        inds = torch.nonzero(
            anchor_weights[:, 0] > 0, as_tuple=False).squeeze(1)
        bbox_deltas_ = bbox_deltas[inds]
        bbox_anchors_ = bbox_anchors[inds]
        bbox_gts_ = bbox_gts[inds]
        anchor_weights_ = anchor_weights[inds]
        pred_anchors_ = self.anchor_coder.decode(
            bbox_anchors_, bbox_deltas_, wh_ratio_clip=1e-6)
        loss_shape = self.loss_shape(
            pred_anchors_,
            bbox_gts_,
            anchor_weights_,
            avg_factor=anchor_total_num)
        return loss_shape

    def loss_loc_single(self, loc_pred, loc_target, loc_weight,
                        loc_avg_factor):
        loss_loc = self.loss_loc(
            loc_pred.reshape(-1, 1),
            loc_target.reshape(-1).long(),
            loc_weight.reshape(-1),
            avg_factor=loc_avg_factor)
        return loss_loc

    @force_fp32(
        apply_to=('cls_scores', 'bbox_preds', 'shape_preds', 'loc_preds'))
    def loss(self,
             cls_scores,
             bbox_preds,
             shape_preds,
             loc_preds,
             gt_bboxes,
             gt_labels,
             img_metas,
             gt_bboxes_ignore=None):
        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
        assert len(featmap_sizes) == self.approx_anchor_generator.num_levels

        device = cls_scores[0].device

        # get loc targets
        loc_targets, loc_weights, loc_avg_factor = self.ga_loc_targets(
            gt_bboxes, featmap_sizes)

        # get sampled approxes
        approxs_list, inside_flag_list = self.get_sampled_approxs(
            featmap_sizes, img_metas, device=device)
        # get squares and guided anchors
        squares_list, guided_anchors_list, _ = self.get_anchors(
            featmap_sizes, shape_preds, loc_preds, img_metas, device=device)

        # get shape targets
        shape_targets = self.ga_shape_targets(approxs_list, inside_flag_list,
                                              squares_list, gt_bboxes,
                                              img_metas)
        if shape_targets is None:
            return None
        (bbox_anchors_list, bbox_gts_list, anchor_weights_list, anchor_fg_num,
         anchor_bg_num) = shape_targets
        anchor_total_num = (
            anchor_fg_num if not self.ga_sampling else anchor_fg_num +
            anchor_bg_num)

        # get anchor targets
        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
        cls_reg_targets = self.get_targets(
            guided_anchors_list,
            inside_flag_list,
            gt_bboxes,
            img_metas,
            gt_bboxes_ignore_list=gt_bboxes_ignore,
            gt_labels_list=gt_labels,
            label_channels=label_channels)
        if cls_reg_targets is None:
            return None
        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
         num_total_pos, num_total_neg) = cls_reg_targets
        num_total_samples = (
            num_total_pos + num_total_neg if self.sampling else num_total_pos)

        # anchor number of multi levels
        num_level_anchors = [
            anchors.size(0) for anchors in guided_anchors_list[0]
        ]
        # concat all level anchors to a single tensor
        concat_anchor_list = []
        for i in range(len(guided_anchors_list)):
            concat_anchor_list.append(torch.cat(guided_anchors_list[i]))
        all_anchor_list = images_to_levels(concat_anchor_list,
                                           num_level_anchors)

        # get classification and bbox regression losses
        losses_cls, losses_bbox = multi_apply(
            self.loss_single,
            cls_scores,
            bbox_preds,
            all_anchor_list,
            labels_list,
            label_weights_list,
            bbox_targets_list,
            bbox_weights_list,
            num_total_samples=num_total_samples)

        # get anchor location loss
        losses_loc = []
        for i in range(len(loc_preds)):
            loss_loc = self.loss_loc_single(
                loc_preds[i],
                loc_targets[i],
                loc_weights[i],
                loc_avg_factor=loc_avg_factor)
            losses_loc.append(loss_loc)

        # get anchor shape loss
        losses_shape = []
        for i in range(len(shape_preds)):
            loss_shape = self.loss_shape_single(
                shape_preds[i],
                bbox_anchors_list[i],
                bbox_gts_list[i],
                anchor_weights_list[i],
                anchor_total_num=anchor_total_num)
            losses_shape.append(loss_shape)

        return dict(
            loss_cls=losses_cls,
            loss_bbox=losses_bbox,
            loss_shape=losses_shape,
            loss_loc=losses_loc)

    @force_fp32(
        apply_to=('cls_scores', 'bbox_preds', 'shape_preds', 'loc_preds'))
    def get_bboxes(self,
                   cls_scores,
                   bbox_preds,
                   shape_preds,
                   loc_preds,
                   img_metas,
                   cfg=None,
                   rescale=False):
        assert len(cls_scores) == len(bbox_preds) == len(shape_preds) == len(
            loc_preds)
        num_levels = len(cls_scores)
        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
        device = cls_scores[0].device
        # get guided anchors
        _, guided_anchors, loc_masks = self.get_anchors(
            featmap_sizes,
            shape_preds,
            loc_preds,
            img_metas,
            use_loc_filter=not self.training,
            device=device)
        result_list = []
        for img_id in range(len(img_metas)):
            cls_score_list = [
                cls_scores[i][img_id].detach() for i in range(num_levels)
            ]
            bbox_pred_list = [
                bbox_preds[i][img_id].detach() for i in range(num_levels)
            ]
            guided_anchor_list = [
                guided_anchors[img_id][i].detach() for i in range(num_levels)
            ]
            loc_mask_list = [
                loc_masks[img_id][i].detach() for i in range(num_levels)
            ]
            img_shape = img_metas[img_id]['img_shape']
            scale_factor = img_metas[img_id]['scale_factor']
            proposals = self._get_bboxes_single(cls_score_list, bbox_pred_list,
                                                guided_anchor_list,
                                                loc_mask_list, img_shape,
                                                scale_factor, cfg, rescale)
            result_list.append(proposals)
        return result_list

    def _get_bboxes_single(self,
                           cls_scores,
                           bbox_preds,
                           mlvl_anchors,
                           mlvl_masks,
                           img_shape,
                           scale_factor,
                           cfg,
                           rescale=False):
        cfg = self.test_cfg if cfg is None else cfg
        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
        mlvl_bboxes = []
        mlvl_scores = []
        for cls_score, bbox_pred, anchors, mask in zip(cls_scores, bbox_preds,
                                                       mlvl_anchors,
                                                       mlvl_masks):
            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
            # if no location is kept, end.
            if mask.sum() == 0:
                continue
            # reshape scores and bbox_pred
            cls_score = cls_score.permute(1, 2,
                                          0).reshape(-1, self.cls_out_channels)
            if self.use_sigmoid_cls:
                scores = cls_score.sigmoid()
            else:
                scores = cls_score.softmax(-1)
            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
            # filter scores, bbox_pred w.r.t. mask.
            # anchors are filtered in get_anchors() beforehand.
            scores = scores[mask, :]
            bbox_pred = bbox_pred[mask, :]
            if scores.dim() == 0:
                anchors = anchors.unsqueeze(0)
                scores = scores.unsqueeze(0)
                bbox_pred = bbox_pred.unsqueeze(0)
            # filter anchors, bbox_pred, scores w.r.t. scores
            nms_pre = cfg.get('nms_pre', -1)
            if nms_pre > 0 and scores.shape[0] > nms_pre:
                if self.use_sigmoid_cls:
                    max_scores, _ = scores.max(dim=1)
                else:
                    # remind that we set FG labels to [0, num_class-1]
                    # since mmdet v2.0
                    # BG cat_id: num_class
                    max_scores, _ = scores[:, :-1].max(dim=1)
                _, topk_inds = max_scores.topk(nms_pre)
                anchors = anchors[topk_inds, :]
                bbox_pred = bbox_pred[topk_inds, :]
                scores = scores[topk_inds, :]
            bboxes = self.bbox_coder.decode(
                anchors, bbox_pred, max_shape=img_shape)
            mlvl_bboxes.append(bboxes)
            mlvl_scores.append(scores)
        mlvl_bboxes = torch.cat(mlvl_bboxes)
        if rescale:
            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
        mlvl_scores = torch.cat(mlvl_scores)
        if self.use_sigmoid_cls:
            # Add a dummy background class to the backend when using sigmoid
            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
            # BG cat_id: num_class
            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
        # multi class NMS
        det_bboxes, det_labels = multiclass_nms(mlvl_bboxes, mlvl_scores,
                                                cfg.score_thr, cfg.nms,
                                                cfg.max_per_img)
        return det_bboxes, det_labels