from typing import ValuesView import torch.nn as nn import torch import numpy as np import copy import torch.nn.functional as F from pcdet.ops.iou3d_nms import iou3d_nms_utils from ...utils import common_utils, loss_utils from .roi_head_template import RoIHeadTemplate from ..model_utils.mppnet_utils import build_transformer, PointNet, MLP from .target_assigner.proposal_target_layer import ProposalTargetLayer from pcdet.ops.pointnet2.pointnet2_stack import pointnet2_modules as pointnet2_stack_modules class ProposalTargetLayerMPPNet(ProposalTargetLayer): def __init__(self, roi_sampler_cfg): super().__init__(roi_sampler_cfg = roi_sampler_cfg) def forward(self, batch_dict): """ Args: batch_dict: batch_size: rois: (B, num_rois, 7 + C) roi_scores: (B, num_rois) gt_boxes: (B, N, 7 + C + 1) roi_labels: (B, num_rois) Returns: batch_dict: rois: (B, M, 7 + C) gt_of_rois: (B, M, 7 + C) gt_iou_of_rois: (B, M) roi_scores: (B, M) roi_labels: (B, M) reg_valid_mask: (B, M) rcnn_cls_labels: (B, M) """ batch_rois, batch_gt_of_rois, batch_roi_ious, batch_roi_scores, batch_roi_labels, \ batch_trajectory_rois,batch_valid_length = self.sample_rois_for_mppnet(batch_dict=batch_dict) # regression valid mask reg_valid_mask = (batch_roi_ious > self.roi_sampler_cfg.REG_FG_THRESH).long() # classification label if self.roi_sampler_cfg.CLS_SCORE_TYPE == 'cls': batch_cls_labels = (batch_roi_ious > self.roi_sampler_cfg.CLS_FG_THRESH).long() ignore_mask = (batch_roi_ious > self.roi_sampler_cfg.CLS_BG_THRESH) & \ (batch_roi_ious < self.roi_sampler_cfg.CLS_FG_THRESH) batch_cls_labels[ignore_mask > 0] = -1 elif self.roi_sampler_cfg.CLS_SCORE_TYPE == 'roi_iou': iou_bg_thresh = self.roi_sampler_cfg.CLS_BG_THRESH iou_fg_thresh = self.roi_sampler_cfg.CLS_FG_THRESH fg_mask = batch_roi_ious > iou_fg_thresh bg_mask = batch_roi_ious < iou_bg_thresh interval_mask = (fg_mask == 0) & (bg_mask == 0) batch_cls_labels = (fg_mask > 0).float() batch_cls_labels[interval_mask] = \ (batch_roi_ious[interval_mask] - iou_bg_thresh) / (iou_fg_thresh - iou_bg_thresh) else: raise NotImplementedError targets_dict = {'rois': batch_rois, 'gt_of_rois': batch_gt_of_rois, 'gt_iou_of_rois': batch_roi_ious,'roi_scores': batch_roi_scores, 'roi_labels': batch_roi_labels,'reg_valid_mask': reg_valid_mask, 'rcnn_cls_labels': batch_cls_labels,'trajectory_rois':batch_trajectory_rois, 'valid_length': batch_valid_length, } return targets_dict def sample_rois_for_mppnet(self, batch_dict): """ Args: batch_dict: batch_size: rois: (B, num_rois, 7 + C) roi_scores: (B, num_rois) gt_boxes: (B, N, 7 + C + 1) roi_labels: (B, num_rois) Returns: """ cur_frame_idx = 0 batch_size = batch_dict['batch_size'] rois = batch_dict['trajectory_rois'][:,cur_frame_idx,:,:] roi_scores = batch_dict['roi_scores'][:,:,cur_frame_idx] roi_labels = batch_dict['roi_labels'] gt_boxes = batch_dict['gt_boxes'] code_size = rois.shape[-1] batch_rois = rois.new_zeros(batch_size, self.roi_sampler_cfg.ROI_PER_IMAGE, code_size) batch_gt_of_rois = rois.new_zeros(batch_size, self.roi_sampler_cfg.ROI_PER_IMAGE, gt_boxes.shape[-1]) batch_roi_ious = rois.new_zeros(batch_size, self.roi_sampler_cfg.ROI_PER_IMAGE) batch_roi_scores = rois.new_zeros(batch_size, self.roi_sampler_cfg.ROI_PER_IMAGE) batch_roi_labels = rois.new_zeros((batch_size, self.roi_sampler_cfg.ROI_PER_IMAGE), dtype=torch.long) trajectory_rois = batch_dict['trajectory_rois'] batch_trajectory_rois = rois.new_zeros(batch_size, trajectory_rois.shape[1],self.roi_sampler_cfg.ROI_PER_IMAGE,trajectory_rois.shape[-1]) valid_length = batch_dict['valid_length'] batch_valid_length = rois.new_zeros((batch_size, batch_dict['trajectory_rois'].shape[1], self.roi_sampler_cfg.ROI_PER_IMAGE)) for index in range(batch_size): cur_trajectory_rois = trajectory_rois[index] cur_roi, cur_gt, cur_roi_labels, cur_roi_scores = rois[index],gt_boxes[index], roi_labels[index], roi_scores[index] if 'valid_length' in batch_dict.keys(): cur_valid_length = valid_length[index] k = cur_gt.__len__() - 1 while k > 0 and cur_gt[k].sum() == 0: k -= 1 cur_gt = cur_gt[:k + 1] cur_gt = cur_gt.new_zeros((1, cur_gt.shape[1])) if len(cur_gt) == 0 else cur_gt if self.roi_sampler_cfg.get('SAMPLE_ROI_BY_EACH_CLASS', False): max_overlaps, gt_assignment = self.get_max_iou_with_same_class( rois=cur_roi, roi_labels=cur_roi_labels, gt_boxes=cur_gt[:, 0:7], gt_labels=cur_gt[:, -1].long() ) else: iou3d = iou3d_nms_utils.boxes_iou3d_gpu(cur_roi, cur_gt[:, 0:7]) # (M, N) max_overlaps, gt_assignment = torch.max(iou3d, dim=1) sampled_inds,fg_inds, bg_inds = self.subsample_rois(max_overlaps=max_overlaps) batch_roi_labels[index] = cur_roi_labels[sampled_inds.long()] if self.roi_sampler_cfg.get('USE_ROI_AUG',False): fg_rois, fg_iou3d = self.aug_roi_by_noise_torch(cur_roi[fg_inds], cur_gt[gt_assignment[fg_inds]], max_overlaps[fg_inds], aug_times=self.roi_sampler_cfg.ROI_FG_AUG_TIMES) bg_rois = cur_roi[bg_inds] bg_iou3d = max_overlaps[bg_inds] batch_rois[index] = torch.cat([fg_rois,bg_rois],0) batch_roi_ious[index] = torch.cat([fg_iou3d,bg_iou3d],0) batch_gt_of_rois[index] = cur_gt[gt_assignment[sampled_inds]] else: batch_rois[index] = cur_roi[sampled_inds] batch_roi_ious[index] = max_overlaps[sampled_inds] batch_gt_of_rois[index] = cur_gt[gt_assignment[sampled_inds]] batch_roi_scores[index] = cur_roi_scores[sampled_inds] if 'valid_length' in batch_dict.keys(): batch_valid_length[index] = cur_valid_length[:,sampled_inds] if self.roi_sampler_cfg.USE_TRAJ_AUG.ENABLED: batch_trajectory_rois_list = [] for idx in range(0,batch_dict['num_frames']): if idx== cur_frame_idx: batch_trajectory_rois_list.append(cur_trajectory_rois[cur_frame_idx:cur_frame_idx+1,sampled_inds]) continue fg_trajs, _ = self.aug_roi_by_noise_torch(cur_trajectory_rois[idx,fg_inds], cur_trajectory_rois[idx,fg_inds][:,:8], max_overlaps[fg_inds], \ aug_times=self.roi_sampler_cfg.ROI_FG_AUG_TIMES,pos_thresh=self.roi_sampler_cfg.USE_TRAJ_AUG.THRESHOD) bg_trajs = cur_trajectory_rois[idx,bg_inds] batch_trajectory_rois_list.append(torch.cat([fg_trajs,bg_trajs],0)[None,:,:]) batch_trajectory_rois[index] = torch.cat(batch_trajectory_rois_list,0) else: batch_trajectory_rois[index] = cur_trajectory_rois[:,sampled_inds] return batch_rois, batch_gt_of_rois, batch_roi_ious, batch_roi_scores, batch_roi_labels, batch_trajectory_rois,batch_valid_length def subsample_rois(self, max_overlaps): # sample fg, easy_bg, hard_bg fg_rois_per_image = int(np.round(self.roi_sampler_cfg.FG_RATIO * self.roi_sampler_cfg.ROI_PER_IMAGE)) fg_thresh = min(self.roi_sampler_cfg.REG_FG_THRESH, self.roi_sampler_cfg.CLS_FG_THRESH) fg_inds = ((max_overlaps >= fg_thresh)).nonzero().view(-1) easy_bg_inds = ((max_overlaps < self.roi_sampler_cfg.CLS_BG_THRESH_LO)).nonzero().view(-1) hard_bg_inds = ((max_overlaps < self.roi_sampler_cfg.REG_FG_THRESH) & (max_overlaps >= self.roi_sampler_cfg.CLS_BG_THRESH_LO)).nonzero().view(-1) fg_num_rois = fg_inds.numel() bg_num_rois = hard_bg_inds.numel() + easy_bg_inds.numel() if fg_num_rois > 0 and bg_num_rois > 0: # sampling fg fg_rois_per_this_image = min(fg_rois_per_image, fg_num_rois) rand_num = torch.from_numpy(np.random.permutation(fg_num_rois)).type_as(max_overlaps).long() fg_inds = fg_inds[rand_num[:fg_rois_per_this_image]] # sampling bg bg_rois_per_this_image = self.roi_sampler_cfg.ROI_PER_IMAGE - fg_rois_per_this_image bg_inds = self.sample_bg_inds( hard_bg_inds, easy_bg_inds, bg_rois_per_this_image, self.roi_sampler_cfg.HARD_BG_RATIO ) elif fg_num_rois > 0 and bg_num_rois == 0: # sampling fg rand_num = np.floor(np.random.rand(self.roi_sampler_cfg.ROI_PER_IMAGE) * fg_num_rois) rand_num = torch.from_numpy(rand_num).type_as(max_overlaps).long() fg_inds = fg_inds[rand_num] bg_inds = torch.tensor([]).type_as(fg_inds) elif bg_num_rois > 0 and fg_num_rois == 0: # sampling bg bg_rois_per_this_image = self.roi_sampler_cfg.ROI_PER_IMAGE bg_inds = self.sample_bg_inds( hard_bg_inds, easy_bg_inds, bg_rois_per_this_image, self.roi_sampler_cfg.HARD_BG_RATIO ) else: print('maxoverlaps:(min=%f, max=%f)' % (max_overlaps.min().item(), max_overlaps.max().item())) print('ERROR: FG=%d, BG=%d' % (fg_num_rois, bg_num_rois)) raise NotImplementedError sampled_inds = torch.cat((fg_inds, bg_inds), dim=0) return sampled_inds.long(), fg_inds.long(), bg_inds.long() def aug_roi_by_noise_torch(self,roi_boxes3d, gt_boxes3d, iou3d_src, aug_times=10, pos_thresh=None): iou_of_rois = torch.zeros(roi_boxes3d.shape[0]).type_as(gt_boxes3d) if pos_thresh is None: pos_thresh = min(self.roi_sampler_cfg.REG_FG_THRESH, self.roi_sampler_cfg.CLS_FG_THRESH) for k in range(roi_boxes3d.shape[0]): temp_iou = cnt = 0 roi_box3d = roi_boxes3d[k] gt_box3d = gt_boxes3d[k].view(1, gt_boxes3d.shape[-1]) aug_box3d = roi_box3d keep = True while temp_iou < pos_thresh and cnt < aug_times: if np.random.rand() <= self.roi_sampler_cfg.RATIO: aug_box3d = roi_box3d # p=RATIO to keep the original roi box keep = True else: aug_box3d = self.random_aug_box3d(roi_box3d) keep = False aug_box3d = aug_box3d.view((1, aug_box3d.shape[-1])) iou3d = iou3d_nms_utils.boxes_iou3d_gpu(aug_box3d[:,:7], gt_box3d[:,:7]) temp_iou = iou3d[0][0] cnt += 1 roi_boxes3d[k] = aug_box3d.view(-1) if cnt == 0 or keep: iou_of_rois[k] = iou3d_src[k] else: iou_of_rois[k] = temp_iou return roi_boxes3d, iou_of_rois def random_aug_box3d(self,box3d): """ :param box3d: (7) [x, y, z, h, w, l, ry] random shift, scale, orientation """ if self.roi_sampler_cfg.REG_AUG_METHOD == 'single': pos_shift = (torch.rand(3, device=box3d.device) - 0.5) # [-0.5 ~ 0.5] hwl_scale = (torch.rand(3, device=box3d.device) - 0.5) / (0.5 / 0.15) + 1.0 # angle_rot = (torch.rand(1, device=box3d.device) - 0.5) / (0.5 / (np.pi / 12)) # [-pi/12 ~ pi/12] aug_box3d = torch.cat([box3d[0:3] + pos_shift, box3d[3:6] * hwl_scale, box3d[6:7] + angle_rot, box3d[7:]], dim=0) return aug_box3d elif self.roi_sampler_cfg.REG_AUG_METHOD == 'multiple': # pos_range, hwl_range, angle_range, mean_iou range_config = [[0.2, 0.1, np.pi / 12, 0.7], [0.3, 0.15, np.pi / 12, 0.6], [0.5, 0.15, np.pi / 9, 0.5], [0.8, 0.15, np.pi / 6, 0.3], [1.0, 0.15, np.pi / 3, 0.2]] idx = torch.randint(low=0, high=len(range_config), size=(1,))[0].long() pos_shift = ((torch.rand(3, device=box3d.device) - 0.5) / 0.5) * range_config[idx][0] hwl_scale = ((torch.rand(3, device=box3d.device) - 0.5) / 0.5) * range_config[idx][1] + 1.0 angle_rot = ((torch.rand(1, device=box3d.device) - 0.5) / 0.5) * range_config[idx][2] aug_box3d = torch.cat([box3d[0:3] + pos_shift, box3d[3:6] * hwl_scale, box3d[6:7] + angle_rot], dim=0) return aug_box3d elif self.roi_sampler_cfg.REG_AUG_METHOD == 'normal': x_shift = np.random.normal(loc=0, scale=0.3) y_shift = np.random.normal(loc=0, scale=0.2) z_shift = np.random.normal(loc=0, scale=0.3) h_shift = np.random.normal(loc=0, scale=0.25) w_shift = np.random.normal(loc=0, scale=0.15) l_shift = np.random.normal(loc=0, scale=0.5) ry_shift = ((torch.rand() - 0.5) / 0.5) * np.pi / 12 aug_box3d = np.array([box3d[0] + x_shift, box3d[1] + y_shift, box3d[2] + z_shift, box3d[3] + h_shift, box3d[4] + w_shift, box3d[5] + l_shift, box3d[6] + ry_shift], dtype=np.float32) aug_box3d = torch.from_numpy(aug_box3d).type_as(box3d) return aug_box3d else: raise NotImplementedError class MPPNetHead(RoIHeadTemplate): def __init__(self,model_cfg, num_class=1,**kwargs): super().__init__(num_class=num_class, model_cfg=model_cfg) self.model_cfg = model_cfg self.proposal_target_layer = ProposalTargetLayerMPPNet(roi_sampler_cfg=self.model_cfg.TARGET_CONFIG) self.use_time_stamp = self.model_cfg.get('USE_TIMESTAMP',None) self.num_lidar_points = self.model_cfg.Transformer.num_lidar_points self.avg_stage1_score = self.model_cfg.get('AVG_STAGE1_SCORE', None) self.nhead = model_cfg.Transformer.nheads self.num_enc_layer = model_cfg.Transformer.enc_layers hidden_dim = model_cfg.TRANS_INPUT self.hidden_dim = model_cfg.TRANS_INPUT self.num_groups = model_cfg.Transformer.num_groups self.grid_size = model_cfg.ROI_GRID_POOL.GRID_SIZE self.num_proxy_points = model_cfg.Transformer.num_proxy_points self.seqboxembed = PointNet(8,model_cfg=self.model_cfg) self.jointembed = MLP(self.hidden_dim*(self.num_groups+1), model_cfg.Transformer.hidden_dim, self.box_coder.code_size * self.num_class, 4) num_radius = len(self.model_cfg.ROI_GRID_POOL.POOL_RADIUS) self.up_dimension_geometry = MLP(input_dim = 29, hidden_dim = 64, output_dim =hidden_dim//num_radius, num_layers = 3) self.up_dimension_motion = MLP(input_dim = 30, hidden_dim = 64, output_dim = hidden_dim, num_layers = 3) self.transformer = build_transformer(model_cfg.Transformer) self.roi_grid_pool_layer = pointnet2_stack_modules.StackSAModuleMSG( radii=self.model_cfg.ROI_GRID_POOL.POOL_RADIUS, nsamples=self.model_cfg.ROI_GRID_POOL.NSAMPLE, mlps=self.model_cfg.ROI_GRID_POOL.MLPS, use_xyz=True, pool_method=self.model_cfg.ROI_GRID_POOL.POOL_METHOD, ) self.class_embed = nn.ModuleList() self.class_embed.append(nn.Linear(model_cfg.Transformer.hidden_dim, 1)) self.bbox_embed = nn.ModuleList() for _ in range(self.num_groups): self.bbox_embed.append(MLP(model_cfg.Transformer.hidden_dim, model_cfg.Transformer.hidden_dim, self.box_coder.code_size * self.num_class, 4)) if self.model_cfg.Transformer.use_grid_pos.enabled: if self.model_cfg.Transformer.use_grid_pos.init_type == 'index': self.grid_index = torch.cat([i.reshape(-1,1)for i in torch.meshgrid(torch.arange(self.grid_size), torch.arange(self.grid_size), torch.arange(self.grid_size))],1).float().cuda() self.grid_pos_embeded = MLP(input_dim = 3, hidden_dim = 256, output_dim = hidden_dim, num_layers = 2) else: self.pos = nn.Parameter(torch.zeros(1, self.num_grid_points, 256)) def init_weights(self, weight_init='xavier'): if weight_init == 'kaiming': init_func = nn.init.kaiming_normal_ elif weight_init == 'xavier': init_func = nn.init.xavier_normal_ elif weight_init == 'normal': init_func = nn.init.normal_ else: raise NotImplementedError for m in self.modules(): if isinstance(m, nn.Conv2d) or isinstance(m, nn.Conv1d): if weight_init == 'normal': init_func(m.weight, mean=0, std=0.001) else: init_func(m.weight) if m.bias is not None: nn.init.constant_(m.bias, 0) nn.init.normal_(self.bbox_embed.layers[-1].weight, mean=0, std=0.001) def get_corner_points_of_roi(self, rois): rois = rois.view(-1, rois.shape[-1]) batch_size_rcnn = rois.shape[0] local_roi_grid_points = self.get_corner_points(rois, batch_size_rcnn) local_roi_grid_points = common_utils.rotate_points_along_z( local_roi_grid_points.clone(), rois[:, 6] ).squeeze(dim=1) global_center = rois[:, 0:3].clone() global_roi_grid_points = local_roi_grid_points + global_center.unsqueeze(dim=1) return global_roi_grid_points, local_roi_grid_points @staticmethod def get_dense_grid_points(rois, batch_size_rcnn, grid_size): faked_features = rois.new_ones((grid_size, grid_size, grid_size)) dense_idx = faked_features.nonzero() dense_idx = dense_idx.repeat(batch_size_rcnn, 1, 1).float() local_roi_size = rois.view(batch_size_rcnn, -1)[:, 3:6] roi_grid_points = (dense_idx + 0.5) / grid_size * local_roi_size.unsqueeze(dim=1) \ - (local_roi_size.unsqueeze(dim=1) / 2) return roi_grid_points @staticmethod def get_corner_points(rois, batch_size_rcnn): faked_features = rois.new_ones((2, 2, 2)) dense_idx = faked_features.nonzero() dense_idx = dense_idx.repeat(batch_size_rcnn, 1, 1).float() local_roi_size = rois.view(batch_size_rcnn, -1)[:, 3:6] roi_grid_points = dense_idx * local_roi_size.unsqueeze(dim=1) \ - (local_roi_size.unsqueeze(dim=1) / 2) return roi_grid_points def roi_grid_pool(self, batch_size, rois, point_coords, point_features,batch_dict=None,batch_cnt=None): num_frames = batch_dict['num_frames'] num_rois = rois.shape[2]*rois.shape[1] global_roi_proxy_points, local_roi_proxy_points = self.get_proxy_points_of_roi( rois.permute(0,2,1,3).contiguous(), grid_size=self.grid_size ) global_roi_proxy_points = global_roi_proxy_points.view(batch_size, -1, 3) point_coords = point_coords.view(point_coords.shape[0]*num_frames,point_coords.shape[1]//num_frames,point_coords.shape[-1]) xyz = point_coords[:, :, 0:3].view(-1,3) num_points = point_coords.shape[1] num_proxy_points = self.num_proxy_points if batch_cnt is None: xyz_batch_cnt = torch.tensor([num_points]*num_rois*batch_size).cuda().int() else: xyz_batch_cnt = torch.tensor(batch_cnt).cuda().int() new_xyz_batch_cnt = torch.tensor([num_proxy_points]*num_rois*batch_size).cuda().int() new_xyz = global_roi_proxy_points.view(-1, 3) _, pooled_features = self.roi_grid_pool_layer( xyz=xyz.contiguous(), xyz_batch_cnt=xyz_batch_cnt, new_xyz=new_xyz, new_xyz_batch_cnt=new_xyz_batch_cnt, features=point_features.view(-1,point_features.shape[-1]).contiguous(), ) features = pooled_features.view( point_features.shape[0], num_frames*self.num_proxy_points, pooled_features.shape[-1]).contiguous() return features,global_roi_proxy_points.view(batch_size*rois.shape[2], num_frames*num_proxy_points,3).contiguous() def get_proxy_points_of_roi(self, rois, grid_size): rois = rois.view(-1, rois.shape[-1]) batch_size_rcnn = rois.shape[0] local_roi_grid_points = self.get_dense_grid_points(rois, batch_size_rcnn, grid_size) local_roi_grid_points = common_utils.rotate_points_along_z(local_roi_grid_points.clone(), rois[:, 6]).squeeze(dim=1) global_center = rois[:, 0:3].clone() global_roi_grid_points = local_roi_grid_points + global_center.unsqueeze(dim=1) return global_roi_grid_points, local_roi_grid_points def spherical_coordinate(self, src, diag_dist): assert (src.shape[-1] == 27) device = src.device indices_x = torch.LongTensor([0,3,6,9,12,15,18,21,24]).to(device) # indices_y = torch.LongTensor([1,4,7,10,13,16,19,22,25]).to(device) # indices_z = torch.LongTensor([2,5,8,11,14,17,20,23,26]).to(device) src_x = torch.index_select(src, -1, indices_x) src_y = torch.index_select(src, -1, indices_y) src_z = torch.index_select(src, -1, indices_z) dis = (src_x ** 2 + src_y ** 2 + src_z ** 2) ** 0.5 phi = torch.atan(src_y / (src_x + 1e-5)) the = torch.acos(src_z / (dis + 1e-5)) dis = dis / (diag_dist + 1e-5) src = torch.cat([dis, phi, the], dim = -1) return src def crop_current_frame_points(self, src, batch_size,trajectory_rois,num_rois,batch_dict): for bs_idx in range(batch_size): cur_batch_boxes = trajectory_rois[bs_idx,0,:,:7].view(-1,7) cur_radiis = torch.sqrt((cur_batch_boxes[:,3]/2) ** 2 + (cur_batch_boxes[:,4]/2) ** 2) * 1.1 cur_points = batch_dict['points'][(batch_dict['points'][:, 0] == bs_idx)][:,1:] dis = torch.norm((cur_points[:,:2].unsqueeze(0) - cur_batch_boxes[:,:2].unsqueeze(1).repeat(1,cur_points.shape[0],1)), dim = 2) point_mask = (dis <= cur_radiis.unsqueeze(-1)) sampled_idx = torch.topk(point_mask.float(),128)[1] sampled_idx_buffer = sampled_idx[:, 0:1].repeat(1, 128) roi_idx = torch.arange(num_rois)[:, None].repeat(1, 128) sampled_mask = point_mask[roi_idx, sampled_idx] sampled_idx_buffer[sampled_mask] = sampled_idx[sampled_mask] src[bs_idx] = cur_points[sampled_idx_buffer][:,:,:5] empty_flag = sampled_mask.sum(-1)==0 src[bs_idx,empty_flag] = 0 src = src.repeat([1,1,trajectory_rois.shape[1],1]) return src def crop_previous_frame_points(self,src,batch_size,trajectory_rois,num_rois,valid_length,batch_dict): for bs_idx in range(batch_size): cur_points = batch_dict['points'][(batch_dict['points'][:, 0] == bs_idx)][:,1:] for idx in range(1,trajectory_rois.shape[1]): time_mask = (cur_points[:,-1] - idx*0.1).abs() < 1e-3 cur_time_points = cur_points[time_mask] cur_batch_boxes = trajectory_rois[bs_idx,idx,:,:7].view(-1,7) cur_radiis = torch.sqrt((cur_batch_boxes[:,3]/2) ** 2 + (cur_batch_boxes[:,4]/2) ** 2) * 1.1 if not self.training and cur_batch_boxes.shape[0] > 32: length_iter= cur_batch_boxes.shape[0]//32 dis_list = [] for i in range(length_iter+1): dis = torch.norm((cur_time_points[:,:2].unsqueeze(0) - \ cur_batch_boxes[32*i:32*(i+1),:2].unsqueeze(1).repeat(1,cur_time_points.shape[0],1)), dim = 2) dis_list.append(dis) dis = torch.cat(dis_list,0) else: dis = torch.norm((cur_time_points[:,:2].unsqueeze(0) - \ cur_batch_boxes[:,:2].unsqueeze(1).repeat(1,cur_time_points.shape[0],1)), dim = 2) point_mask = (dis <= cur_radiis.unsqueeze(-1)).view(trajectory_rois.shape[2],-1) for roi_box_idx in range(0, num_rois): if not valid_length[bs_idx,idx,roi_box_idx]: continue cur_roi_points = cur_time_points[point_mask[roi_box_idx]] if cur_roi_points.shape[0] > self.num_lidar_points: np.random.seed(0) choice = np.random.choice(cur_roi_points.shape[0], self.num_lidar_points, replace=True) cur_roi_points_sample = cur_roi_points[choice] elif cur_roi_points.shape[0] == 0: cur_roi_points_sample = cur_roi_points.new_zeros(self.num_lidar_points, 6) else: empty_num = self.num_lidar_points - cur_roi_points.shape[0] add_zeros = cur_roi_points.new_zeros(empty_num, 6) add_zeros = cur_roi_points[0].repeat(empty_num, 1) cur_roi_points_sample = torch.cat([cur_roi_points, add_zeros], dim = 0) if not self.use_time_stamp: cur_roi_points_sample = cur_roi_points_sample[:,:-1] src[bs_idx, roi_box_idx, self.num_lidar_points*idx:self.num_lidar_points*(idx+1), :] = cur_roi_points_sample return src def get_proposal_aware_geometry_feature(self,src, batch_size,trajectory_rois,num_rois,batch_dict): proposal_aware_feat_list = [] for i in range(trajectory_rois.shape[1]): corner_points, _ = self.get_corner_points_of_roi(trajectory_rois[:,i,:,:].contiguous()) corner_points = corner_points.view(batch_size, num_rois, -1, corner_points.shape[-1]) corner_points = corner_points.view(batch_size * num_rois, -1) trajectory_roi_center = trajectory_rois[:,i,:,:].contiguous().reshape(batch_size * num_rois, -1)[:,:3] corner_add_center_points = torch.cat([corner_points, trajectory_roi_center], dim = -1) proposal_aware_feat = src[:,i*self.num_lidar_points:(i+1)*self.num_lidar_points,:3].repeat(1,1,9) - \ corner_add_center_points.unsqueeze(1).repeat(1,self.num_lidar_points,1) lwh = trajectory_rois[:,i,:,:].reshape(batch_size * num_rois, -1)[:,3:6].unsqueeze(1).repeat(1,proposal_aware_feat.shape[1],1) diag_dist = (lwh[:,:,0]**2 + lwh[:,:,1]**2 + lwh[:,:,2]**2) ** 0.5 proposal_aware_feat = self.spherical_coordinate(proposal_aware_feat, diag_dist = diag_dist.unsqueeze(-1)) proposal_aware_feat_list.append(proposal_aware_feat) proposal_aware_feat = torch.cat(proposal_aware_feat_list,dim=1) proposal_aware_feat = torch.cat([proposal_aware_feat, src[:,:,3:]], dim = -1) src_gemoetry = self.up_dimension_geometry(proposal_aware_feat) proxy_point_geometry, proxy_points = self.roi_grid_pool(batch_size,trajectory_rois,src,src_gemoetry,batch_dict,batch_cnt=None) return proxy_point_geometry,proxy_points def get_proposal_aware_motion_feature(self,proxy_point,batch_size,trajectory_rois,num_rois,batch_dict): time_stamp = torch.ones([proxy_point.shape[0],proxy_point.shape[1],1]).cuda() padding_zero = torch.zeros([proxy_point.shape[0],proxy_point.shape[1],2]).cuda() proxy_point_time_padding = torch.cat([padding_zero,time_stamp],-1) num_frames = trajectory_rois.shape[1] for i in range(num_frames): proxy_point_time_padding[:,i*self.num_proxy_points:(i+1)*self.num_proxy_points,-1] = i*0.1 corner_points, _ = self.get_corner_points_of_roi(trajectory_rois[:,0,:,:].contiguous()) corner_points = corner_points.view(batch_size, num_rois, -1, corner_points.shape[-1]) corner_points = corner_points.view(batch_size * num_rois, -1) trajectory_roi_center = trajectory_rois[:,0,:,:].reshape(batch_size * num_rois, -1)[:,:3] corner_add_center_points = torch.cat([corner_points, trajectory_roi_center], dim = -1) proposal_aware_feat = proxy_point[:,:,:3].repeat(1,1,9) - corner_add_center_points.unsqueeze(1) lwh = trajectory_rois[:,0,:,:].reshape(batch_size * num_rois, -1)[:,3:6].unsqueeze(1).repeat(1,proxy_point.shape[1],1) diag_dist = (lwh[:,:,0]**2 + lwh[:,:,1]**2 + lwh[:,:,2]**2) ** 0.5 proposal_aware_feat = self.spherical_coordinate(proposal_aware_feat, diag_dist = diag_dist.unsqueeze(-1)) proposal_aware_feat = torch.cat([proposal_aware_feat,proxy_point_time_padding],-1) proxy_point_motion_feat = self.up_dimension_motion(proposal_aware_feat) return proxy_point_motion_feat def trajectories_auxiliary_branch(self,trajectory_rois): time_stamp = torch.ones([trajectory_rois.shape[0],trajectory_rois.shape[1],trajectory_rois.shape[2],1]).cuda() for i in range(time_stamp.shape[1]): time_stamp[:,i,:] = i*0.1 box_seq = torch.cat([trajectory_rois[:,:,:,:7],time_stamp],-1) box_seq[:, :, :,0:3] = box_seq[:, :, :,0:3] - box_seq[:, 0:1, :, 0:3] roi_ry = box_seq[:,:,:,6] % (2 * np.pi) roi_ry_t0 = roi_ry[:,0] roi_ry_t0 = roi_ry_t0.repeat(1,box_seq.shape[1]) box_seq = common_utils.rotate_points_along_z( points=box_seq.view(-1, 1, box_seq.shape[-1]), angle=-roi_ry_t0.view(-1) ).view(box_seq.shape[0],box_seq.shape[1], -1, box_seq.shape[-1]) box_seq[:, :, :, 6] = 0 batch_rcnn = box_seq.shape[0]*box_seq.shape[2] box_reg, box_feat, _ = self.seqboxembed(box_seq.permute(0,2,3,1).contiguous().view(batch_rcnn,box_seq.shape[-1],box_seq.shape[1])) return box_reg, box_feat def generate_trajectory(self,cur_batch_boxes,proposals_list,batch_dict): trajectory_rois = cur_batch_boxes[:,None,:,:].repeat(1,batch_dict['rois'].shape[-2],1,1) trajectory_rois[:,0,:,:]= cur_batch_boxes valid_length = torch.zeros([batch_dict['batch_size'],batch_dict['rois'].shape[-2],trajectory_rois.shape[2]]) valid_length[:,0] = 1 num_frames = batch_dict['rois'].shape[-2] for i in range(1,num_frames): frame = torch.zeros_like(cur_batch_boxes) frame[:,:,0:2] = trajectory_rois[:,i-1,:,0:2] + trajectory_rois[:,i-1,:,7:9] frame[:,:,2:] = trajectory_rois[:,i-1,:,2:] for bs_idx in range( batch_dict['batch_size']): iou3d = iou3d_nms_utils.boxes_iou3d_gpu(frame[bs_idx,:,:7], proposals_list[bs_idx,i,:,:7]) max_overlaps, traj_assignment = torch.max(iou3d, dim=1) fg_inds = ((max_overlaps >= 0.5)).nonzero().view(-1) valid_length[bs_idx,i,fg_inds] = 1 trajectory_rois[bs_idx,i,fg_inds,:] = proposals_list[bs_idx,i,traj_assignment[fg_inds]] batch_dict['valid_length'] = valid_length return trajectory_rois,valid_length def forward(self, batch_dict): """ :param input_data: input dict :return: """ batch_dict['rois'] = batch_dict['proposals_list'].permute(0,2,1,3) num_rois = batch_dict['rois'].shape[1] batch_dict['num_frames'] = batch_dict['rois'].shape[2] batch_dict['roi_scores'] = batch_dict['roi_scores'].permute(0,2,1) batch_dict['roi_labels'] = batch_dict['roi_labels'][:,0,:].long() proposals_list = batch_dict['proposals_list'] batch_size = batch_dict['batch_size'] cur_batch_boxes = copy.deepcopy(batch_dict['rois'].detach())[:,:,0] batch_dict['cur_frame_idx'] = 0 trajectory_rois,valid_length = self.generate_trajectory(cur_batch_boxes,proposals_list,batch_dict) batch_dict['traj_memory'] = trajectory_rois batch_dict['has_class_labels'] = True batch_dict['trajectory_rois'] = trajectory_rois if self.training: targets_dict = self.assign_targets(batch_dict) batch_dict['rois'] = targets_dict['rois'] batch_dict['roi_scores'] = targets_dict['roi_scores'] batch_dict['roi_labels'] = targets_dict['roi_labels'] targets_dict['trajectory_rois'][:,batch_dict['cur_frame_idx'],:,:] = batch_dict['rois'] trajectory_rois = targets_dict['trajectory_rois'] valid_length = targets_dict['valid_length'] empty_mask = batch_dict['rois'][:,:,:6].sum(-1)==0 else: empty_mask = batch_dict['rois'][:,:,0,:6].sum(-1)==0 batch_dict['valid_traj_mask'] = ~empty_mask rois = batch_dict['rois'] num_rois = batch_dict['rois'].shape[1] num_sample = self.num_lidar_points src = rois.new_zeros(batch_size, num_rois, num_sample, 5) src = self.crop_current_frame_points(src, batch_size, trajectory_rois, num_rois,batch_dict) src = self.crop_previous_frame_points(src, batch_size,trajectory_rois, num_rois,valid_length,batch_dict) src = src.view(batch_size * num_rois, -1, src.shape[-1]) src_geometry_feature,proxy_points = self.get_proposal_aware_geometry_feature(src,batch_size,trajectory_rois,num_rois,batch_dict) src_motion_feature = self.get_proposal_aware_motion_feature(proxy_points,batch_size,trajectory_rois,num_rois,batch_dict) src = src_geometry_feature + src_motion_feature box_reg, feat_box = self.trajectories_auxiliary_branch(trajectory_rois) if self.model_cfg.get('USE_TRAJ_EMPTY_MASK',None): src[empty_mask.view(-1)] = 0 if self.model_cfg.Transformer.use_grid_pos.init_type == 'index': pos = self.grid_pos_embeded(self.grid_index.cuda())[None,:,:] pos = torch.cat([torch.zeros(1,1,self.hidden_dim).cuda(),pos],1) else: pos=None hs, tokens = self.transformer(src,pos=pos) point_cls_list = [] point_reg_list = [] for i in range(self.num_enc_layer): point_cls_list.append(self.class_embed[0](tokens[i][0])) for i in range(hs.shape[0]): for j in range(self.num_enc_layer): point_reg_list.append(self.bbox_embed[i](tokens[j][i])) point_cls = torch.cat(point_cls_list,0) point_reg = torch.cat(point_reg_list,0) hs = hs.permute(1,0,2).reshape(hs.shape[1],-1) joint_reg = self.jointembed(torch.cat([hs,feat_box],-1)) rcnn_cls = point_cls rcnn_reg = joint_reg if not self.training: batch_dict['rois'] = batch_dict['rois'][:,:,0].contiguous() rcnn_cls = rcnn_cls[-rcnn_cls.shape[0]//self.num_enc_layer:] batch_cls_preds, batch_box_preds = self.generate_predicted_boxes( batch_size=batch_dict['batch_size'], rois=batch_dict['rois'], cls_preds=rcnn_cls, box_preds=rcnn_reg ) batch_dict['batch_box_preds'] = batch_box_preds batch_dict['cls_preds_normalized'] = False if self.avg_stage1_score: stage1_score = batch_dict['roi_scores'][:,:,:1] batch_cls_preds = F.sigmoid(batch_cls_preds) if self.model_cfg.get('IOU_WEIGHT', None): batch_box_preds_list = [] roi_labels_list = [] batch_cls_preds_list = [] for bs_idx in range(batch_size): car_mask = batch_dict['roi_labels'][bs_idx] ==1 batch_cls_preds_car = batch_cls_preds[bs_idx].pow(self.model_cfg.IOU_WEIGHT[0])* \ stage1_score[bs_idx].pow(1-self.model_cfg.IOU_WEIGHT[0]) batch_cls_preds_car = batch_cls_preds_car[car_mask][None] batch_cls_preds_pedcyc = batch_cls_preds[bs_idx].pow(self.model_cfg.IOU_WEIGHT[1])* \ stage1_score[bs_idx].pow(1-self.model_cfg.IOU_WEIGHT[1]) batch_cls_preds_pedcyc = batch_cls_preds_pedcyc[~car_mask][None] cls_preds = torch.cat([batch_cls_preds_car,batch_cls_preds_pedcyc],1) box_preds = torch.cat([batch_dict['batch_box_preds'][bs_idx][car_mask], batch_dict['batch_box_preds'][bs_idx][~car_mask]],0)[None] roi_labels = torch.cat([batch_dict['roi_labels'][bs_idx][car_mask], batch_dict['roi_labels'][bs_idx][~car_mask]],0)[None] batch_box_preds_list.append(box_preds) roi_labels_list.append(roi_labels) batch_cls_preds_list.append(cls_preds) batch_dict['batch_box_preds'] = torch.cat(batch_box_preds_list,0) batch_dict['roi_labels'] = torch.cat(roi_labels_list,0) batch_cls_preds = torch.cat(batch_cls_preds_list,0) else: batch_cls_preds = torch.sqrt(batch_cls_preds*stage1_score) batch_dict['cls_preds_normalized'] = True batch_dict['batch_cls_preds'] = batch_cls_preds else: targets_dict['batch_size'] = batch_size targets_dict['rcnn_cls'] = rcnn_cls targets_dict['rcnn_reg'] = rcnn_reg targets_dict['box_reg'] = box_reg targets_dict['point_reg'] = point_reg targets_dict['point_cls'] = point_cls self.forward_ret_dict = targets_dict return batch_dict def get_loss(self, tb_dict=None): tb_dict = {} if tb_dict is None else tb_dict rcnn_loss = 0 rcnn_loss_cls, cls_tb_dict = self.get_box_cls_layer_loss(self.forward_ret_dict) rcnn_loss += rcnn_loss_cls tb_dict.update(cls_tb_dict) rcnn_loss_reg, reg_tb_dict = self.get_box_reg_layer_loss(self.forward_ret_dict) rcnn_loss += rcnn_loss_reg tb_dict.update(reg_tb_dict) tb_dict['rcnn_loss'] = rcnn_loss.item() return rcnn_loss, tb_dict def get_box_reg_layer_loss(self, forward_ret_dict): loss_cfgs = self.model_cfg.LOSS_CONFIG code_size = self.box_coder.code_size reg_valid_mask = forward_ret_dict['reg_valid_mask'].view(-1) batch_size = forward_ret_dict['batch_size'] gt_boxes3d_ct = forward_ret_dict['gt_of_rois'][..., 0:code_size] gt_of_rois_src = forward_ret_dict['gt_of_rois_src'][..., 0:code_size].view(-1, code_size) rcnn_reg = forward_ret_dict['rcnn_reg'] roi_boxes3d = forward_ret_dict['rois'] rcnn_batch_size = gt_boxes3d_ct.view(-1, code_size).shape[0] fg_mask = (reg_valid_mask > 0) fg_sum = fg_mask.long().sum().item() tb_dict = {} if loss_cfgs.REG_LOSS == 'smooth-l1': rois_anchor = roi_boxes3d.clone().detach()[:,:,:7].contiguous().view(-1, code_size) rois_anchor[:, 0:3] = 0 rois_anchor[:, 6] = 0 reg_targets = self.box_coder.encode_torch( gt_boxes3d_ct.view(rcnn_batch_size, code_size), rois_anchor ) rcnn_loss_reg = self.reg_loss_func( rcnn_reg.view(rcnn_batch_size, -1).unsqueeze(dim=0), reg_targets.unsqueeze(dim=0), ) # [B, M, 7] rcnn_loss_reg = (rcnn_loss_reg.view(rcnn_batch_size, -1) * fg_mask.unsqueeze(dim=-1).float()).sum() / max(fg_sum, 1) rcnn_loss_reg = rcnn_loss_reg * loss_cfgs.LOSS_WEIGHTS['rcnn_reg_weight']*loss_cfgs.LOSS_WEIGHTS['traj_reg_weight'][0] tb_dict['rcnn_loss_reg'] = rcnn_loss_reg.item() if self.model_cfg.USE_AUX_LOSS: point_reg = forward_ret_dict['point_reg'] groups = point_reg.shape[0]//reg_targets.shape[0] if groups != 1 : point_loss_regs = 0 slice = reg_targets.shape[0] for i in range(groups): point_loss_reg = self.reg_loss_func( point_reg[i*slice:(i+1)*slice].view(slice, -1).unsqueeze(dim=0),reg_targets.unsqueeze(dim=0),) point_loss_reg = (point_loss_reg.view(slice, -1) * fg_mask.unsqueeze(dim=-1).float()).sum() / max(fg_sum, 1) point_loss_reg = point_loss_reg * loss_cfgs.LOSS_WEIGHTS['rcnn_reg_weight']*loss_cfgs.LOSS_WEIGHTS['traj_reg_weight'][2] point_loss_regs += point_loss_reg point_loss_regs = point_loss_regs / groups tb_dict['point_loss_reg'] = point_loss_regs.item() rcnn_loss_reg += point_loss_regs else: point_loss_reg = self.reg_loss_func(point_reg.view(rcnn_batch_size, -1).unsqueeze(dim=0),reg_targets.unsqueeze(dim=0),) point_loss_reg = (point_loss_reg.view(rcnn_batch_size, -1) * fg_mask.unsqueeze(dim=-1).float()).sum() / max(fg_sum, 1) point_loss_reg = point_loss_reg * loss_cfgs.LOSS_WEIGHTS['rcnn_reg_weight']*loss_cfgs.LOSS_WEIGHTS['traj_reg_weight'][2] tb_dict['point_loss_reg'] = point_loss_reg.item() rcnn_loss_reg += point_loss_reg seqbox_reg = forward_ret_dict['box_reg'] seqbox_loss_reg = self.reg_loss_func(seqbox_reg.view(rcnn_batch_size, -1).unsqueeze(dim=0),reg_targets.unsqueeze(dim=0),) seqbox_loss_reg = (seqbox_loss_reg.view(rcnn_batch_size, -1) * fg_mask.unsqueeze(dim=-1).float()).sum() / max(fg_sum, 1) seqbox_loss_reg = seqbox_loss_reg * loss_cfgs.LOSS_WEIGHTS['rcnn_reg_weight']*loss_cfgs.LOSS_WEIGHTS['traj_reg_weight'][1] tb_dict['seqbox_loss_reg'] = seqbox_loss_reg.item() rcnn_loss_reg += seqbox_loss_reg if loss_cfgs.CORNER_LOSS_REGULARIZATION and fg_sum > 0: fg_rcnn_reg = rcnn_reg.view(rcnn_batch_size, -1)[fg_mask] fg_roi_boxes3d = roi_boxes3d[:,:,:7].contiguous().view(-1, code_size)[fg_mask] fg_roi_boxes3d = fg_roi_boxes3d.view(1, -1, code_size) batch_anchors = fg_roi_boxes3d.clone().detach() roi_ry = fg_roi_boxes3d[:, :, 6].view(-1) roi_xyz = fg_roi_boxes3d[:, :, 0:3].view(-1, 3) batch_anchors[:, :, 0:3] = 0 rcnn_boxes3d = self.box_coder.decode_torch( fg_rcnn_reg.view(batch_anchors.shape[0], -1, code_size), batch_anchors ).view(-1, code_size) rcnn_boxes3d = common_utils.rotate_points_along_z( rcnn_boxes3d.unsqueeze(dim=1), roi_ry ).squeeze(dim=1) rcnn_boxes3d[:, 0:3] += roi_xyz corner_loss_func = loss_utils.get_corner_loss_lidar loss_corner = corner_loss_func( rcnn_boxes3d[:, 0:7], gt_of_rois_src[fg_mask][:, 0:7]) loss_corner = loss_corner.mean() loss_corner = loss_corner * loss_cfgs.LOSS_WEIGHTS['rcnn_corner_weight'] rcnn_loss_reg += loss_corner tb_dict['rcnn_loss_corner'] = loss_corner.item() else: raise NotImplementedError return rcnn_loss_reg, tb_dict def get_box_cls_layer_loss(self, forward_ret_dict): loss_cfgs = self.model_cfg.LOSS_CONFIG rcnn_cls = forward_ret_dict['rcnn_cls'] rcnn_cls_labels = forward_ret_dict['rcnn_cls_labels'].view(-1) if loss_cfgs.CLS_LOSS == 'BinaryCrossEntropy': rcnn_cls_flat = rcnn_cls.view(-1) groups = rcnn_cls_flat.shape[0] // rcnn_cls_labels.shape[0] if groups != 1: rcnn_loss_cls = 0 slice = rcnn_cls_labels.shape[0] for i in range(groups): batch_loss_cls = F.binary_cross_entropy(torch.sigmoid(rcnn_cls_flat[i*slice:(i+1)*slice]), rcnn_cls_labels.float(), reduction='none') cls_valid_mask = (rcnn_cls_labels >= 0).float() rcnn_loss_cls = rcnn_loss_cls + (batch_loss_cls * cls_valid_mask).sum() / torch.clamp(cls_valid_mask.sum(), min=1.0) rcnn_loss_cls = rcnn_loss_cls / groups else: batch_loss_cls = F.binary_cross_entropy(torch.sigmoid(rcnn_cls_flat), rcnn_cls_labels.float(), reduction='none') cls_valid_mask = (rcnn_cls_labels >= 0).float() rcnn_loss_cls = (batch_loss_cls * cls_valid_mask).sum() / torch.clamp(cls_valid_mask.sum(), min=1.0) elif loss_cfgs.CLS_LOSS == 'CrossEntropy': batch_loss_cls = F.cross_entropy(rcnn_cls, rcnn_cls_labels, reduction='none', ignore_index=-1) cls_valid_mask = (rcnn_cls_labels >= 0).float() rcnn_loss_cls = (batch_loss_cls * cls_valid_mask).sum() / torch.clamp(cls_valid_mask.sum(), min=1.0) else: raise NotImplementedError rcnn_loss_cls = rcnn_loss_cls * loss_cfgs.LOSS_WEIGHTS['rcnn_cls_weight'] tb_dict = {'rcnn_loss_cls': rcnn_loss_cls.item()} return rcnn_loss_cls, tb_dict def generate_predicted_boxes(self, batch_size, rois, cls_preds=None, box_preds=None): """ Args: batch_size: rois: (B, N, 7) cls_preds: (BN, num_class) box_preds: (BN, code_size) Returns: """ code_size = self.box_coder.code_size if cls_preds is not None: batch_cls_preds = cls_preds.view(batch_size, -1, cls_preds.shape[-1]) else: batch_cls_preds = None batch_box_preds = box_preds.view(batch_size, -1, code_size) roi_ry = rois[:, :, 6].view(-1) roi_xyz = rois[:, :, 0:3].view(-1, 3) local_rois = rois.clone().detach() local_rois[:, :, 0:3] = 0 batch_box_preds = self.box_coder.decode_torch(batch_box_preds, local_rois).view(-1, code_size) batch_box_preds = common_utils.rotate_points_along_z( batch_box_preds.unsqueeze(dim=1), roi_ry ).squeeze(dim=1) batch_box_preds[:, 0:3] += roi_xyz batch_box_preds = batch_box_preds.view(batch_size, -1, code_size) batch_box_preds = torch.cat([batch_box_preds,rois[:,:,7:]],-1) return batch_cls_preds, batch_box_preds