Merge branch 'master' into develop

pcdet/datasets/kitti/kitti_object_eval_python/eval.py

@@ -694,10 +694,10 @@ def get_official_eval_result(gt_annos, dt_annos, current_classes, PR_detail_dict
                 result += print_str((f"aos  AP:{mAPaos[j, 0, i]:.2f}, "
                                      f"{mAPaos[j, 1, i]:.2f}, "
                                      f"{mAPaos[j, 2, i]:.2f}"))
-                if i == 0:
-                   ret_dict['%s_aos_easy' % class_to_name[curcls]] = mAPaos[j, 0, 0]
-                   ret_dict['%s_aos_moderate' % class_to_name[curcls]] = mAPaos[j, 1, 0]
-                   ret_dict['%s_aos_hard' % class_to_name[curcls]] = mAPaos[j, 2, 0]
+                # if i == 0:
+                   # ret_dict['%s_aos/easy' % class_to_name[curcls]] = mAPaos[j, 0, 0]
+                   # ret_dict['%s_aos/moderate' % class_to_name[curcls]] = mAPaos[j, 1, 0]
+                   # ret_dict['%s_aos/hard' % class_to_name[curcls]] = mAPaos[j, 2, 0]
 
             result += print_str(
                 (f"{class_to_name[curcls]} "
@@ -716,30 +716,30 @@ def get_official_eval_result(gt_annos, dt_annos, current_classes, PR_detail_dict
                                      f"{mAPaos_R40[j, 1, i]:.2f}, "
                                      f"{mAPaos_R40[j, 2, i]:.2f}"))
                 if i == 0:
-                   ret_dict['%s_aos_easy_R40' % class_to_name[curcls]] = mAPaos_R40[j, 0, 0]
-                   ret_dict['%s_aos_moderate_R40' % class_to_name[curcls]] = mAPaos_R40[j, 1, 0]
-                   ret_dict['%s_aos_hard_R40' % class_to_name[curcls]] = mAPaos_R40[j, 2, 0]
+                   ret_dict['%s_aos/easy_R40' % class_to_name[curcls]] = mAPaos_R40[j, 0, 0]
+                   ret_dict['%s_aos/moderate_R40' % class_to_name[curcls]] = mAPaos_R40[j, 1, 0]
+                   ret_dict['%s_aos/hard_R40' % class_to_name[curcls]] = mAPaos_R40[j, 2, 0]
 
             if i == 0:
-                ret_dict['%s_3d_easy' % class_to_name[curcls]] = mAP3d[j, 0, 0]
-                ret_dict['%s_3d_moderate' % class_to_name[curcls]] = mAP3d[j, 1, 0]
-                ret_dict['%s_3d_hard' % class_to_name[curcls]] = mAP3d[j, 2, 0]
-                ret_dict['%s_bev_easy' % class_to_name[curcls]] = mAPbev[j, 0, 0]
-                ret_dict['%s_bev_moderate' % class_to_name[curcls]] = mAPbev[j, 1, 0]
-                ret_dict['%s_bev_hard' % class_to_name[curcls]] = mAPbev[j, 2, 0]
-                ret_dict['%s_image_easy' % class_to_name[curcls]] = mAPbbox[j, 0, 0]
-                ret_dict['%s_image_moderate' % class_to_name[curcls]] = mAPbbox[j, 1, 0]
-                ret_dict['%s_image_hard' % class_to_name[curcls]] = mAPbbox[j, 2, 0]
-
-                ret_dict['%s_3d_easy_R40' % class_to_name[curcls]] = mAP3d_R40[j, 0, 0]
-                ret_dict['%s_3d_moderate_R40' % class_to_name[curcls]] = mAP3d_R40[j, 1, 0]
-                ret_dict['%s_3d_hard_R40' % class_to_name[curcls]] = mAP3d_R40[j, 2, 0]
-                ret_dict['%s_bev_easy_R40' % class_to_name[curcls]] = mAPbev_R40[j, 0, 0]
-                ret_dict['%s_bev_moderate_R40' % class_to_name[curcls]] = mAPbev_R40[j, 1, 0]
-                ret_dict['%s_bev_hard_R40' % class_to_name[curcls]] = mAPbev_R40[j, 2, 0]
-                ret_dict['%s_image_easy_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 0, 0]
-                ret_dict['%s_image_moderate_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 1, 0]
-                ret_dict['%s_image_hard_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 2, 0]
+                # ret_dict['%s_3d/easy' % class_to_name[curcls]] = mAP3d[j, 0, 0]
+                # ret_dict['%s_3d/moderate' % class_to_name[curcls]] = mAP3d[j, 1, 0]
+                # ret_dict['%s_3d/hard' % class_to_name[curcls]] = mAP3d[j, 2, 0]
+                # ret_dict['%s_bev/easy' % class_to_name[curcls]] = mAPbev[j, 0, 0]
+                # ret_dict['%s_bev/moderate' % class_to_name[curcls]] = mAPbev[j, 1, 0]
+                # ret_dict['%s_bev/hard' % class_to_name[curcls]] = mAPbev[j, 2, 0]
+                # ret_dict['%s_image/easy' % class_to_name[curcls]] = mAPbbox[j, 0, 0]
+                # ret_dict['%s_image/moderate' % class_to_name[curcls]] = mAPbbox[j, 1, 0]
+                # ret_dict['%s_image/hard' % class_to_name[curcls]] = mAPbbox[j, 2, 0]
+
+                ret_dict['%s_3d/easy_R40' % class_to_name[curcls]] = mAP3d_R40[j, 0, 0]
+                ret_dict['%s_3d/moderate_R40' % class_to_name[curcls]] = mAP3d_R40[j, 1, 0]
+                ret_dict['%s_3d/hard_R40' % class_to_name[curcls]] = mAP3d_R40[j, 2, 0]
+                ret_dict['%s_bev/easy_R40' % class_to_name[curcls]] = mAPbev_R40[j, 0, 0]
+                ret_dict['%s_bev/moderate_R40' % class_to_name[curcls]] = mAPbev_R40[j, 1, 0]
+                ret_dict['%s_bev/hard_R40' % class_to_name[curcls]] = mAPbev_R40[j, 2, 0]
+                ret_dict['%s_image/easy_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 0, 0]
+                ret_dict['%s_image/moderate_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 1, 0]
+                ret_dict['%s_image/hard_R40' % class_to_name[curcls]] = mAPbbox_R40[j, 2, 0]
 
     return result, ret_dict
 

pcdet/models/dense_heads/anchor_head_multi.py

@@ -4,6 +4,7 @@ from .anchor_head_template import AnchorHeadTemplate
 from ..backbones_2d import BaseBEVBackbone
 import torch
 
+
 class SingleHead(BaseBEVBackbone):
     def __init__(self, model_cfg, input_channels, num_class, num_anchors_per_location, code_size, encode_conv_cfg=None):
         super().__init__(encode_conv_cfg, input_channels)
@@ -75,6 +76,7 @@ class SingleHead(BaseBEVBackbone):
 
         return ret_dict
 
+
 class AnchorHeadMulti(AnchorHeadTemplate):
     def __init__(self, model_cfg, input_channels, num_class, class_names, grid_size, point_cloud_range, predict_boxes_when_training=True):
         super().__init__(
@@ -83,7 +85,6 @@ class AnchorHeadMulti(AnchorHeadTemplate):
         self.model_cfg = model_cfg
         self.make_multihead(input_channels)
 
-
     def make_multihead(self, input_channels):
         rpn_head_cfgs = self.model_cfg.RPN_HEAD_CFGS
         rpn_heads = []
@@ -123,7 +124,8 @@ class AnchorHeadMulti(AnchorHeadTemplate):
                 gt_boxes=data_dict['gt_boxes']
             )
             self.forward_ret_dict.update(targets_dict)
-        else:
+
+        if not self.training or self.predict_boxes_when_training:
             batch_cls_preds, batch_box_preds = self.generate_predicted_boxes(
                 batch_size=data_dict['batch_size'],
                 cls_preds=cls_preds, box_preds=box_preds, dir_cls_preds=dir_cls_preds

pcdet/models/roi_heads/roi_head_template.py

@@ -39,6 +39,7 @@ class RoIHeadTemplate(nn.Module):
         fc_layers = nn.Sequential(*fc_layers)
         return fc_layers
 
+    @torch.no_grad()
     def proposal_layer(self, batch_dict, nms_config):
         """
         Args:

tools/eval_utils/eval_utils.py

@@ -95,8 +95,8 @@ def eval_one_epoch(cfg, model, dataloader, epoch_id, logger, dist_test=False, sa
         cur_rcnn_recall = metric['recall_rcnn_%s' % str(cur_thresh)] / max(gt_num_cnt, 1)
         logger.info('recall_roi_%s: %f' % (cur_thresh, cur_roi_recall))
         logger.info('recall_rcnn_%s: %f' % (cur_thresh, cur_rcnn_recall))
-        ret_dict['recall_roi_%s' % str(cur_thresh)] = cur_roi_recall
-        ret_dict['recall_rcnn_%s' % str(cur_thresh)] = cur_rcnn_recall
+        ret_dict['recall/roi_%s' % str(cur_thresh)] = cur_roi_recall
+        ret_dict['recall/rcnn_%s' % str(cur_thresh)] = cur_rcnn_recall
 
     total_pred_objects = 0
     for anno in det_annos:

tools/train_utils/train_utils.py

@@ -29,7 +29,7 @@ def train_one_epoch(model, optimizer, train_loader, model_func, lr_scheduler, ac
             cur_lr = optimizer.param_groups[0]['lr']
 
         if tb_log is not None:
-            tb_log.add_scalar('learning_rate', cur_lr, accumulated_iter)
+            tb_log.add_scalar('meta_data/learning_rate', cur_lr, accumulated_iter)
 
         model.train()
         optimizer.zero_grad()
@@ -51,10 +51,10 @@ def train_one_epoch(model, optimizer, train_loader, model_func, lr_scheduler, ac
             tbar.refresh()
 
             if tb_log is not None:
-                tb_log.add_scalar('train_loss', loss, accumulated_iter)
-                tb_log.add_scalar('learning_rate', cur_lr, accumulated_iter)
+                tb_log.add_scalar('train/loss', loss, accumulated_iter)
+                tb_log.add_scalar('meta_data/learning_rate', cur_lr, accumulated_iter)
                 for key, val in tb_dict.items():
-                    tb_log.add_scalar('train_' + key, val, accumulated_iter)
+                    tb_log.add_scalar('train/' + key, val, accumulated_iter)
     if rank == 0:
         pbar.close()
     return accumulated_iter