refactor/rename some parts of the codes

pcdet/models/backbones_3d/focal_sparse_conv/focal_sparse_conv.py

@@ -2,7 +2,7 @@ import torch
 import torch.nn as nn
 import spconv.pytorch as spconv
 from pcdet.ops.roiaware_pool3d.roiaware_pool3d_utils import points_in_boxes_gpu
-from pcdet.models.backbones_3d.focal_sparse_conv.utils import split_voxels, check_repeat, FocalLoss
+from pcdet.models.backbones_3d.focal_sparse_conv.focal_sparse_utils import split_voxels, check_repeat, FocalLoss
 from pcdet.utils import common_utils
 
 
@@ -221,5 +221,4 @@ class FocalSparseConv(spconv.SparseModule):
         out = out.replace_feature(self.bn1(out.features))
         out = out.replace_feature(self.relu(out.features))
 
-        batch_dict['loss_box_of_pts'] += loss_box_of_pts
-        return out, batch_dict
+        return out, batch_dict, loss_box_of_pts

pcdet/models/backbones_3d/spconv_backbone_focal.py

@@ -196,6 +196,15 @@ class VoxelBackBone8xFocal(nn.Module):
             'x_conv4': 64
         }
         
+        self.forward_ret_dict = {}
+        
+    def get_loss(self, tb_dict=None):
+        loss = self.forward_ret_dict['loss_box_of_pts']
+        if tb_dict is None:
+            tb_dict = {}
+        tb_dict['loss_box_of_pts'] = loss.item()
+        return loss, tb_dict
+
     def forward(self, batch_dict):
         """
         Args:
@@ -221,9 +230,12 @@ class VoxelBackBone8xFocal(nn.Module):
         x = self.conv_input(input_sp_tensor)
         x_conv1, batch_dict = self.conv1(x, batch_dict)
 
+        loss_box_of_pts = 0
         if self.use_img:
             x_image = self.semseg(batch_dict['images'])['layer1_feat2d']
-            x_conv1, batch_dict = self.conv_focal_multimodal(x_conv1, batch_dict, x_image)
+            x_conv1, batch_dict, loss_box_of_pts = self.conv_focal_multimodal(x_conv1, batch_dict, x_image)
+
+        self.forward_ret_dict['loss_box_of_pts'] = loss_box_of_pts
 
         x_conv2, batch_dict = self.conv2(x_conv1, batch_dict)
         x_conv3, batch_dict = self.conv3(x_conv2, batch_dict)

pcdet/models/detectors/pv_rcnn.py

@@ -12,9 +12,6 @@ class PVRCNN(Detector3DTemplate):
 
         if self.training:
             loss, tb_dict, disp_dict = self.get_training_loss()
-            if 'loss_box_of_pts' in batch_dict:
-                loss += batch_dict['loss_box_of_pts']
-                tb_dict['loss_box_of_pts'] = batch_dict['loss_box_of_pts']
 
             ret_dict = {
                 'loss': loss
@@ -31,4 +28,9 @@ class PVRCNN(Detector3DTemplate):
         loss_rcnn, tb_dict = self.roi_head.get_loss(tb_dict)
 
         loss = loss_rpn + loss_point + loss_rcnn
+        
+        if hasattr(self.backbone_3d, 'get_loss'):
+            loss_backbone3d, tb_dict = self.backbone_3d.get_loss(tb_dict)
+            loss += loss_backbone3d
+        
         return loss, tb_dict, disp_dict

pcdet/models/detectors/voxel_rcnn.py

@@ -13,10 +13,6 @@ class VoxelRCNN(Detector3DTemplate):
         if self.training:
             loss, tb_dict, disp_dict = self.get_training_loss()
 
-            if 'loss_box_of_pts' in batch_dict:
-                loss += batch_dict['loss_box_of_pts']
-                tb_dict['loss_box_of_pts'] = batch_dict['loss_box_of_pts']
-
             ret_dict = {
                 'loss': loss
             }
@@ -33,4 +29,9 @@ class VoxelRCNN(Detector3DTemplate):
         loss_rcnn, tb_dict = self.roi_head.get_loss(tb_dict)
 
         loss = loss + loss_rpn + loss_rcnn
+        
+        if hasattr(self.backbone_3d, 'get_loss'):
+            loss_backbone3d, tb_dict = self.backbone_3d.get_loss(tb_dict)
+            loss += loss_backbone3d
+            
         return loss, tb_dict, disp_dict

pcdet/utils/box2d_utils.py

-import torch
-
-
-def area(box) -> torch.Tensor:
-    """
-    Computes the area of all the boxes.
-
-    Returns:
-        torch.Tensor: a vector with areas of each box.
-    """
-    area = (box[:, 2] - box[:, 0]) * (box[:, 3] - box[:, 1])
-    return area
-
-# implementation from https://github.com/kuangliu/torchcv/blob/master/torchcv/utils/box.py
-# with slight modifications
-def pairwise_iou(boxes1, boxes2) -> torch.Tensor:
-    """
-    Given two lists of boxes of size N and M,
-    compute the IoU (intersection over union)
-    between __all__ N x M pairs of boxes.
-    The box order must be (xmin, ymin, xmax, ymax).
-
-    Args:
-        boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively.
-
-    Returns:
-        Tensor: IoU, sized [N,M].
-    """
-    area1 = area(boxes1)
-    area2 = area(boxes2)
-
-    width_height = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) - torch.max(
-        boxes1[:, None, :2], boxes2[:, :2]
-    )  # [N,M,2]
-
-    width_height.clamp_(min=0)  # [N,M,2]
-    inter = width_height.prod(dim=2)  # [N,M]
-    del width_height
-
-    # handle empty boxes
-    iou = torch.where(
-        inter > 0,
-        inter / (area1[:, None] + area2 - inter),
-        torch.zeros(1, dtype=inter.dtype, device=inter.device),
-    )
-    return iou
\ No newline at end of file

pcdet/utils/box_utils.py

@@ -333,3 +333,49 @@ def boxes3d_nearest_bev_iou(boxes_a, boxes_b):
     boxes_bev_b = boxes3d_lidar_to_aligned_bev_boxes(boxes_b)
 
     return boxes_iou_normal(boxes_bev_a, boxes_bev_b)
+
+
+def area(box) -> torch.Tensor:
+    """
+    Computes the area of all the boxes.
+
+    Returns:
+        torch.Tensor: a vector with areas of each box.
+    """
+    area = (box[:, 2] - box[:, 0]) * (box[:, 3] - box[:, 1])
+    return area
+
+
+# implementation from https://github.com/kuangliu/torchcv/blob/master/torchcv/utils/box.py
+# with slight modifications
+def pairwise_iou(boxes1, boxes2) -> torch.Tensor:
+    """
+    Given two lists of boxes of size N and M,
+    compute the IoU (intersection over union)
+    between __all__ N x M pairs of boxes.
+    The box order must be (xmin, ymin, xmax, ymax).
+
+    Args:
+        boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively.
+
+    Returns:
+        Tensor: IoU, sized [N,M].
+    """
+    area1 = area(boxes1)
+    area2 = area(boxes2)
+
+    width_height = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) - torch.max(
+        boxes1[:, None, :2], boxes2[:, :2]
+    )  # [N,M,2]
+
+    width_height.clamp_(min=0)  # [N,M,2]
+    inter = width_height.prod(dim=2)  # [N,M]
+    del width_height
+
+    # handle empty boxes
+    iou = torch.where(
+        inter > 0,
+        inter / (area1[:, None] + area2 - inter),
+        torch.zeros(1, dtype=inter.dtype, device=inter.device),
+    )
+    return iou