fix bugs [occured in different environments]

README.md

@@ -26,6 +26,13 @@ A Simple PointPillars PyTorch Implenmentation for 3D Lidar(KITTI) Detection. [[Z
 ![](./figures/pc_pred_000134.png)
 ![](./figures/img_3dbbox_000134.png)
 
+## [Compile] 
+
+```
+cd ops
+python setup.py develop
+```
+
 ## [Datasets]
 
 1. Download
@@ -74,13 +81,6 @@ A Simple PointPillars PyTorch Implenmentation for 3D Lidar(KITTI) Detection. [[Z
     
     ```
 
-## [Compile] 
-
-```
-cd ops
-python setup.py develop
-```
-
 ## [Training]
 
 ```

model/anchors.py

@@ -174,7 +174,7 @@ def anchor_target(batched_anchors, batched_gt_bboxes, batched_gt_labels, assigne
             neg_flag = assigned_gt_inds == 0
             # 1. anchor labels
             assigned_gt_labels = torch.zeros_like(cur_anchors[:, 0], dtype=torch.long) + nclasses # -1 is not optimal, for some bboxes are with labels -1
-            assigned_gt_labels[pos_flag] = gt_labels[assigned_gt_inds[pos_flag] - 1]
+            assigned_gt_labels[pos_flag] = gt_labels[assigned_gt_inds[pos_flag] - 1].long()
             assigned_gt_labels_weights = torch.zeros_like(cur_anchors[:, 0])
             assigned_gt_labels_weights[pos_flag] = 1
             assigned_gt_labels_weights[neg_flag] = 1

model/pointpillars.py

@@ -183,7 +183,7 @@ def forward(self, x):
 
 
 class Head(nn.Module):
-    def __init__(self, in_channel, n_anchors=6, n_classes=3):
+    def __init__(self, in_channel, n_anchors, n_classes):
         super().__init__()
 
         self.conv_cls = nn.Conv2d(in_channel, n_anchors*n_classes, 1)
@@ -240,7 +240,7 @@ def __init__(self,
         self.neck = Neck(in_channels=[64, 128, 256], 
                          upsample_strides=[1, 2, 4], 
                          out_channels=[128, 128, 128])
-        self.head = Head(in_channel=384)
+        self.head = Head(in_channel=384, n_anchors=2*nclasses, n_classes=nclasses)
 
         # anchors
         ranges = [[0, -39.68, -0.6, 69.12, 39.68, -0.6],

pre_process_kitti.py

@@ -33,6 +33,7 @@ def judge_difficulty(annotation_dict):
 
 
 def create_data_info_pkl(data_root, data_type, prefix, label=True, db=False):
+    sep = os.path.sep
     print(f"Processing {data_type} data..")
     ids_file = os.path.join(CUR, 'dataset', 'ImageSets', f'{data_type}.txt')
     with open(ids_file, 'r') as f:
@@ -50,13 +51,13 @@ def create_data_info_pkl(data_root, data_type, prefix, label=True, db=False):
         img_path = os.path.join(data_root, split, 'image_2', f'{id}.png')
         lidar_path = os.path.join(data_root, split, 'velodyne', f'{id}.bin')
         calib_path = os.path.join(data_root, split, 'calib', f'{id}.txt') 
-        cur_info_dict['velodyne_path'] = '/'.join(lidar_path.split('/')[-3:])
+        cur_info_dict['velodyne_path'] = sep.join(lidar_path.split(sep)[-3:])
 
         img = cv2.imread(img_path)
         image_shape = img.shape[:2]
         cur_info_dict['image'] = {
             'image_shape': image_shape,
-            'image_path': '/'.join(img_path.split('/')[-3:]), 
+            'image_path': sep.join(img_path.split(sep)[-3:]), 
             'image_idx': int(id),
         }