[Fix] Fix unit test of datasets (#1865)

* fix ci of dataset and visualizer by regenerate pkl

* decrease channels in ut for reducing cuda memory

* still use sample_id in the kitti submitted files

* minor changes

* use waymo-open-dataset-tf-2-6-0

* use waymo-open-dataset-tf-2-5-0 to be compabible with python3.6

* use waymo-open-dataset-tf-2-4-0 to be compabible with python3.6

* no dependency about waymo-open-dataset

* Revert "no dependency about waymo-open-dataset"

This reverts commit 581d4eefe26851a295b3190bcfa73d6fb2ba54aa.

* not install waymo-open-dataset for ci

* max python version is 3.8 because of trimesh

* Revert "max python version is 2.8 because of trimesh"

This reverts commit 6d8a1851b22a2059e7de75b576c2b7f633ad7cfa.

* fix networkx in ci

* update docstring of data_preprocessor

* modify version of typingextensions

mmdet3d/evaluation/metrics/kitti_metric.py

@@ -168,7 +168,7 @@ class KittiMetric(BaseMetric):
         """Compute the metrics from processed results.
 
         Args:
-            results (list): The processed results of each batch.
+            results (list): The processed results of the whole dataset.
 
         Returns:
             Dict[str, float]: The computed metrics. The keys are the names of

mmdet3d/models/data_preprocessors/data_preprocessor.py

@@ -104,10 +104,10 @@ class Det3DDataPreprocessor(DetDataPreprocessor):
         ``BaseDataPreprocessor``.
 
         Args:
-            data (List[dict] | List[List[dict]]): data from dataloader.
-                The outer list always represent the batch size, when it is
-                a list[list[dict]], the inter list indicate test time
-                augmentation.
+            data (dict | List[dict]): data from dataloader.
+                The dict contains the whole batch data, when it is
+                a list[dict], the list indicate test time augmentation.
+
             training (bool): Whether to enable training time augmentation.
                 Defaults to False.
 
@@ -293,7 +293,7 @@ class Det3DDataPreprocessor(DetDataPreprocessor):
         else:
             raise TypeError('Output of `cast_data` should be a list of dict '
                             'or a tuple with inputs and data_samples, but got'
-                            f'{type(data)}： {data}')
+                            f'{type(data)}: {data}')
         return batch_pad_shape
 
     @torch.no_grad()

mmdet3d/visualization/local_visualizer.py

@@ -606,9 +606,8 @@ class Det3DLocalVisualizer(DetLocalVisualizer):
                 win_name=name,
                 wait_time=wait_time)
 
-        mkdir_or_exist(out_file)
-
         if out_file is not None:
+            mkdir_or_exist(out_file)
             if drawn_img_3d is not None:
                 mmcv.imwrite(drawn_img_3d[..., ::-1], out_file + '.jpg')
             if drawn_img is not None:

requirements/optional.txt

+black==20.8b1 # be compatible with typing-extensions 3.7.4
 spconv
-waymo-open-dataset-tf-2-1-0==1.2.0
+typing-extensions==3.7.4 # required by tensorflow before version 2.6
+waymo-open-dataset-tf-2-4-0

requirements/runtime.txt

 lyft_dataset_sdk
-networkx>=2.2,<2.3
+networkx>=2.5
 numba==0.53.0
 numpy
 nuscenes-devkit

tests/test_models/test_dense_heads/test_fcos_mono3d_head.py

@@ -44,9 +44,9 @@ class TestFCOSMono3DHead(TestCase):
 
         fcos_mono3d_head = FCOSMono3DHead(
             num_classes=10,
-            in_channels=256,
+            in_channels=32,
             stacked_convs=2,
-            feat_channels=256,
+            feat_channels=32,
             use_direction_classifier=True,
             diff_rad_by_sin=True,
             pred_attrs=True,
@@ -55,16 +55,16 @@ class TestFCOSMono3DHead(TestCase):
             dir_limit_offset=0,
             strides=[8, 16, 32, 64, 128],
             group_reg_dims=(2, 1, 3, 1, 2),  # offset, depth, size, rot, velo
-            cls_branch=(256, ),
+            cls_branch=(32, ),
             reg_branch=(
-                (256, ),  # offset
-                (256, ),  # depth
-                (256, ),  # size
-                (256, ),  # rot
+                (32, ),  # offset
+                (32, ),  # depth
+                (32, ),  # size
+                (32, ),  # rot
                 ()  # velo
             ),
-            dir_branch=(256, ),
-            attr_branch=(256, ),
+            dir_branch=(32, ),
+            attr_branch=(32, ),
             loss_cls=dict(
                 type='mmdet.FocalLoss',
                 use_sigmoid=True,
@@ -96,11 +96,11 @@ class TestFCOSMono3DHead(TestCase):
 
         # FCOS3D head expects a multiple levels of features per image
         feats = [
-            torch.rand([1, 256, 116, 200], dtype=torch.float32),
-            torch.rand([1, 256, 58, 100], dtype=torch.float32),
-            torch.rand([1, 256, 29, 50], dtype=torch.float32),
-            torch.rand([1, 256, 15, 25], dtype=torch.float32),
-            torch.rand([1, 256, 8, 13], dtype=torch.float32)
+            torch.rand([1, 32, 116, 200], dtype=torch.float32),
+            torch.rand([1, 32, 58, 100], dtype=torch.float32),
+            torch.rand([1, 32, 29, 50], dtype=torch.float32),
+            torch.rand([1, 32, 15, 25], dtype=torch.float32),
+            torch.rand([1, 32, 8, 13], dtype=torch.float32)
         ]
 
         # Test forward

tests/test_models/test_dense_heads/test_freeanchors.py

@@ -19,6 +19,14 @@ class TestFreeAnchor(unittest.TestCase):
         freeanchor_cfg = _get_detector_cfg(
             'free_anchor/pointpillars_hv_regnet-1.6gf_fpn_head-free-anchor'
             '_sbn-all_8xb4-2x_nus-3d.py')
+        # decrease channels to reduce cuda memory.
+        freeanchor_cfg.pts_voxel_encoder.feat_channels = [1, 1]
+        freeanchor_cfg.pts_middle_encoder.in_channels = 1
+        freeanchor_cfg.pts_backbone.base_channels = 1
+        freeanchor_cfg.pts_backbone.stem_channels = 1
+        freeanchor_cfg.pts_neck.out_channels = 1
+        freeanchor_cfg.pts_bbox_head.feat_channels = 1
+        freeanchor_cfg.pts_bbox_head.in_channels = 1
         model = MODELS.build(freeanchor_cfg)
         num_gt_instance = 3
         packed_inputs = _create_detector_inputs(

tests/test_models/test_dense_heads/test_ssn.py

@@ -18,6 +18,14 @@ class TestSSN(unittest.TestCase):
         _setup_seed(0)
         ssn_cfg = _get_detector_cfg(
             'ssn/ssn_hv_secfpn_sbn-all_16xb2-2x_nus-3d.py')
+        ssn_cfg.pts_voxel_encoder.feat_channels = [1, 1]
+        ssn_cfg.pts_middle_encoder.in_channels = 1
+        ssn_cfg.pts_backbone.in_channels = 1
+        ssn_cfg.pts_backbone.out_channels = [1, 1, 1]
+        ssn_cfg.pts_neck.in_channels = [1, 1, 1]
+        ssn_cfg.pts_neck.out_channels = [1, 1, 1]
+        ssn_cfg.pts_bbox_head.in_channels = 3
+        ssn_cfg.pts_bbox_head.feat_channels = 1
         model = MODELS.build(ssn_cfg)
         num_gt_instance = 50
         packed_inputs = _create_detector_inputs(