[Fix]update pp cfg and fix MultiScaleFlipAug3D

configs/_base_/datasets/kitti-3d-3class.py

@@ -71,9 +71,9 @@ test_pipeline = [
                 translation_std=[0, 0, 0]),
             dict(type='RandomFlip3D'),
             dict(
-                type='PointsRangeFilter', point_cloud_range=point_cloud_range),
-            dict(type='Pack3DDetInputs', keys=['points']),
-        ])
+                type='PointsRangeFilter', point_cloud_range=point_cloud_range)
+        ]),
+    dict(type='Pack3DDetInputs', keys=['points'])
 ]
 # construct a pipeline for data and gt loading in show function
 # please keep its loading function consistent with test_pipeline (e.g. client)
@@ -117,13 +117,27 @@ val_dataloader = dict(
         data_root=data_root,
         data_prefix=dict(pts='training/velodyne_reduced'),
         ann_file='kitti_infos_val.pkl',
-        pipeline=eval_pipeline,
+        pipeline=test_pipeline,
+        modality=input_modality,
+        test_mode=True,
+        metainfo=metainfo,
+        box_type_3d='LiDAR'))
+test_dataloader = dict(
+    batch_size=1,
+    num_workers=1,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(pts='training/velodyne_reduced'),
+        ann_file='kitti_infos_val.pkl',
+        pipeline=test_pipeline,
         modality=input_modality,
         test_mode=True,
         metainfo=metainfo,
         box_type_3d='LiDAR'))
-test_dataloader = val_dataloader
-test_dataloader['dataset']['pipeline'] = test_pipeline
 val_evaluator = dict(
     type='KittiMetric',
     ann_file=data_root + 'kitti_infos_val.pkl',

configs/_base_/datasets/kitti-3d-car.py

@@ -115,13 +115,27 @@ val_dataloader = dict(
         data_root=data_root,
         data_prefix=dict(pts='training/velodyne_reduced'),
         ann_file='kitti_infos_val.pkl',
-        pipeline=eval_pipeline,
+        pipeline=test_pipeline,
+        modality=input_modality,
+        test_mode=True,
+        metainfo=metainfo,
+        box_type_3d='LiDAR'))
+test_dataloader = dict(
+    batch_size=1,
+    num_workers=1,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(pts='training/velodyne_reduced'),
+        ann_file='kitti_infos_val.pkl',
+        pipeline=test_pipeline,
         modality=input_modality,
         test_mode=True,
         metainfo=metainfo,
         box_type_3d='LiDAR'))
-test_dataloader = val_dataloader
-test_dataloader['dataset']['pipeline'] = test_pipeline
 val_evaluator = dict(
     type='KittiMetric',
     ann_file=data_root + 'kitti_infos_val.pkl',

configs/_base_/schedules/cyclic_20e.py

@@ -4,7 +4,6 @@
 # use a default schedule.
 # optimizer
 lr = 1e-4
-iter_num_in_epoch = 3712
 # This schedule is mainly used by models on nuScenes dataset
 # max_norm=10 is better for SECOND
 optim_wrapper = dict(
@@ -15,35 +14,39 @@ optim_wrapper = dict(
 param_scheduler = [
     dict(
         type='CosineAnnealingLR',
-        T_max=8 * iter_num_in_epoch,
+        T_max=8,
         eta_min=lr * 10,
-        by_epoch=False,
         begin=0,
-        end=8 * iter_num_in_epoch),
+        end=8,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingLR',
-        T_max=12 * iter_num_in_epoch,
+        T_max=12,
         eta_min=lr * 1e-4,
-        by_epoch=False,
-        begin=8 * iter_num_in_epoch,
-        end=20 * iter_num_in_epoch),
+        begin=8,
+        end=20,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=8 * iter_num_in_epoch,
+        T_max=8,
         eta_min=0.85 / 0.95,
-        by_epoch=False,
         begin=0,
-        end=8 * iter_num_in_epoch),
+        end=8,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=12 * iter_num_in_epoch,
+        T_max=12,
         eta_min=1,
-        by_epoch=False,
-        begin=8 * iter_num_in_epoch,
-        end=20 * iter_num_in_epoch)
+        begin=8,
+        end=20,
+        by_epoch=True,
+        convert_to_iter_based=True)
 ]
 
 # runtime settings
-train_cfg = dict(by_epoch=True, max_epochs=20)
-val_cfg = dict(interval=1)
+train_cfg = dict(by_epoch=True, max_epochs=20, val_interval=1)
+val_cfg = dict()
 test_cfg = dict()

configs/_base_/schedules/cyclic_40e.py

@@ -3,7 +3,6 @@
 # rather than the max learning rate. Since the target_ratio is (10, 1e-4),
 # the learning rate will change from 0.0018 to 0.018, than go to 0.0018*1e-4
 lr = 0.0018
-iter_num_in_epoch = 3712
 # The optimizer follows the setting in SECOND.Pytorch, but here we use
 # the official AdamW optimizer implemented by PyTorch.
 optim_wrapper = dict(
@@ -14,38 +13,42 @@ optim_wrapper = dict(
 param_scheduler = [
     dict(
         type='CosineAnnealingLR',
-        T_max=16 * iter_num_in_epoch,
+        T_max=16,
         eta_min=lr * 10,
-        by_epoch=False,
         begin=0,
-        end=16 * iter_num_in_epoch),
+        end=16,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingLR',
-        T_max=24 * iter_num_in_epoch,
+        T_max=24,
         eta_min=lr * 1e-4,
-        by_epoch=False,
-        begin=16 * iter_num_in_epoch,
-        end=40 * iter_num_in_epoch),
+        begin=16,
+        end=40,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=16 * iter_num_in_epoch,
+        T_max=16,
         eta_min=0.85 / 0.95,
-        by_epoch=False,
         begin=0,
-        end=16 * iter_num_in_epoch),
+        end=16,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=24 * iter_num_in_epoch,
+        T_max=24,
         eta_min=1,
-        by_epoch=False,
-        begin=16 * iter_num_in_epoch,
-        end=40 * iter_num_in_epoch)
+        begin=16,
+        end=40,
+        by_epoch=True,
+        convert_to_iter_based=True)
 ]
 
 # Runtime settings，training schedule for 40e
 # Although the max_epochs is 40, this schedule is usually used we
 # RepeatDataset with repeat ratio N, thus the actual max epoch
 # number could be Nx40
-train_cfg = dict(by_epoch=True, max_epochs=40)
-val_cfg = dict(interval=1)
+train_cfg = dict(by_epoch=True, max_epochs=40, val_interval=1)
+val_cfg = dict()
 test_cfg = dict()

configs/pointpillars/hv_pointpillars_secfpn_6x8_160e_kitti-3d-3class.py

@@ -52,51 +52,53 @@ test_pipeline = [
                 translation_std=[0, 0, 0]),
             dict(type='RandomFlip3D'),
             dict(
-                type='PointsRangeFilter', point_cloud_range=point_cloud_range),
-            dict(type='Pack3DDetInputs', keys=['points'])
-        ])
+                type='PointsRangeFilter', point_cloud_range=point_cloud_range)
+        ]),
+    dict(type='Pack3DDetInputs', keys=['points'])
 ]
 
 train_dataloader = dict(
     dataset=dict(dataset=dict(pipeline=train_pipeline, metainfo=metainfo)))
-test_dataloader = dict(dataset=dict(metainfo=metainfo))
-val_dataloader = dict(dataset=dict(metainfo=metainfo))
+test_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))
+val_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))
 # In practice PointPillars also uses a different schedule
 # optimizer
 lr = 0.001
 epoch_num = 80
-iter_num_in_epoch = 3712
 optim_wrapper = dict(
     optimizer=dict(lr=lr), clip_grad=dict(max_norm=35, norm_type=2))
 param_scheduler = [
     dict(
         type='CosineAnnealingLR',
-        T_max=epoch_num * 0.4 * iter_num_in_epoch,
+        T_max=epoch_num * 0.4,
         eta_min=lr * 10,
-        by_epoch=False,
         begin=0,
-        end=epoch_num * 0.4 * iter_num_in_epoch),
+        end=epoch_num * 0.4,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingLR',
-        T_max=epoch_num * 0.6 * iter_num_in_epoch,
+        T_max=epoch_num * 0.6,
         eta_min=lr * 1e-4,
-        by_epoch=False,
-        begin=epoch_num * 0.4 * iter_num_in_epoch,
-        end=epoch_num * 1 * iter_num_in_epoch),
+        begin=epoch_num * 0.4,
+        end=epoch_num * 1,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=epoch_num * 0.4 * iter_num_in_epoch,
+        T_max=epoch_num * 0.4,
         eta_min=0.85 / 0.95,
-        by_epoch=False,
         begin=0,
-        end=epoch_num * 0.4 * iter_num_in_epoch),
+        end=epoch_num * 0.4,
+        by_epoch=True,
+        convert_to_iter_based=True),
     dict(
         type='CosineAnnealingBetas',
-        T_max=epoch_num * 0.6 * iter_num_in_epoch,
+        T_max=epoch_num * 0.6,
         eta_min=1,
-        by_epoch=False,
-        begin=epoch_num * 0.4 * iter_num_in_epoch,
-        end=epoch_num * 1 * iter_num_in_epoch)
+        begin=epoch_num * 0.4,
+        end=epoch_num * 1,
+        convert_to_iter_based=True)
 ]
 # max_norm=35 is slightly better than 10 for PointPillars in the earlier
 # development of the codebase thus we keep the setting. But we does not
@@ -104,6 +106,6 @@ param_scheduler = [
 # PointPillars usually need longer schedule than second, we simply double
 # the training schedule. Do remind that since we use RepeatDataset and
 # repeat factor is 2, so we actually train 160 epochs.
-train_cfg = dict(by_epoch=True, max_epochs=epoch_num)
-val_cfg = dict(interval=2)
+train_cfg = dict(by_epoch=True, max_epochs=epoch_num, val_interval=2)
+val_cfg = dict()
 test_cfg = dict()

configs/pointpillars/hv_pointpillars_secfpn_6x8_160e_kitti-3d-car.py

 # model settings
 _base_ = './hv_pointpillars_secfpn_6x8_160e_kitti-3d-3class.py'
+# dataset settings
+dataset_type = 'KittiDataset'
+data_root = 'data/kitti/'
+class_names = ['Car']
+metainfo = dict(CLASSES=class_names)
 
 point_cloud_range = [0, -39.68, -3, 69.12, 39.68, 1]
+
 model = dict(
     bbox_head=dict(
         type='Anchor3DHead',
@@ -26,11 +32,7 @@ model = dict(
         allowed_border=0,
         pos_weight=-1,
         debug=False))
-# dataset settings
-dataset_type = 'KittiDataset'
-data_root = 'data/kitti/'
-class_names = ['Car']
-metainfo = dict(CLASSES=class_names)
+
 db_sampler = dict(
     data_root=data_root,
     info_path=data_root + 'kitti_dbinfos_train.pkl',
@@ -70,15 +72,12 @@ test_pipeline = [
                 translation_std=[0, 0, 0]),
             dict(type='RandomFlip3D'),
             dict(
-                type='PointsRangeFilter', point_cloud_range=point_cloud_range),
-            dict(type='Pack3DDetInputs', keys=['points'])
-        ])
+                type='PointsRangeFilter', point_cloud_range=point_cloud_range)
+        ]),
+    dict(type='Pack3DDetInputs', keys=['points'])
 ]
 
 train_dataloader = dict(
-    dataset=dict(
-        type='RepeatDataset',
-        times=2,
-        dataset=dict(pipeline=train_pipeline, metainfo=metainfo)))
-test_dataloader = dict(dataset=dict(metainfo=metainfo))
-val_dataloader = dict(dataset=dict(metainfo=metainfo))
+    dataset=dict(dataset=dict(pipeline=train_pipeline, metainfo=metainfo)))
+test_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))
+val_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))

mmdet3d/datasets/pipelines/formating.py

@@ -69,6 +69,9 @@ class Pack3DDetInputs(BaseTransform):
         """
         # augtest
         if isinstance(results, list):
+            if len(results) == 1:
+                # simple test
+                return self.pack_single_results(results[0])
             pack_results = []
             for single_result in results:
                 pack_results.append(self.pack_single_results(single_result))
@@ -113,7 +116,8 @@ class Pack3DDetInputs(BaseTransform):
                 img = results['img']
                 if len(img.shape) < 3:
                     img = np.expand_dims(img, -1)
-                results['img'] = np.ascontiguousarray(img.transpose(2, 0, 1))
+                results['img'] = to_tensor(
+                    np.ascontiguousarray(img.transpose(2, 0, 1)))
 
         for key in [
                 'proposals', 'gt_bboxes', 'gt_bboxes_ignore', 'gt_labels',

mmdet3d/datasets/pipelines/test_time_aug.py

@@ -88,6 +88,8 @@ class MultiScaleFlipAug3D(BaseTransform):
         pcd_vertical_flip_aug = [False, True] \
             if self.flip and self.pcd_vertical_flip else [False]
         for scale in self.img_scale:
+            # TODO refactor according to augtest docs
+            self.transforms.transforms[0].scale = scale
             for pts_scale_ratio in self.pts_scale_ratio:
                 for flip in flip_aug:
                     for pcd_horizontal_flip in pcd_horizontal_flip_aug:

mmdet3d/metrics/kitti_metric.py

@@ -159,7 +159,7 @@ class KittiMetric(BaseMetric):
             for pred_result in pred:
                 for attr_name in pred[pred_result]:
                     pred[pred_result][attr_name] = pred[pred_result][
-                        attr_name].to('cpu')
+                        attr_name].to(self.collect_device)
                 result[pred_result] = pred[pred_result]
                 sample_idx = data['data_sample']['sample_idx']
                 result['sample_idx'] = sample_idx

mmdet3d/models/detectors/base.py

@@ -21,9 +21,10 @@ class Base3DDetector(BaseDetector):
     """
 
     def __init__(self,
-                 data_processor: OptConfigType = None,
+                 data_preprocessor: OptConfigType = None,
                  init_cfg: OptMultiConfig = None) -> None:
-        super().__init__(data_preprocessor=data_processor, init_cfg=init_cfg)
+        super().__init__(
+            data_preprocessor=data_preprocessor, init_cfg=init_cfg)
 
     def forward(self,
                 inputs: Union[dict, List[dict]],

mmdet3d/models/detectors/single_stage.py

@@ -42,7 +42,8 @@ class SingleStage3DDetector(Base3DDetector):
                  test_cfg: OptConfigType = None,
                  data_preprocessor: OptConfigType = None,
                  init_cfg: OptMultiConfig = None) -> None:
-        super().__init__(data_processor=data_preprocessor, init_cfg=init_cfg)
+        super().__init__(
+            data_preprocessor=data_preprocessor, init_cfg=init_cfg)
         self.backbone = MODELS.build(backbone)
         if neck is not None:
             self.neck = MODELS.build(neck)