add lidarseg config && support torchsparse++

mmdet3d/configs/_base_/datasets/nus_seg.py

@@ -0,0 +1,221 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmcv.transforms import TestTimeAug
+from mmengine.dataset.sampler import DefaultSampler
+from mmengine.visualization import LocalVisBackend
+
+from mmdet3d.datasets import NuScenesSegDataset
+from mmdet3d.datasets.transforms import (GlobalRotScaleTrans,
+                                         LoadAnnotations3D, LoadPointsFromFile,
+                                         Pack3DDetInputs, PointSegClassMapping,
+                                         RandomFlip3D)
+from mmdet3d.evaluation import SegMetric
+from mmdet3d.models import Seg3DTTAModel
+from mmdet3d.visualization import Det3DLocalVisualizer
+
+# For nuScenes we usually do 16-class segmentation.
+# For labels_map we follow the uniform format of MMDetection & MMSegmentation
+# i.e. we consider the unlabeled class as the last one, which is different
+# from the original implementation of some methods e.g. Cylinder3D.
+data_root = 'data/nuscenes/'
+class_names = [
+    'barrier', 'bicycle', 'bus', 'car', 'construction_vehicle', 'motorcycle',
+    'pedestrian', 'traffic_cone', 'trailer', 'truck', 'driveable_surface',
+    'other_flat', 'sidewalk', 'terrain', 'manmade', 'vegetation'
+]
+labels_map = {
+    0: 16,
+    1: 16,
+    2: 6,
+    3: 6,
+    4: 6,
+    5: 16,
+    6: 6,
+    7: 16,
+    8: 16,
+    9: 0,
+    10: 16,
+    11: 16,
+    12: 7,
+    13: 16,
+    14: 1,
+    15: 2,
+    16: 2,
+    17: 3,
+    18: 4,
+    19: 16,
+    20: 16,
+    21: 5,
+    22: 8,
+    23: 9,
+    24: 10,
+    25: 11,
+    26: 12,
+    27: 13,
+    28: 14,
+    29: 16,
+    30: 15,
+    31: 16
+}
+
+metainfo = dict(
+    classes=class_names, seg_label_mapping=labels_map, max_label=31)
+
+input_modality = dict(use_lidar=True, use_camera=False)
+data_prefix = dict(
+    pts='samples/LIDAR_TOP',
+    img='',
+    pts_semantic_mask='lidarseg/v1.0-trainval')
+
+# Example to use different file client
+# Method 1: simply set the data root and let the file I/O module
+# automatically infer from prefix (not support LMDB and Memcache yet)
+
+# data_root = 's3://openmmlab/datasets/detection3d/nuscenes/'
+
+# Method 2: Use backend_args, file_client_args in versions before 1.1.0
+# backend_args = dict(
+#     backend='petrel',
+#     path_mapping=dict({
+#         './data/': 's3://openmmlab/datasets/detection3d/',
+#          'data/': 's3://openmmlab/datasets/detection3d/'
+#      }))
+backend_args = None
+
+train_pipeline = [
+    dict(
+        type=LoadPointsFromFile,
+        coord_type='LIDAR',
+        load_dim=5,
+        use_dim=4,
+        backend_args=backend_args),
+    dict(
+        type=LoadAnnotations3D,
+        with_bbox_3d=False,
+        with_label_3d=False,
+        with_seg_3d=True,
+        seg_3d_dtype='np.uint8',
+        backend_args=backend_args),
+    dict(type=PointSegClassMapping),
+    dict(
+        type=RandomFlip3D,
+        sync_2d=False,
+        flip_ratio_bev_horizontal=0.5,
+        flip_ratio_bev_vertical=0.5),
+    dict(
+        type=GlobalRotScaleTrans,
+        rot_range=[-0.78539816, 0.78539816],
+        scale_ratio_range=[0.95, 1.05],
+        translation_std=[0.1, 0.1, 0.1]),
+    dict(type=Pack3DDetInputs, keys=['points', 'pts_semantic_mask'])
+]
+test_pipeline = [
+    dict(
+        type=LoadPointsFromFile,
+        coord_type='LIDAR',
+        load_dim=5,
+        use_dim=4,
+        backend_args=backend_args),
+    dict(
+        type=LoadAnnotations3D,
+        with_bbox_3d=False,
+        with_label_3d=False,
+        with_seg_3d=True,
+        seg_3d_dtype='np.uint8',
+        backend_args=backend_args),
+    dict(type=PointSegClassMapping),
+    dict(type=Pack3DDetInputs, keys=['points'])
+]
+tta_pipeline = [
+    dict(
+        type=LoadPointsFromFile,
+        coord_type='LIDAR',
+        load_dim=5,
+        use_dim=4,
+        backend_args=backend_args),
+    dict(
+        type=LoadAnnotations3D,
+        with_bbox_3d=False,
+        with_label_3d=False,
+        with_seg_3d=True,
+        seg_3d_dtype='np.uint8',
+        backend_args=backend_args),
+    dict(type=PointSegClassMapping),
+    dict(
+        type=TestTimeAug,
+        transforms=[[
+            dict(
+                type=RandomFlip3D,
+                sync_2d=False,
+                flip_ratio_bev_horizontal=0.,
+                flip_ratio_bev_vertical=0.),
+            dict(
+                type=RandomFlip3D,
+                sync_2d=False,
+                flip_ratio_bev_horizontal=0.,
+                flip_ratio_bev_vertical=1.),
+            dict(
+                type=RandomFlip3D,
+                sync_2d=False,
+                flip_ratio_bev_horizontal=1.,
+                flip_ratio_bev_vertical=0.),
+            dict(
+                type=RandomFlip3D,
+                sync_2d=False,
+                flip_ratio_bev_horizontal=1.,
+                flip_ratio_bev_vertical=1.)
+        ],
+                    [
+                        dict(
+                            type=GlobalRotScaleTrans,
+                            rot_range=[pcd_rotate_range, pcd_rotate_range],
+                            scale_ratio_range=[
+                                pcd_scale_factor, pcd_scale_factor
+                            ],
+                            translation_std=[0, 0, 0])
+                        for pcd_rotate_range in [-0.78539816, 0.0, 0.78539816]
+                        for pcd_scale_factor in [0.95, 1.0, 1.05]
+                    ], [dict(type=Pack3DDetInputs, keys=['points'])]])
+]
+
+train_dataloader = dict(
+    batch_size=2,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    dataset=dict(
+        type=NuScenesSegDataset,
+        data_root=data_root,
+        ann_file='nuscenes_infos_train.pkl',
+        data_prefix=data_prefix,
+        pipeline=train_pipeline,
+        metainfo=metainfo,
+        modality=input_modality,
+        ignore_index=16,
+        backend_args=backend_args))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=1,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type=DefaultSampler, shuffle=False),
+    dataset=dict(
+        type=NuScenesSegDataset,
+        data_root=data_root,
+        ann_file='nuscenes_infos_val.pkl',
+        data_prefix=data_prefix,
+        pipeline=test_pipeline,
+        metainfo=metainfo,
+        modality=input_modality,
+        ignore_index=16,
+        test_mode=True,
+        backend_args=backend_args))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type=SegMetric)
+test_evaluator = val_evaluator
+
+vis_backends = [dict(type=LocalVisBackend)]
+visualizer = dict(
+    type=Det3DLocalVisualizer, vis_backends=vis_backends, name='visualizer')
+
+tta_model = dict(type=Seg3DTTAModel)

mmdet3d/configs/_base_/schedules/lidarseg_50e.py

@@ -0,0 +1,34 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.optim.optimizer.optimizer_wrapper import OptimWrapper
+from mmengine.optim.scheduler.lr_scheduler import OneCycleLR
+from mmengine.runner.loops import EpochBasedTrainLoop, TestLoop, ValLoop
+from torch.optim.adamw import AdamW
+
+# training schedule for 50e
+train_cfg = dict(type=EpochBasedTrainLoop, max_epochs=50, val_interval=1)
+val_cfg = dict(type=ValLoop)
+test_cfg = dict(type=TestLoop)
+
+# learning rate
+lr = 0.01
+param_scheduler = [
+    dict(
+        type=OneCycleLR,
+        eta_max=lr,
+        pct_start=0.2,
+        div_factor=25.0,
+        final_div_factor=100.0,
+        by_epoch=False)
+]
+
+# optimizer
+optim_wrapper = dict(
+    type=OptimWrapper,
+    optimizer=dict(
+        type=AdamW, lr=lr, betas=(0.9, 0.999), weight_decay=0.01, eps=1e-6))
+
+# Default setting for scaling LR automatically
+#   - `enable` means enable scaling LR automatically
+#       or not by default.
+#   - `base_batch_size` = (8 GPUs) x (2 samples per GPU).
+auto_scale_lr = dict(enable=False, base_batch_size=16)

mmdet3d/configs/_base_/schedules/lidarseg_80e.py

@@ -0,0 +1,34 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.optim.optimizer.optimizer_wrapper import OptimWrapper
+from mmengine.optim.scheduler.lr_scheduler import OneCycleLR
+from mmengine.runner.loops import EpochBasedTrainLoop, TestLoop, ValLoop
+from torch.optim.adamw import AdamW
+
+# training schedule for 80e
+train_cfg = dict(type=EpochBasedTrainLoop, max_epochs=80, val_interval=1)
+val_cfg = dict(type=ValLoop)
+test_cfg = dict(type=TestLoop)
+
+# learning rate
+lr = 0.01
+param_scheduler = [
+    dict(
+        type=OneCycleLR,
+        eta_max=lr,
+        pct_start=0.2,
+        div_factor=25.0,
+        final_div_factor=100.0,
+        by_epoch=False)
+]
+
+# optimizer
+optim_wrapper = dict(
+    type=OptimWrapper,
+    optimizer=dict(
+        type=AdamW, lr=lr, betas=(0.9, 0.999), weight_decay=0.01, eps=1e-6))
+
+# Default setting for scaling LR automatically
+#   - `enable` means enable scaling LR automatically
+#       or not by default.
+#   - `base_batch_size` = (8 GPUs) x (2 samples per GPU).
+auto_scale_lr = dict(enable=False, base_batch_size=16)

mmdet3d/configs/lidarseg_benchmark/cylinder3d_8xb2_50e_semantickitti.py

@@ -0,0 +1,17 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine import read_base
+
+with read_base():
+    from .._base_.datasets.semantickitti import *
+    from .._base_.models.cylinder3d import *
+    from .._base_.schedules.lidarseg_50e import *
+    from .._base_.default_runtime import *
+
+from mmengine.hooks.checkpoint_hook import CheckpointHook
+from mmengine.visualization.vis_backend import LocalVisBackend, WandbVisBackend
+
+visualizer.update(
+    dict(vis_backends=[dict(type=LocalVisBackend),
+                       dict(type=WandbVisBackend)]))
+default_hooks.update(
+    dict(checkpoint=dict(type=CheckpointHook, save_best='miou')))

mmdet3d/configs/lidarseg_benchmark/cylinder3d_8xb2_80e_nus_seg.py

@@ -0,0 +1,23 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine import read_base
+
+with read_base():
+    from .._base_.datasets.nus_seg import *
+    from .._base_.models.cylinder3d import *
+    from .._base_.schedules.lidarseg_80e import *
+    from .._base_.default_runtime import *
+
+from mmengine.hooks.checkpoint_hook import CheckpointHook
+from mmengine.visualization.vis_backend import LocalVisBackend, WandbVisBackend
+
+model.update(dict(decode_head=dict(num_classes=17, ignore_index=16)))
+
+train_dataloader.update(
+    dict(dataset=dict(ann_file='nuscenes_lidarseg_infos_train.pkl')))
+test_dataloader.update(
+    dict(dataset=dict(ann_file='nuscenes_lidarseg_infos_val.pkl')))
+visualizer.update(
+    dict(vis_backends=[dict(type=LocalVisBackend),
+                       dict(type=WandbVisBackend)]))
+default_hooks.update(
+    dict(checkpoint=dict(type=CheckpointHook, save_best='miou')))

mmdet3d/configs/lidarseg_benchmark/cylinder3d_8xb2_amp_50e_semantickitti.py

@@ -0,0 +1,7 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine import read_base
+
+with read_base():
+    from .cylinder3d_8xb2_50e_semantickitti import *
+
+optim_wrapper.update(dict(type='AmpOptimWrapper', loss_scale='dynamic'))

mmdet3d/configs/lidarseg_benchmark/cylinder3d_8xb2_amp_80e_nus_seg.py

@@ -0,0 +1,7 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine import read_base
+
+with read_base():
+    from .cylinder3d_8xb2_80e_nus_seg import *
+
+optim_wrapper.update(dict(type='AmpOptimWrapper', loss_scale='dynamic'))