Skip to content

GitLab

Commit d2b71343 authored by 雍大凯's avatar 雍大凯
Browse files

add code

parent 69e57885
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 4717 additions and 0 deletions
docker-hub/FlashOCC/Flashocc/lib/dvr/dvr.cu 0 → 100644
View file @ d2b71343
This diff is collapsed.
docker-hub/FlashOCC/Flashocc/lib/dvr/dvr.hip 0 → 100644
View file @ d2b71343
This diff is collapsed.
mmdetection3d @ c9541b0d
Subproject commit c9541b0db89498fdea5cafd05b7b17f7b625b858
docker-hub/FlashOCC/Flashocc/projects/configs/bevdet_occ/bevdet-occ-r50-4d-stereo.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (256, 704),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 0.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 32
multi_adj_frame_id_cfg = (1, 1+1, 1)
model = dict(
type='BEVStereo4DOCC',
align_after_view_transfromation=False,
num_adj=len(range(*multi_adj_frame_id_cfg)),
img_backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
with_cp=True,
style='pytorch'),
img_neck=dict(
type='CustomFPN',
in_channels=[1024, 2048],
out_channels=256,
num_outs=1,
start_level=0,
out_ids=[0]),
img_view_transformer=dict(
type='LSSViewTransformerBEVStereo',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=256,
out_channels=numC_Trans,
sid=False,
collapse_z=False,
loss_depth_weight=0.05,
depthnet_cfg=dict(use_dcn=False,
aspp_mid_channels=96,
stereo=True,
bias=5.),
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet3D',
numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
num_layer=[1, 2, 4],
with_cp=False,
num_channels=[numC_Trans, numC_Trans*2, numC_Trans*4],
stride=[1, 2, 2],
backbone_output_ids=[0, 1, 2]),
img_bev_encoder_neck=dict(type='LSSFPN3D',
in_channels=numC_Trans*7,
out_channels=numC_Trans),
pre_process=dict(
type='CustomResNet3D',
numC_input=numC_Trans,
with_cp=False,
num_layer=[1, ],
num_channels=[numC_Trans, ],
stride=[1, ],
backbone_output_ids=[0, ]),
occ_head=dict(
type='BEVOCCHead3D',
in_dim=numC_Trans,
out_dim=32,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar', 'mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
data_root=data_root,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet4d',
multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4,
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
]
load_from = "ckpts/bevdet-r50-4d-stereo-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
evaluation = dict(interval=1, start=20, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=5)
# with_pretrain:
# align_after_view_transfromation=False
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 8.22
# ===> barrier - IoU = 44.21
# ===> bicycle - IoU = 10.34
# ===> bus - IoU = 42.08
# ===> car - IoU = 49.63
# ===> construction_vehicle - IoU = 23.37
# ===> motorcycle - IoU = 17.41
# ===> pedestrian - IoU = 21.49
# ===> traffic_cone - IoU = 19.7
# ===> trailer - IoU = 31.33
# ===> truck - IoU = 37.09
# ===> driveable_surface - IoU = 80.13
# ===> other_flat - IoU = 37.37
# ===> sidewalk - IoU = 50.41
# ===> terrain - IoU = 54.29
# ===> manmade - IoU = 45.56
# ===> vegetation - IoU = 39.59
# ===> mIoU of 6019 samples: 36.01
docker-hub/FlashOCC/Flashocc/projects/configs/bevdet_occ/bevdet-occ-r50.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (256, 704),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 0.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 32
model = dict(
type='BEVDetOCC',
img_backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
with_cp=True,
style='pytorch',
pretrained='torchvision://resnet50',
),
img_neck=dict(
type='CustomFPN',
in_channels=[1024, 2048],
out_channels=256,
num_outs=1,
start_level=0,
out_ids=[0]),
img_view_transformer=dict(
type='LSSViewTransformer',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=256,
out_channels=numC_Trans,
sid=False,
collapse_z=False,
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet3D',
numC_input=numC_Trans,
num_layer=[1, 2, 4],
with_cp=False,
num_channels=[numC_Trans, numC_Trans*2, numC_Trans*4],
stride=[1, 2, 2],
backbone_output_ids=[0, 1, 2]),
img_bev_encoder_neck=dict(type='LSSFPN3D',
in_channels=numC_Trans*7,
out_channels=numC_Trans),
occ_head=dict(
type='BEVOCCHead3D',
in_dim=numC_Trans,
out_dim=32,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar', 'mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
data_root=data_root,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet',
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4,
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
]
load_from = "ckpts/bevdet-r50-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
evaluation = dict(interval=1, start=20, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=5)
# with pretrain
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 6.65
# ===> barrier - IoU = 36.97
# ===> bicycle - IoU = 8.33
# ===> bus - IoU = 38.69
# ===> car - IoU = 44.46
# ===> construction_vehicle - IoU = 15.21
# ===> motorcycle - IoU = 13.67
# ===> pedestrian - IoU = 16.39
# ===> traffic_cone - IoU = 15.27
# ===> trailer - IoU = 27.11
# ===> truck - IoU = 31.04
# ===> driveable_surface - IoU = 78.7
# ===> other_flat - IoU = 36.45
# ===> sidewalk - IoU = 48.27
# ===> terrain - IoU = 51.68
# ===> manmade - IoU = 36.82
# ===> vegetation - IoU = 32.09
# ===> mIoU of 6019 samples: 31.64
# with det pretrain; use_mask=False; class_balance=True
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 4.36
# ===> barrier - IoU = 28.87
# ===> bicycle - IoU = 2.86
# ===> bus - IoU = 29.27
# ===> car - IoU = 32.45
# ===> construction_vehicle - IoU = 11.05
# ===> motorcycle - IoU = 12.82
# ===> pedestrian - IoU = 10.11
# ===> traffic_cone - IoU = 9.47
# ===> trailer - IoU = 7.93
# ===> truck - IoU = 21.58
# ===> driveable_surface - IoU = 49.85
# ===> other_flat - IoU = 25.5
# ===> sidewalk - IoU = 26.78
# ===> terrain - IoU = 21.14
# ===> manmade - IoU = 5.76
# ===> vegetation - IoU = 7.09
# ===> mIoU of 6019 samples: 18.05
\ No newline at end of file
docker-hub/FlashOCC/Flashocc/projects/configs/bevdet_occ/bevdet-occ-stbase-4d-stereo-512x1408.py 0 → 100644
View file @ d2b71343
# Copyright (c) Phigent Robotics. All rights reserved.
# align_after_view_transfromation=True
# align_after_view_transfromation=False
# 1x/12epoch
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 10.12
# ===> barrier - IoU = 48.06
# ===> bicycle - IoU = 0.0
# ===> bus - IoU = 51.19
# ===> car - IoU = 53.61
# ===> construction_vehicle - IoU = 27.15
# ===> motorcycle - IoU = 2.74
# ===> pedestrian - IoU = 28.3
# ===> traffic_cone - IoU = 23.33
# ===> trailer - IoU = 36.24
# ===> truck - IoU = 42.13
# ===> driveable_surface - IoU = 81.77
# ===> other_flat - IoU = 42.43
# ===> sidewalk - IoU = 53.67
# ===> terrain - IoU = 57.31
# ===> manmade - IoU = 48.27
# ===> vegetation - IoU = 43.31
# ===> mIoU of 6019 samples: 38.21
# 2x/24epoch
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 12.15
# ===> barrier - IoU = 49.63
# ===> bicycle - IoU = 25.1
# ===> bus - IoU = 52.02
# ===> car - IoU = 54.46
# ===> construction_vehicle - IoU = 27.87
# ===> motorcycle - IoU = 27.99
# ===> pedestrian - IoU = 28.94
# ===> traffic_cone - IoU = 27.23
# ===> trailer - IoU = 36.43
# ===> truck - IoU = 42.22
# ===> driveable_surface - IoU = 82.31
# ===> other_flat - IoU = 43.29
# ===> sidewalk - IoU = 54.62
# ===> terrain - IoU = 57.9
# ===> manmade - IoU = 48.61
# ===> vegetation - IoU = 43.55
# ===> mIoU of 6019 samples: 42.02
# 3x/36epoch
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 12.37
# ===> barrier - IoU = 50.15
# ===> bicycle - IoU = 26.97
# ===> bus - IoU = 51.86
# ===> car - IoU = 54.65
# ===> construction_vehicle - IoU = 28.38
# ===> motorcycle - IoU = 28.96
# ===> pedestrian - IoU = 29.02
# ===> traffic_cone - IoU = 28.28
# ===> trailer - IoU = 37.05
# ===> truck - IoU = 42.52
# ===> driveable_surface - IoU = 82.55
# ===> other_flat - IoU = 43.15
# ===> sidewalk - IoU = 54.87
# ===> terrain - IoU = 58.33
# ===> manmade - IoU = 48.78
# ===> vegetation - IoU = 43.79
# ===> mIoU of 6019 samples: 42.45
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (512, 1408),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
# Model
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 0.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 32
multi_adj_frame_id_cfg = (1, 1+1, 1)
model = dict(
type='BEVStereo4DOCC',
align_after_view_transfromation=False,
num_adj=len(range(*multi_adj_frame_id_cfg)),
img_backbone=dict(
type='SwinTransformer',
pretrain_img_size=224,
patch_size=4,
window_size=12,
mlp_ratio=4,
embed_dims=128,
depths=[2, 2, 18, 2],
num_heads=[4, 8, 16, 32],
strides=(4, 2, 2, 2),
out_indices=(2, 3),
qkv_bias=True,
qk_scale=None,
patch_norm=True,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.1,
use_abs_pos_embed=False,
return_stereo_feat=True,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN', requires_grad=True),
pretrain_style='official',
output_missing_index_as_none=False),
img_neck=dict(
type='FPN_LSS',
in_channels=512 + 1024,
out_channels=512,
# with_cp=False,
extra_upsample=None,
input_feature_index=(0, 1),
scale_factor=2),
img_view_transformer=dict(
type='LSSViewTransformerBEVStereo',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=512,
out_channels=numC_Trans,
sid=False,
collapse_z=False,
loss_depth_weight=0.05,
depthnet_cfg=dict(use_dcn=False,
aspp_mid_channels=96,
stereo=True,
bias=5.),
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet3D',
numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
num_layer=[1, 2, 4],
with_cp=False,
num_channels=[numC_Trans,numC_Trans*2,numC_Trans*4],
stride=[1,2,2],
backbone_output_ids=[0,1,2]),
img_bev_encoder_neck=dict(type='LSSFPN3D',
in_channels=numC_Trans*7,
out_channels=numC_Trans),
pre_process=dict(
type='CustomResNet3D',
numC_input=numC_Trans,
with_cp=False,
num_layer=[1,],
num_channels=[numC_Trans,],
stride=[1,],
backbone_output_ids=[0,]),
occ_head=dict(
type='BEVOCCHead3D',
in_dim=numC_Trans,
out_dim=32,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar','mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet4d',
multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=1, # with 32 GPU
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=2e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24,])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
dict(
type='SyncbnControlHook',
syncbn_start_epoch=0,
),
]
load_from="ckpts/bevdet-stbase-4d-stereo-512x1408-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-r50-4d-stereo.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams': 6,
'input_size': (256, 704),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 6.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 80
multi_adj_frame_id_cfg = (1, 1+1, 1)
model = dict(
type='BEVStereo4DOCC',
align_after_view_transfromation=False,
num_adj=len(range(*multi_adj_frame_id_cfg)),
img_backbone=dict(
pretrained='torchvision://resnet50',
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
with_cp=True,
style='pytorch'),
img_neck=dict(
type='CustomFPN',
in_channels=[1024, 2048],
out_channels=256,
num_outs=1,
start_level=0,
out_ids=[0]),
img_view_transformer=dict(
type='LSSViewTransformerBEVStereo',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=256,
out_channels=numC_Trans,
sid=True,
loss_depth_weight=0.05,
depthnet_cfg=dict(use_dcn=False,
aspp_mid_channels=96,
stereo=True,
bias=5.),
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet',
numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
img_bev_encoder_neck=dict(
type='FPN_LSS',
in_channels=numC_Trans * 8 + numC_Trans * 2,
out_channels=256),
pre_process=dict(
type='CustomResNet',
numC_input=numC_Trans,
num_layer=[1, ],
num_channels=[numC_Trans, ],
stride=[1, ],
backbone_output_ids=[0, ]),
occ_head=dict(
type='BEVOCCHead2D',
in_dim=256,
out_dim=256,
Dz=16,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar', 'mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
data_root=data_root,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet4d',
multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4,
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
]
load_from = "./ckpts/bevdet-r50-4d-stereo-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
evaluation = dict(interval=1, start=20, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=5)
# with_pretrain:
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 9.08
# ===> barrier - IoU = 46.32
# ===> bicycle - IoU = 17.71
# ===> bus - IoU = 42.7
# ===> car - IoU = 50.64
# ===> construction_vehicle - IoU = 23.72
# ===> motorcycle - IoU = 20.13
# ===> pedestrian - IoU = 22.34
# ===> traffic_cone - IoU = 24.09
# ===> trailer - IoU = 30.26
# ===> truck - IoU = 37.39
# ===> driveable_surface - IoU = 81.68
# ===> other_flat - IoU = 40.13
# ===> sidewalk - IoU = 52.34
# ===> terrain - IoU = 56.46
# ===> manmade - IoU = 47.69
# ===> vegetation - IoU = 40.6
# ===> mIoU of 6019 samples: 37.84
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-r50-M0-trt.py 0 → 100644
View file @ d2b71343
_base_ = ['./flashocc-r50-M0.py',
]
model = dict(
wocc=True,
wdet3d=False,
)
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-r50-M0.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (256, 704),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 6.4],
'depth': [1.0, 45.0, 1.0],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 64
model = dict(
type='BEVDetOCC',
img_backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
with_cp=True,
style='pytorch',
pretrained='torchvision://resnet50',
),
img_neck=dict(
type='CustomFPN',
in_channels=[1024, 2048],
out_channels=256,
num_outs=1,
start_level=0,
out_ids=[0]),
img_view_transformer=dict(
type='LSSViewTransformer',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=256,
out_channels=numC_Trans,
sid=False,
collapse_z=True,
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet',
numC_input=numC_Trans,
num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
img_bev_encoder_neck=dict(
type='FPN_LSS',
in_channels=numC_Trans * 8 + numC_Trans * 2,
out_channels=128),
occ_head=dict(
type='BEVOCCHead2D',
in_dim=128,
out_dim=128,
Dz=16,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5
)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar', 'mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
data_root=data_root,
classes=class_names,
modality=input_modality,
stereo=False,
filter_empty_gt=False,
img_info_prototype='bevdet',
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4,
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
]
load_from = "ckpts/bevdet-r50-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
evaluation = dict(interval=1, start=20, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=5)
# with det pretrain; use_mask=True; out_dim=256,
# ===> per class IoU of 6019 samples:
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 6.21
# ===> barrier - IoU = 39.56
# ===> bicycle - IoU = 11.27
# ===> bus - IoU = 36.31
# ===> car - IoU = 43.96
# ===> construction_vehicle - IoU = 16.25
# ===> motorcycle - IoU = 14.74
# ===> pedestrian - IoU = 16.89
# ===> traffic_cone - IoU = 15.76
# ===> trailer - IoU = 28.56
# ===> truck - IoU = 30.91
# ===> driveable_surface - IoU = 78.16
# ===> other_flat - IoU = 37.52
# ===> sidewalk - IoU = 47.42
# ===> terrain - IoU = 51.35
# ===> manmade - IoU = 36.79
# ===> vegetation - IoU = 31.42
# ===> mIoU of 6019 samples: 31.95
# {'mIoU': array([0.06207982, 0.39564533, 0.11270112, 0.36311426, 0.43955401,
# 0.16252583, 0.14739984, 0.16885096, 0.15757262, 0.28564777,
# 0.30909029, 0.7815907 , 0.37523904, 0.47420705, 0.51351759,
# 0.36789645, 0.31420157, 0.87802724])}
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-r50-trt.py 0 → 100644
View file @ d2b71343
_base_ = ['./flashocc-r50.py',
]
model = dict(
wocc=True,
wdet3d=False,
)
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-r50.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
point_cloud_range = [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (256, 704),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 6.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 64
model = dict(
type='BEVDetOCC',
img_backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
with_cp=True,
style='pytorch',
#pretrained='torchvision://resnet50',
),
img_neck=dict(
type='CustomFPN',
in_channels=[1024, 2048],
out_channels=256,
num_outs=1,
start_level=0,
out_ids=[0]),
img_view_transformer=dict(
type='LSSViewTransformer',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=256,
out_channels=numC_Trans,
sid=False,
collapse_z=True,
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet',
numC_input=numC_Trans,
num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
img_bev_encoder_neck=dict(
type='FPN_LSS',
in_channels=numC_Trans * 8 + numC_Trans * 2,
out_channels=256),
occ_head=dict(
type='BEVOCCHead2D',
in_dim=256,
out_dim=256, # out_dim=128 for M0!!!
Dz=16,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5
)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar', 'mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=False),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
data_root=data_root,
classes=class_names,
modality=input_modality,
stereo=False,
filter_empty_gt=False,
img_info_prototype='bevdet',
)
test_data_config = dict(
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=24,
workers_per_gpu=24,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
]
load_from = "ckpts/bevdet-r50-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
evaluation = dict(interval=1, start=20, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=5)
# with det pretrain; use_mask=True;
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 6.74
# ===> barrier - IoU = 37.65
# ===> bicycle - IoU = 10.26
# ===> bus - IoU = 39.55
# ===> car - IoU = 44.36
# ===> construction_vehicle - IoU = 14.88
# ===> motorcycle - IoU = 13.4
# ===> pedestrian - IoU = 15.79
# ===> traffic_cone - IoU = 15.38
# ===> trailer - IoU = 27.44
# ===> truck - IoU = 31.73
# ===> driveable_surface - IoU = 78.82
# ===> other_flat - IoU = 37.98
# ===> sidewalk - IoU = 48.7
# ===> terrain - IoU = 52.5
# ===> manmade - IoU = 37.89
# ===> vegetation - IoU = 32.24
# ===> mIoU of 6019 samples: 32.08
# with det pretrain; use_mask=False; class_balance=True
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 4.49
# ===> barrier - IoU = 29.59
# ===> bicycle - IoU = 7.38
# ===> bus - IoU = 30.32
# ===> car - IoU = 32.22
# ===> construction_vehicle - IoU = 13.04
# ===> motorcycle - IoU = 11.91
# ===> pedestrian - IoU = 8.61
# ===> traffic_cone - IoU = 8.11
# ===> trailer - IoU = 7.66
# ===> truck - IoU = 20.84
# ===> driveable_surface - IoU = 48.59
# ===> other_flat - IoU = 26.62
# ===> sidewalk - IoU = 26.08
# ===> terrain - IoU = 20.86
# ===> manmade - IoU = 7.62
# ===> vegetation - IoU = 7.14
# ===> mIoU of 6019 samples: 18.3
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-stbase-4d-stereo-512x1408_4x4_1e-2.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (512, 1408),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
# Model
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 6.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 80
multi_adj_frame_id_cfg = (1, 1+1, 1)
model = dict(
type='BEVStereo4DOCC',
align_after_view_transfromation=False,
num_adj=len(range(*multi_adj_frame_id_cfg)),
img_backbone=dict(
type='SwinTransformer',
pretrain_img_size=224,
patch_size=4,
window_size=12,
mlp_ratio=4,
embed_dims=128,
depths=[2, 2, 18, 2],
num_heads=[4, 8, 16, 32],
strides=(4, 2, 2, 2),
out_indices=(2, 3),
qkv_bias=True,
qk_scale=None,
patch_norm=True,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.1,
use_abs_pos_embed=False,
return_stereo_feat=True,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN', requires_grad=True),
pretrain_style='official',
output_missing_index_as_none=False),
img_neck=dict(
type='FPN_LSS',
in_channels=512 + 1024,
out_channels=512,
# with_cp=False,
extra_upsample=None,
input_feature_index=(0, 1),
scale_factor=2),
img_view_transformer=dict(
type='LSSViewTransformerBEVStereo',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=512,
out_channels=numC_Trans,
sid=False,
collapse_z=True,
loss_depth_weight=0.05,
depthnet_cfg=dict(use_dcn=False,
aspp_mid_channels=96,
stereo=True,
bias=5.),
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet',
with_cp=True,
numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
img_bev_encoder_neck=dict(
type='FPN_LSS',
in_channels=numC_Trans * 8 + numC_Trans * 2,
out_channels=256),
pre_process=dict(
type='CustomResNet',
numC_input=numC_Trans,
num_layer=[1, ],
num_channels=[numC_Trans, ],
stride=[1, ],
backbone_output_ids=[0, ]),
occ_head=dict(
type='BEVOCCHead2D',
in_dim=256,
out_dim=256,
Dz=16,
use_mask=True,
num_classes=18,
use_predicter=True,
class_balance=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar','mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet4d',
multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)
test_data_config = dict(
data_root=data_root,
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4, # with 32 GPU
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=1e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
dict(
type='SyncbnControlHook',
syncbn_start_epoch=0,
),
]
evaluation = dict(interval=6, start=0, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=3)
# load_from="ckpts/bevdet-stbase-4d-stereo-512x1408-cbgs.pth"
resume_from="work_dirs/flashocc-stbase-4d-stereo-512x1408_4x4_1e-2/epoch_5.pth"
# fp16 = dict(loss_scale='dynamic')
# bash tools/dist_train.sh projects/configs/flashocc/flashocc-stbase-4d-stereo-512x1408_4x4_1e-2.py 4
\ No newline at end of file
docker-hub/FlashOCC/Flashocc/projects/configs/flashocc/flashocc-stbase-4d-stereo-512x1408_4x4_2e-4.py 0 → 100644
View file @ d2b71343
_base_ = ['../../../mmdetection3d/configs/_base_/datasets/nus-3d.py',
'../../../mmdetection3d/configs/_base_/default_runtime.py']
plugin = True
plugin_dir = 'projects/mmdet3d_plugin/'
# For nuScenes we usually do 10-class detection
class_names = [
'car', 'truck', 'construction_vehicle', 'bus', 'trailer', 'barrier',
'motorcycle', 'bicycle', 'pedestrian', 'traffic_cone'
]
data_config = {
'cams': [
'CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_LEFT',
'CAM_BACK', 'CAM_BACK_RIGHT'
],
'Ncams':
6,
'input_size': (512, 1408),
'src_size': (900, 1600),
# Augmentation
'resize': (-0.06, 0.11),
'rot': (-5.4, 5.4),
'flip': True,
'crop_h': (0.0, 0.0),
'resize_test': 0.00,
}
# Model
grid_config = {
'x': [-40, 40, 0.4],
'y': [-40, 40, 0.4],
'z': [-1, 5.4, 6.4],
'depth': [1.0, 45.0, 0.5],
}
voxel_size = [0.1, 0.1, 0.2]
numC_Trans = 80
multi_adj_frame_id_cfg = (1, 1+1, 1)
model = dict(
type='BEVStereo4DOCC',
align_after_view_transfromation=False,
num_adj=len(range(*multi_adj_frame_id_cfg)),
img_backbone=dict(
type='SwinTransformer',
pretrain_img_size=224,
patch_size=4,
window_size=12,
mlp_ratio=4,
embed_dims=128,
depths=[2, 2, 18, 2],
num_heads=[4, 8, 16, 32],
strides=(4, 2, 2, 2),
out_indices=(2, 3),
qkv_bias=True,
qk_scale=None,
patch_norm=True,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.1,
use_abs_pos_embed=False,
return_stereo_feat=True,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN', requires_grad=True),
pretrain_style='official',
output_missing_index_as_none=False),
img_neck=dict(
type='FPN_LSS',
in_channels=512 + 1024,
out_channels=512,
# with_cp=False,
extra_upsample=None,
input_feature_index=(0, 1),
scale_factor=2),
img_view_transformer=dict(
type='LSSViewTransformerBEVStereo',
grid_config=grid_config,
input_size=data_config['input_size'],
in_channels=512,
out_channels=numC_Trans,
sid=False,
collapse_z=True,
loss_depth_weight=0.05,
depthnet_cfg=dict(use_dcn=False,
aspp_mid_channels=96,
stereo=True,
bias=5.),
downsample=16),
img_bev_encoder_backbone=dict(
type='CustomResNet',
with_cp=True,
numC_input=numC_Trans * (len(range(*multi_adj_frame_id_cfg))+1),
num_channels=[numC_Trans * 2, numC_Trans * 4, numC_Trans * 8]),
img_bev_encoder_neck=dict(
type='FPN_LSS',
in_channels=numC_Trans * 8 + numC_Trans * 2,
out_channels=256),
pre_process=dict(
type='CustomResNet',
numC_input=numC_Trans,
num_layer=[1, ],
num_channels=[numC_Trans, ],
stride=[1, ],
backbone_output_ids=[0, ]),
occ_head=dict(
type='BEVOCCHead2D',
in_dim=256,
out_dim=256,
Dz=16,
use_mask=True,
num_classes=18,
use_predicter=True,
class_wise=False,
loss_occ=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
ignore_index=255,
loss_weight=1.0
),
)
)
# Data
dataset_type = 'NuScenesDatasetOccpancy'
data_root = 'data/nuscenes/'
file_client_args = dict(backend='disk')
bda_aug_conf = dict(
rot_lim=(-0., 0.),
scale_lim=(1., 1.),
flip_dx_ratio=0.5,
flip_dy_ratio=0.5)
train_pipeline = [
dict(
type='PrepareImageInputs',
is_train=True,
data_config=data_config,
sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=True),
dict(type='LoadOccGTFromFile'),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(type='PointToMultiViewDepth', downsample=1, grid_config=grid_config),
dict(type='DefaultFormatBundle3D', class_names=class_names),
dict(
type='Collect3D', keys=['img_inputs', 'gt_depth', 'voxel_semantics',
'mask_lidar','mask_camera'])
]
test_pipeline = [
dict(type='PrepareImageInputs', data_config=data_config, sequential=True),
dict(
type='LoadAnnotationsBEVDepth',
bda_aug_conf=bda_aug_conf,
classes=class_names,
is_train=False),
dict(
type='LoadPointsFromFile',
coord_type='LIDAR',
load_dim=5,
use_dim=5,
file_client_args=file_client_args),
dict(
type='MultiScaleFlipAug3D',
img_scale=(1333, 800),
pts_scale_ratio=1,
flip=False,
transforms=[
dict(
type='DefaultFormatBundle3D',
class_names=class_names,
with_label=False),
dict(type='Collect3D', keys=['points', 'img_inputs'])
])
]
input_modality = dict(
use_lidar=False,
use_camera=True,
use_radar=False,
use_map=False,
use_external=False)
share_data_config = dict(
type=dataset_type,
classes=class_names,
modality=input_modality,
stereo=True,
filter_empty_gt=False,
img_info_prototype='bevdet4d',
multi_adj_frame_id_cfg=multi_adj_frame_id_cfg,
)
test_data_config = dict(
data_root=data_root,
pipeline=test_pipeline,
ann_file=data_root + 'bevdetv2-nuscenes_infos_val.pkl')
data = dict(
samples_per_gpu=4, # with 32 GPU
workers_per_gpu=4,
train=dict(
data_root=data_root,
ann_file=data_root + 'bevdetv2-nuscenes_infos_train.pkl',
pipeline=train_pipeline,
classes=class_names,
test_mode=False,
use_valid_flag=True,
# we use box_type_3d='LiDAR' in kitti and nuscenes dataset
# and box_type_3d='Depth' in sunrgbd and scannet dataset.
box_type_3d='LiDAR'),
val=test_data_config,
test=test_data_config)
for key in ['val', 'train', 'test']:
data[key].update(share_data_config)
# Optimizer
optimizer = dict(type='AdamW', lr=2e-4, weight_decay=1e-2)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=200,
warmup_ratio=0.001,
step=[24, ])
runner = dict(type='EpochBasedRunner', max_epochs=24)
custom_hooks = [
dict(
type='MEGVIIEMAHook',
init_updates=10560,
priority='NORMAL',
),
dict(
type='SyncbnControlHook',
syncbn_start_epoch=0,
),
]
evaluation = dict(interval=6, start=0, pipeline=test_pipeline)
checkpoint_config = dict(interval=1, max_keep_ckpts=3)
load_from="ckpts/bevdet-stbase-4d-stereo-512x1408-cbgs.pth"
# fp16 = dict(loss_scale='dynamic')
# bash tools/dist_train.sh projects/configs/flashocc/flashocc-stbase-4d-stereo-512x1408.py 4
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 13.42
# ===> barrier - IoU = 51.07
# ===> bicycle - IoU = 27.68
# ===> bus - IoU = 51.57
# ===> car - IoU = 56.22
# ===> construction_vehicle - IoU = 27.27
# ===> motorcycle - IoU = 29.98
# ===> pedestrian - IoU = 29.93
# ===> traffic_cone - IoU = 29.8
# ===> trailer - IoU = 37.77
# ===> truck - IoU = 43.52
# ===> driveable_surface - IoU = 83.81
# ===> other_flat - IoU = 46.55
# ===> sidewalk - IoU = 56.15
# ===> terrain - IoU = 59.56
# ===> manmade - IoU = 50.84
# ===> vegetation - IoU = 44.67
# ===> mIoU of 6019 samples: 43.52
# ===> per class IoU of 6019 samples:
# ===> others - IoU = 13.31
# ===> barrier - IoU = 51.62
# ===> bicycle - IoU = 28.07
# ===> bus - IoU = 50.91
# ===> car - IoU = 55.69
# ===> construction_vehicle - IoU = 27.46
# ===> motorcycle - IoU = 31.05
# ===> pedestrian - IoU = 29.98
# ===> traffic_cone - IoU = 29.2
# ===> trailer - IoU = 38.86
# ===> truck - IoU = 43.68
# ===> driveable_surface - IoU = 83.87
# ===> other_flat - IoU = 45.63
# ===> sidewalk - IoU = 56.33
# ===> terrain - IoU = 59.01
# ===> manmade - IoU = 50.63
# ===> vegetation - IoU = 44.56
# ===> mIoU of 6019 samples: 43.52
# {'mIoU': array([0.13311691, 0.51617081, 0.28070517, 0.50911942, 0.55694228,
# 0.27461342, 0.31050779, 0.29979125, 0.29204287, 0.38862984,
# 0.43680049, 0.83872518, 0.45630227, 0.56327839, 0.59008883,
# 0.50627122, 0.44564523, 0.90959399])}
\ No newline at end of file
docker-hub/FlashOCC/Flashocc/projects/configs/panoptic-flashocc/panoptic-flashocc-r50-depth-tiny-pano-trt.py 0 → 100644
View file @ d2b71343
_base_ = ['./flashoccv2-r50-depth-tiny-pano.py',
]
model = dict(
wocc=True,
wdet3d=False,
)
docker-hub/FlashOCC/Flashocc/projects/configs/panoptic-flashocc/panoptic-flashocc-r50-depth-trt.py 0 → 100644
View file @ d2b71343
_base_ = ['./flashoccv2-r50-depth.py',
]
model = dict(
wocc=True,
wdet3d=False,
)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment