File size: 10,838 Bytes
34d1f8b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 |
from collections import OrderedDict
from copy import deepcopy
from typing import Dict, List, Optional, Tuple
import numpy as np
import torch
import torch.distributed as dist
from mmengine.utils import is_list_of
from torch import Tensor
from torch.nn import functional as F
from mmdet3d.models import Base3DDetector
from mmdet3d.registry import MODELS
from mmdet3d.structures import Det3DDataSample
from mmdet3d.utils import OptConfigType, OptMultiConfig, OptSampleList
from .ops import Voxelization
@MODELS.register_module()
class BEVFusion(Base3DDetector):
def __init__(
self,
data_preprocessor: OptConfigType = None,
pts_voxel_encoder: Optional[dict] = None,
pts_middle_encoder: Optional[dict] = None,
fusion_layer: Optional[dict] = None,
img_backbone: Optional[dict] = None,
pts_backbone: Optional[dict] = None,
view_transform: Optional[dict] = None,
img_neck: Optional[dict] = None,
pts_neck: Optional[dict] = None,
bbox_head: Optional[dict] = None,
init_cfg: OptMultiConfig = None,
seg_head: Optional[dict] = None,
**kwargs,
) -> None:
voxelize_cfg = data_preprocessor.pop('voxelize_cfg')
super().__init__(
data_preprocessor=data_preprocessor, init_cfg=init_cfg)
self.voxelize_reduce = voxelize_cfg.pop('voxelize_reduce')
self.pts_voxel_layer = Voxelization(**voxelize_cfg)
self.pts_voxel_encoder = MODELS.build(pts_voxel_encoder)
self.img_backbone = MODELS.build(
img_backbone) if img_backbone is not None else None
self.img_neck = MODELS.build(
img_neck) if img_neck is not None else None
self.view_transform = MODELS.build(
view_transform) if view_transform is not None else None
self.pts_middle_encoder = MODELS.build(pts_middle_encoder)
self.fusion_layer = MODELS.build(
fusion_layer) if fusion_layer is not None else None
self.pts_backbone = MODELS.build(pts_backbone)
self.pts_neck = MODELS.build(pts_neck)
self.bbox_head = MODELS.build(bbox_head)
self.init_weights()
def _forward(self,
batch_inputs: Tensor,
batch_data_samples: OptSampleList = None):
"""Network forward process.
Usually includes backbone, neck and head forward without any post-
processing.
"""
pass
def parse_losses(
self, losses: Dict[str, torch.Tensor]
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
"""Parses the raw outputs (losses) of the network.
Args:
losses (dict): Raw output of the network, which usually contain
losses and other necessary information.
Returns:
tuple[Tensor, dict]: There are two elements. The first is the
loss tensor passed to optim_wrapper which may be a weighted sum
of all losses, and the second is log_vars which will be sent to
the logger.
"""
log_vars = []
for loss_name, loss_value in losses.items():
if isinstance(loss_value, torch.Tensor):
log_vars.append([loss_name, loss_value.mean()])
elif is_list_of(loss_value, torch.Tensor):
log_vars.append(
[loss_name,
sum(_loss.mean() for _loss in loss_value)])
else:
raise TypeError(
f'{loss_name} is not a tensor or list of tensors')
loss = sum(value for key, value in log_vars if 'loss' in key)
log_vars.insert(0, ['loss', loss])
log_vars = OrderedDict(log_vars) # type: ignore
for loss_name, loss_value in log_vars.items():
# reduce loss when distributed training
if dist.is_available() and dist.is_initialized():
loss_value = loss_value.data.clone()
dist.all_reduce(loss_value.div_(dist.get_world_size()))
log_vars[loss_name] = loss_value.item()
return loss, log_vars # type: ignore
def init_weights(self) -> None:
if self.img_backbone is not None:
self.img_backbone.init_weights()
@property
def with_bbox_head(self):
"""bool: Whether the detector has a box head."""
return hasattr(self, 'bbox_head') and self.bbox_head is not None
@property
def with_seg_head(self):
"""bool: Whether the detector has a segmentation head.
"""
return hasattr(self, 'seg_head') and self.seg_head is not None
def extract_img_feat(
self,
x,
points,
lidar2image,
camera_intrinsics,
camera2lidar,
img_aug_matrix,
lidar_aug_matrix,
img_metas,
) -> torch.Tensor:
B, N, C, H, W = x.size()
x = x.view(B * N, C, H, W).contiguous()
x = self.img_backbone(x)
x = self.img_neck(x)
if not isinstance(x, torch.Tensor):
x = x[0]
BN, C, H, W = x.size()
x = x.view(B, int(BN / B), C, H, W)
with torch.autocast(device_type='cuda', dtype=torch.float32):
x = self.view_transform(
x,
points,
lidar2image,
camera_intrinsics,
camera2lidar,
img_aug_matrix,
lidar_aug_matrix,
img_metas,
)
return x
def extract_pts_feat(self, batch_inputs_dict) -> torch.Tensor:
points = batch_inputs_dict['points']
with torch.autocast('cuda', enabled=False):
points = [point.float() for point in points]
feats, coords, sizes = self.voxelize(points)
batch_size = coords[-1, 0] + 1
x = self.pts_middle_encoder(feats, coords, batch_size)
return x
@torch.no_grad()
def voxelize(self, points):
feats, coords, sizes = [], [], []
for k, res in enumerate(points):
ret = self.pts_voxel_layer(res)
if len(ret) == 3:
# hard voxelize
f, c, n = ret
else:
assert len(ret) == 2
f, c = ret
n = None
feats.append(f)
coords.append(F.pad(c, (1, 0), mode='constant', value=k))
if n is not None:
sizes.append(n)
feats = torch.cat(feats, dim=0)
coords = torch.cat(coords, dim=0)
if len(sizes) > 0:
sizes = torch.cat(sizes, dim=0)
if self.voxelize_reduce:
feats = feats.sum(
dim=1, keepdim=False) / sizes.type_as(feats).view(-1, 1)
feats = feats.contiguous()
return feats, coords, sizes
def predict(self, batch_inputs_dict: Dict[str, Optional[Tensor]],
batch_data_samples: List[Det3DDataSample],
**kwargs) -> List[Det3DDataSample]:
"""Forward of testing.
Args:
batch_inputs_dict (dict): The model input dict which include
'points' keys.
- points (list[torch.Tensor]): Point cloud of each sample.
batch_data_samples (List[:obj:`Det3DDataSample`]): The Data
Samples. It usually includes information such as
`gt_instance_3d`.
Returns:
list[:obj:`Det3DDataSample`]: Detection results of the
input sample. Each Det3DDataSample usually contain
'pred_instances_3d'. And the ``pred_instances_3d`` usually
contains following keys.
- scores_3d (Tensor): Classification scores, has a shape
(num_instances, )
- labels_3d (Tensor): Labels of bboxes, has a shape
(num_instances, ).
- bbox_3d (:obj:`BaseInstance3DBoxes`): Prediction of bboxes,
contains a tensor with shape (num_instances, 7).
"""
batch_input_metas = [item.metainfo for item in batch_data_samples]
feats = self.extract_feat(batch_inputs_dict, batch_input_metas)
if self.with_bbox_head:
outputs = self.bbox_head.predict(feats, batch_input_metas)
res = self.add_pred_to_datasample(batch_data_samples, outputs)
return res
def extract_feat(
self,
batch_inputs_dict,
batch_input_metas,
**kwargs,
):
imgs = batch_inputs_dict.get('imgs', None)
points = batch_inputs_dict.get('points', None)
features = []
if imgs is not None:
imgs = imgs.contiguous()
lidar2image, camera_intrinsics, camera2lidar = [], [], []
img_aug_matrix, lidar_aug_matrix = [], []
for i, meta in enumerate(batch_input_metas):
lidar2image.append(meta['lidar2img'])
camera_intrinsics.append(meta['cam2img'])
camera2lidar.append(meta['cam2lidar'])
img_aug_matrix.append(meta.get('img_aug_matrix', np.eye(4)))
lidar_aug_matrix.append(
meta.get('lidar_aug_matrix', np.eye(4)))
lidar2image = imgs.new_tensor(np.asarray(lidar2image))
camera_intrinsics = imgs.new_tensor(np.array(camera_intrinsics))
camera2lidar = imgs.new_tensor(np.asarray(camera2lidar))
img_aug_matrix = imgs.new_tensor(np.asarray(img_aug_matrix))
lidar_aug_matrix = imgs.new_tensor(np.asarray(lidar_aug_matrix))
img_feature = self.extract_img_feat(imgs, deepcopy(points),
lidar2image, camera_intrinsics,
camera2lidar, img_aug_matrix,
lidar_aug_matrix,
batch_input_metas)
features.append(img_feature)
pts_feature = self.extract_pts_feat(batch_inputs_dict)
features.append(pts_feature)
if self.fusion_layer is not None:
x = self.fusion_layer(features)
else:
assert len(features) == 1, features
x = features[0]
x = self.pts_backbone(x)
x = self.pts_neck(x)
return x
def loss(self, batch_inputs_dict: Dict[str, Optional[Tensor]],
batch_data_samples: List[Det3DDataSample],
**kwargs) -> List[Det3DDataSample]:
batch_input_metas = [item.metainfo for item in batch_data_samples]
feats = self.extract_feat(batch_inputs_dict, batch_input_metas)
losses = dict()
if self.with_bbox_head:
bbox_loss = self.bbox_head.loss(feats, batch_data_samples)
losses.update(bbox_loss)
return losses
|