def build_net_loss(model_cfg, target_assigner):
num_class, use_mcnms, use_rotate_nms, nms_pre_max_sizes, nms_post_max_sizes, \
nms_score_thresholds, nms_iou_thresholds = load_target_assigner_param(model_cfg)
losses = losses_builder.build(model_cfg.loss)
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
net_loss = PointPillarsLoss(target_assigner,
nms_score_thresholds=nms_score_thresholds,
nms_iou_thresholds=nms_iou_thresholds,
nms_pre_max_sizes=nms_pre_max_sizes,
nms_post_max_sizes=nms_post_max_sizes,
cls_loss_ftor=cls_loss_ftor,
loc_loss_ftor=loc_loss_ftor,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight)
return net_loss
def build(model_cfg: second_pb2.VoxelNet,
voxel_generator,
target_assigner,
measure_time=False):
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]]
classes_cfg = model_cfg.target_assigner.class_settings
num_class = len(classes_cfg)
use_mcnms = [c.use_multi_class_nms for c in classes_cfg]
use_rotate_nms = [c.use_rotate_nms for c in classes_cfg]
if len(model_cfg.target_assigner.nms_pre_max_sizes) != 0:
nms_pre_max_sizes = list(model_cfg.target_assigner.nms_pre_max_sizes)
assert len(nms_pre_max_sizes) == num_class
else:
nms_pre_max_sizes = [c.nms_pre_max_size for c in classes_cfg]
if len(model_cfg.target_assigner.nms_post_max_sizes) != 0:
nms_post_max_sizes = list(model_cfg.target_assigner.nms_post_max_sizes)
assert len(nms_post_max_sizes) == num_class
else:
nms_post_max_sizes = [c.nms_post_max_size for c in classes_cfg]
if len(model_cfg.target_assigner.nms_score_thresholds) != 0:
nms_score_thresholds = list(
model_cfg.target_assigner.nms_score_thresholds)
assert len(nms_score_thresholds) == num_class
else:
nms_score_thresholds = [c.nms_score_threshold for c in classes_cfg]
if len(model_cfg.target_assigner.nms_iou_thresholds) != 0:
nms_iou_thresholds = list(model_cfg.target_assigner.nms_iou_thresholds)
assert len(nms_iou_thresholds) == num_class
else:
nms_iou_thresholds = [c.nms_iou_threshold for c in classes_cfg]
assert all(use_mcnms) or all([not b for b in use_mcnms]), "not implemented"
assert all(use_rotate_nms) or all([not b for b in use_rotate_nms
]), "not implemented"
if all([not b for b in use_mcnms]):
assert all([e == nms_pre_max_sizes[0] for e in nms_pre_max_sizes])
assert all([e == nms_post_max_sizes[0] for e in nms_post_max_sizes])
assert all(
[e == nms_score_thresholds[0] for e in nms_score_thresholds])
assert all([e == nms_iou_thresholds[0] for e in nms_iou_thresholds])
num_input_features = model_cfg.num_point_features
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
3: LossNormType.DontNorm,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
sin_error_factor = model_cfg.sin_error_factor
if sin_error_factor == 0:
sin_error_factor = 1.0
print("CLASS NAME:", model_cfg.network_class_name)
net = get_voxelnet_class(model_cfg.network_class_name)(
dense_shape,
num_class=num_class,
vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
vfe_num_filters=vfe_num_filters,
middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
middle_num_input_features=model_cfg.middle_feature_extractor.
num_input_features,
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name,
rpn_num_input_features=model_cfg.rpn.num_input_features,
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_rotate_nms=all(use_rotate_nms),
multiclass_nms=all(use_mcnms),
nms_score_thresholds=nms_score_thresholds,
nms_pre_max_sizes=nms_pre_max_sizes,
nms_post_max_sizes=nms_post_max_sizes,
nms_iou_thresholds=nms_iou_thresholds,
use_sigmoid_score=model_cfg.use_sigmoid_score,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
measure_time=measure_time,
voxel_generator=voxel_generator,
post_center_range=list(model_cfg.post_center_limit_range),
dir_offset=model_cfg.direction_offset,
sin_error_factor=sin_error_factor,
nms_class_agnostic=model_cfg.nms_class_agnostic,
num_direction_bins=model_cfg.num_direction_bins,
direction_limit_offset=model_cfg.direction_limit_offset,
)
return net
def build(model_cfg: second_pb2.VoxelNet,
voxel_generator,
target_assigner,
measure_time=False) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]]
num_class = len(target_assigner.classes)
class_cfg = model_cfg.target_assigner
use_iou_param_partaa = class_cfg.use_iou_param_partaa
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, iou_loss_ftor, \
cls_weight, loc_weight, iou_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
# if model_cfg.use_rpn:
net = VoxelNet(
dense_shape,
num_class=num_class,
vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
vfe_num_filters=vfe_num_filters,
middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
middle_num_input_features=model_cfg.middle_feature_extractor.
num_input_features,
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name,
rpn_num_input_features=model_cfg.rpn.num_input_features,
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_voxel_classifier=model_cfg.use_aux_classifier,
use_rotate_nms=model_cfg.use_rotate_nms,
multiclass_nms=model_cfg.use_multi_class_nms,
nms_score_threshold=model_cfg.nms_score_threshold,
nms_pre_max_size=model_cfg.nms_pre_max_size,
nms_post_max_size=model_cfg.nms_post_max_size,
nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
use_sparse_rpn=False,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
use_bev=model_cfg.use_bev,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
measure_time=measure_time,
use_iou_branch=model_cfg.use_iou_branch,
iou_dict=model_cfg.iou,
use_iou_param_partaa=use_iou_param_partaa,
iou_loss_weight=iou_weight,
iou_loss_ftor=iou_loss_ftor)
return net
def build(model_cfg: second_pb2.VoxelNet, fv_generator, #voxel_generator,
save_path=None, RGB_embedding=False) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
# grid_size = voxel_generator.grid_size
fv_dim = fv_generator.fv_dim
dense_shape = [1] + fv_dim[::-1].tolist() + [vfe_num_filters[-1]]
num_class = model_cfg.num_class
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
net = VoxelNet(
dense_shape,
num_class=num_class,
# vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name,
# vfe_num_filters=vfe_num_filters,
# middle_class_name=model_cfg.middle_feature_extractor.module_class_name,
# middle_num_filters_d1=list(
# model_cfg.middle_feature_extractor.num_filters_down1),
# middle_num_filters_d2=list(
# model_cfg.middle_feature_extractor.num_filters_down2),
# rpn_class_name=model_cfg.rpn.module_class_name,
# rpn_layer_nums=list(model_cfg.rpn.layer_nums),
# rpn_layer_strides=list(model_cfg.rpn.layer_strides),
# rpn_num_filters=list(model_cfg.rpn.num_filters),
# rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
# rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
# use_norm=True,
# use_rotate_nms=model_cfg.use_rotate_nms,
# multiclass_nms=model_cfg.use_multi_class_nms,
# nms_score_threshold=model_cfg.nms_score_threshold,
# nms_pre_max_size=model_cfg.nms_pre_max_size,
# nms_post_max_size=model_cfg.nms_post_max_size,
# nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
# use_direction_classifier=model_cfg.use_direction_classifier,
# use_bev=model_cfg.use_bev,
# num_input_features=num_input_features,
# num_groups=model_cfg.rpn.num_groups,
# use_groupnorm=model_cfg.rpn.use_groupnorm,
# with_distance=vfe_with_distance,
# cls_loss_weight=cls_weight,
# loc_loss_weight=loc_weight,
# pos_cls_weight=pos_cls_weight,
# neg_cls_weight=neg_cls_weight,
# direction_loss_weight=direction_loss_weight,
# loss_norm_type=loss_norm_type,
# encode_rad_error_by_sin=encode_rad_error_by_sin,
# loc_loss_ftor=loc_loss_ftor,
# cls_loss_ftor=cls_loss_ftor,
# voxel_size=voxel_generator.voxel_size,
# pc_range=voxel_generator.point_cloud_range,
# grid_size=fv_generator.grid_size,
# cartesian_coord_range=fv_generator.cartesian_coord_range,
save_path=save_path,
RGB_embedding=RGB_embedding
)
return net
def build(model_cfg: second_pb2.VoxelNet, voxel_generator,
target_assigner) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters)
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]]
num_class = model_cfg.num_class
num_input_features = model_cfg.num_point_features
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
net = VoxelNetForCudaImplementation(
dense_shape,
num_class=num_class,
vfe_num_filters=vfe_num_filters,
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_rotate_nms=model_cfg.use_rotate_nms,
multiclass_nms=model_cfg.use_multi_class_nms,
nms_score_threshold=model_cfg.nms_score_threshold,
nms_pre_max_size=model_cfg.nms_pre_max_size,
nms_post_max_size=model_cfg.nms_post_max_size,
nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score,
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
use_bev=model_cfg.use_bev,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight,
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
voxel_size=voxel_generator.voxel_size,
pc_range=voxel_generator.point_cloud_range)
return net
def build(model_cfg: second_pb2.VoxelNet, voxel_generator,
target_assigner) -> VoxelNet:
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters) #64
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance
grid_size = voxel_generator.grid_size #[432,496,1]
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]
] #[1,1,496,432,64]
num_class = model_cfg.num_class
num_input_features = model_cfg.num_point_features #4
if model_cfg.without_reflectivity:
num_input_features = 3
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss) # 建立损失函数
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight
neg_cls_weight = model_cfg.neg_class_weight
direction_loss_weight = model_cfg.direction_loss_weight
net = VoxelNet( # 以VoxelNet为基础搭建网络
dense_shape, # 输入VFE的张量格式
num_class=num_class, # 训练类别
vfe_class_name=model_cfg.voxel_feature_extractor.
module_class_name, # VFE配置
vfe_num_filters=vfe_num_filters, # 64
middle_class_name=model_cfg.middle_feature_extractor.
module_class_name, # 中间特征提取层配置
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name, # SSD网络配置
rpn_layer_nums=list(model_cfg.rpn.layer_nums),
rpn_layer_strides=list(model_cfg.rpn.layer_strides),
rpn_num_filters=list(model_cfg.rpn.num_filters),
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters),
use_norm=True,
use_rotate_nms=model_cfg.use_rotate_nms, # NMS配置
multiclass_nms=model_cfg.use_multi_class_nms,
nms_score_threshold=model_cfg.nms_score_threshold,
nms_pre_max_size=model_cfg.nms_pre_max_size,
nms_post_max_size=model_cfg.nms_post_max_size,
nms_iou_threshold=model_cfg.nms_iou_threshold,
use_sigmoid_score=model_cfg.use_sigmoid_score, # 得分设置
encode_background_as_zeros=model_cfg.encode_background_as_zeros,
use_direction_classifier=model_cfg.use_direction_classifier,
use_bev=model_cfg.use_bev,
num_input_features=num_input_features,
num_groups=model_cfg.rpn.num_groups,
use_groupnorm=model_cfg.rpn.use_groupnorm,
with_distance=vfe_with_distance,
cls_loss_weight=cls_weight, # 不同损失权重设置
loc_loss_weight=loc_weight,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
direction_loss_weight=direction_loss_weight,
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin,
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner, # 训练目标
voxel_size=voxel_generator.voxel_size, # 体素参数配置
pc_range=voxel_generator.point_cloud_range)
return net
def build(model_cfg: second_pb2.VoxelNet, voxel_generator,
target_assigner, measure_time=False):
"""build second pytorch instance.
"""
if not isinstance(model_cfg, second_pb2.VoxelNet):
raise ValueError('model_cfg not of type ' 'second_pb2.VoxelNet.')
vfe_num_filters = list(model_cfg.voxel_feature_extractor.num_filters) #64
vfe_with_distance = model_cfg.voxel_feature_extractor.with_distance #false
grid_size = voxel_generator.grid_size #[400,400,1] max=400
dense_shape = [1] + grid_size[::-1].tolist() + [vfe_num_filters[-1]] #[1,1,400,400,64]
classes_cfg = model_cfg.target_assigner.class_settings
num_class = len(classes_cfg) #10
use_mcnms = [c.use_multi_class_nms for c in classes_cfg] #false
use_rotate_nms = [c.use_rotate_nms for c in classes_cfg] #false
if len(model_cfg.target_assigner.nms_pre_max_sizes) != 0:
nms_pre_max_sizes = list(model_cfg.target_assigner.nms_pre_max_sizes)
assert len(nms_pre_max_sizes) == num_class
else:
nms_pre_max_sizes = [c.nms_pre_max_size for c in classes_cfg]
if len(model_cfg.target_assigner.nms_post_max_sizes) != 0:
nms_post_max_sizes = list(model_cfg.target_assigner.nms_post_max_sizes)
assert len(nms_post_max_sizes) == num_class
else:
nms_post_max_sizes = [c.nms_post_max_size for c in classes_cfg]
if len(model_cfg.target_assigner.nms_score_thresholds) != 0:
nms_score_thresholds = list(model_cfg.target_assigner.nms_score_thresholds)
assert len(nms_score_thresholds) == num_class
else:
nms_score_thresholds = [c.nms_score_threshold for c in classes_cfg]
if len(model_cfg.target_assigner.nms_iou_thresholds) != 0:
nms_iou_thresholds = list(model_cfg.target_assigner.nms_iou_thresholds)
assert len(nms_iou_thresholds) == num_class
else:
nms_iou_thresholds = [c.nms_iou_threshold for c in classes_cfg]
assert all(use_mcnms) or all([not b for b in use_mcnms]), "not implemented" #false
assert all(use_rotate_nms) or all([not b for b in use_rotate_nms]), "not implemented"
if all([not b for b in use_mcnms]):
assert all([e == nms_pre_max_sizes[0] for e in nms_pre_max_sizes]) #e:1000
assert all([e == nms_post_max_sizes[0] for e in nms_post_max_sizes]) #e:300
assert all([e == nms_score_thresholds[0] for e in nms_score_thresholds]) #e:0.05
assert all([e == nms_iou_thresholds[0] for e in nms_iou_thresholds]) #e:0.5
num_input_features = model_cfg.num_point_features #4
loss_norm_type_dict = {
0: LossNormType.NormByNumExamples,
1: LossNormType.NormByNumPositives,
2: LossNormType.NormByNumPosNeg,
3: LossNormType.DontNorm,
}
loss_norm_type = loss_norm_type_dict[model_cfg.loss_norm_type]
losses = losses_builder.build(model_cfg.loss)
encode_rad_error_by_sin = model_cfg.encode_rad_error_by_sin
cls_loss_ftor, loc_loss_ftor, cls_weight, loc_weight, _ = losses
pos_cls_weight = model_cfg.pos_class_weight #1.0
neg_cls_weight = model_cfg.neg_class_weight #1.0
direction_loss_weight = model_cfg.direction_loss_weight #0.2
sin_error_factor = model_cfg.sin_error_factor #1.0
if sin_error_factor == 0:
sin_error_factor = 1.0
net = get_voxelnet_class(model_cfg.network_class_name)( #network_class_name :VoxelNet
dense_shape, #[1,1,400,400,64]
num_class=num_class,
vfe_class_name=model_cfg.voxel_feature_extractor.module_class_name, #module_calss_name:pillarfeaturenet
# vfe_radar_class_name = model_cfg.voxel_radar_feature_extractor.module_radar_class_name,
vfe_num_filters=vfe_num_filters, #64
middle_class_name=model_cfg.middle_feature_extractor.module_class_name, #pointpillarscatter
middle_num_input_features=model_cfg.middle_feature_extractor.num_input_features, #64
middle_num_filters_d1=list(
model_cfg.middle_feature_extractor.num_filters_down1),
middle_num_filters_d2=list(
model_cfg.middle_feature_extractor.num_filters_down2),
rpn_class_name=model_cfg.rpn.module_class_name, #RPNV2
rpn_num_input_features=model_cfg.rpn.num_input_features, #64
rpn_layer_nums=list(model_cfg.rpn.layer_nums), #[3,5,5]
rpn_layer_strides=list(model_cfg.rpn.layer_strides), #[2,2,2]
rpn_num_filters=list(model_cfg.rpn.num_filters), #[64,128,256]
rpn_upsample_strides=list(model_cfg.rpn.upsample_strides),#[0.25,0.5,1]
rpn_num_upsample_filters=list(model_cfg.rpn.num_upsample_filters), #[128,128,128]
use_norm=True,
use_rotate_nms=all(use_rotate_nms),
multiclass_nms=all(use_mcnms),
nms_score_thresholds=nms_score_thresholds, #[0.05]
nms_pre_max_sizes=nms_pre_max_sizes, #1000
nms_post_max_sizes=nms_post_max_sizes, #300
nms_iou_thresholds=nms_iou_thresholds, #0.5
use_sigmoid_score=model_cfg.use_sigmoid_score, #true
encode_background_as_zeros=model_cfg.encode_background_as_zeros, #true
use_direction_classifier=model_cfg.use_direction_classifier, #true
num_input_features=num_input_features, #4
num_groups=model_cfg.rpn.num_groups, #32
use_groupnorm=model_cfg.rpn.use_groupnorm, #false
with_distance=vfe_with_distance, #false
cls_loss_weight=cls_weight, #1
loc_loss_weight=loc_weight, #1
pos_cls_weight=pos_cls_weight, #1
neg_cls_weight=neg_cls_weight, #1
direction_loss_weight=direction_loss_weight, #0.2
loss_norm_type=loss_norm_type,
encode_rad_error_by_sin=encode_rad_error_by_sin, #true
loc_loss_ftor=loc_loss_ftor,
cls_loss_ftor=cls_loss_ftor,
target_assigner=target_assigner,
measure_time=measure_time, #false
voxel_generator=voxel_generator, #voxel_size高20，底面0.25,0.25
post_center_range=list(model_cfg.post_center_limit_range), #[-59.599998474121094, -59.599998474121094, -10.0, 59.599998474121094, 59.599998474121094, 10.0]
dir_offset=model_cfg.direction_offset, #0.779
sin_error_factor=sin_error_factor, #1.0
nms_class_agnostic=model_cfg.nms_class_agnostic, #false
num_direction_bins=model_cfg.num_direction_bins, #2
direction_limit_offset=model_cfg.direction_limit_offset, #0.0
)
return net
